Plain-English definitions of the AI terms you'll hear around legal and document AI — what each one means, why it matters, and a concrete example.
Explore the visual mapShowing 977 of 977 terms
A way to compare two versions by showing each to a different group of people and seeing which performs better.
It replaces guessing with evidence, letting you prove a change is actually an improvement before rolling it out to everyone.
Showing half of users a new answer-summary format and half the old one to see which gets more clicks.
A test where you remove one part of a system to see how much that part actually mattered to the result.
It tells you which components are doing real work and which are dead weight, so you don't credit a feature that wasn't helping.
Turning off the document-reranking step in a search tool to measure how much answer quality drops without it.
A specialized computer chip built to run AI math much faster than an ordinary processor.
These chips are why large models can train and answer quickly; without them, modern AI would be far too slow and expensive.
Running an AI search engine on GPUs so it can scan thousands of documents in seconds.
The duty of a person or organization to answer for what an AI system does, take responsibility for it, and explain the results openly.
Without clear accountability, no one owns the harm an AI system causes, which makes problems hard to fix and trust hard to earn.
A firm using an AI tool to flag relevant documents assigns a named reviewer who must sign off on and explain the tool's selections.
The share of predictions a model got right out of all the predictions it made.
It is the simplest scorecard for a model, though it can be misleading when one outcome is much rarer than the other.
A document classifier that correctly labels 92 out of 100 contracts has 92% accuracy.
A move an AI agent chooses to make in its environment to try to reach a goal.
Actions are the agent's only way to affect the world, so learning which actions pay off is the core of reinforcement learning.
In a game-playing agent, moving a piece left or right is an action.
A small math rule inside a neural network that decides how strongly a signal passes to the next layer.
It lets networks learn complicated, non-straight-line patterns instead of only simple ones.
The ReLU activation passes positive values through unchanged and turns negative values into zero.
A training approach where the model asks humans to label only the examples it finds most confusing.
It cuts the cost of labeling by focusing human effort where it helps the model learn the most.
An AI flags the 50 ambiguous exhibits it is least sure about and asks a lawyer to label just those.
A learning program that gets to choose which actions to take and which data to learn from, instead of just following a fixed set of rules.
Choosing its own informative examples lets the agent learn faster and with less labeled data than a passive system that takes whatever it is given.
A system reviewing documents asks a human to label the cases it is most unsure about, learning more from each answer.
A unit of work an organization performs as part of a business process, which can be a single step or several steps combined.
Breaking work into named activities makes it possible to map, measure, and automate how a process actually runs.
In a document intake process, 'scan the file' and 'log the case number' are two activities.
A training method that automatically gives each setting in a model its own learning speed.
It helps models learn rare but important features without manually tuning the speed for every part.
During training, AdaGrad slows down updates to settings that have already changed a lot and speeds up rarely touched ones.
A popular training method that adjusts how fast a model learns by tracking the recent history of its updates.
It is the default optimizer for most modern AI because it trains reliably with little manual tuning.
A team trains a text classifier with Adam and gets stable progress without hand-setting the learning speed.
A reinforcement-learning method where an agent learns the rules of how its world changes and then uses those rules to figure out the best long-term choices.
By learning the world's transition model rather than just memorizing rewards, the agent can plan ahead and reach good decisions more efficiently.
A delivery robot learns how its moves change its position, then plans the shortest reliable route through a building.
Updating a prediction model on the fly while it runs so it keeps up when the patterns in incoming data change.
Real-world data shifts over time, and a model that never updates slowly becomes wrong; adaptive learning keeps it current.
A spam filter that keeps retraining as spammers change their tactics so it stays accurate.
An input that has been tweaked just slightly, on purpose, so that an AI model gives the wrong answer even though a human sees nothing unusual.
Adversarial examples reveal that models can be fooled by tiny changes, which is a serious security risk for systems that must be reliable.
A few pixels altered on a stop sign cause a self-driving car's vision model to read it as a speed-limit sign.
A defense where a model is deliberately trained on tricky, tampered inputs designed to fool it, so it learns to handle them and becomes harder to deceive.
It hardens a model against attackers who feed it cleverly altered inputs, which matters anywhere a wrong answer could be exploited.
A document classifier could be trained on subtly altered scans so a forger can't slip a mislabeled exhibit past it.
A formal notice telling someone that they were denied credit, had an account changed for the worse, or were refused a credit increase.
When AI helps make lending decisions, the law often requires telling affected people what happened, so these notices are key to fair, transparent automation.
A loan application scored by an AI model is rejected, and the lender sends the applicant a notice explaining the denial.
A number comparing how often a favored group gets a good outcome versus an unfavored group, regardless of anyone's intent.
It is a standard way to spot whether an AI decision is treating one group much worse than another, even when no one meant to discriminate.
If 80% of one group is approved for a loan but only 40% of another, the ratio of 0.5 signals a possible fairness problem.
An AI system that can take multiple steps and use tools to complete a goal, not just answer one question.
Agents can chain actions (search, read, compare, draft), which automates multi-step work — but each step should stay reviewable.
An agent searches a case file, extracts every key date, cross-checks them, and assembles a timeline — in one run.
The agent's current understanding of its situation, which it uses to decide its next action.
Good decisions depend on a good read of the situation, so how an agent represents its state shapes how well it performs.
A navigation agent's state includes its position and which paths it has already tried.
Describes an AI that can plan multiple steps and take actions on its own to finish a task, not just answer one prompt.
Agentic behavior lets AI handle multi-step jobs like researching, deciding, and acting, which is far more powerful but needs careful oversight.
An agentic assistant that reads a case file, searches for missing exhibits, and drafts a summary without step-by-step instructions.
A document-answering setup where an AI can decide on its own which sources to search and how many times to look before answering.
It handles harder questions than plain retrieval because the AI can plan, search, and refine instead of doing a single lookup.
To answer a complex case question, the AI first searches depositions, then decides it also needs the expert report and searches again.
A grouping method that starts with every item alone, then repeatedly merges the closest groups until larger clusters form.
It reveals natural groupings in data without you having to say in advance how many groups there are.
Grouping thousands of case documents by similarity, merging closely related ones step by step.
A way of building software in small, frequent steps so the product grows piece by piece instead of all at once.
Delivering in increments lets teams get feedback early and adjust quickly, instead of discovering problems only at the end.
A team ships a basic document search feature in two weeks, then adds filters and highlights in later short cycles.
The broad field of building software that performs tasks normally requiring human intelligence.
"AI" is an umbrella term, not one product. Knowing it covers many techniques helps you ask what is actually under the hood of any tool.
Reading a scanned report, answering a question about a contract, and flagging an inconsistency are all AI tasks — handled by different methods.
The work of making sure an AI system's goals and behavior match what its human designers and users actually intend.
A capable AI pursuing the wrong objective can cause harm even while working as built, so aligning it to human intent is essential.
An aligned document assistant sticks to summarizing what the files say instead of inventing favorable conclusions the user did not ask for.
A set of shared values that guide how AI should be built and used responsibly, such as fairness, transparency, safety, and accountability.
Principles give teams a north star for making trustworthy AI, turning broad ideals into concrete design and review choices.
Before deploying a document AI tool, a team checks it against principles like transparency and accountability.
Low-quality, generic, or filler content churned out by AI that adds little real value.
Recognizing it matters because flooding documents or the web with AI slop wastes reader time and erodes trust.
An AI-generated case summary padded with vague boilerplate that never cites a specific fact is AI slop.
A clear set of step-by-step rules for solving a problem, usually carried out by a computer.
Algorithms are the core recipes behind every software feature and AI model, defining exactly how a task gets done.
A sorting algorithm puts a list of case files into date order by repeatedly comparing and swapping entries.
The tendency for people to distrust or reject an algorithm's advice more harshly than they would the same advice from a human.
If users dismiss good AI recommendations out of bias, organizations lose the value of the tool, so understanding this reaction matters for adoption.
A reviewer ignores an AI tool's accurate document ranking simply because it came from software rather than a colleague.
Making sure a powerful AI reliably does what people actually intend and value, covering both how to encode those values technically and which values to choose.
An AI that pursues the wrong goal can cause harm even while working as built, so alignment is central to safe, trustworthy systems.
Tuning an assistant so it refuses to fabricate citations rather than just producing confident-sounding answers.
Making something larger or more intense, such as an AI system magnifying a pattern or bias present in its data.
AI can amplify small biases in training data into large, systematic effects, so spotting amplification is important for fairness.
If past hiring data slightly favored one group, a model trained on it can amplify that into a strong, repeated preference.
Using scientific and mathematical methods to study data and draw useful conclusions, covering what happened, what will happen, and what to do about it.
Analytics turns raw data into insight and action, which is the foundation of data-driven decisions in any organization.
A team analyzes past case outcomes to predict which new matters are likely to settle and plan staffing accordingly.
An AI built to handle one specific task or narrow problem, rather than thinking broadly like a human.
Almost all AI in use today is narrow, so recognizing this sets realistic expectations about what a given tool can and can't do.
A tool that only extracts dates from documents is narrow AI; it can't argue a case or hold a conversation.
Extra notes or documentation attached to something, such as a requirement, to explain or clarify it.
Annotations capture the context and reasoning behind an item so future readers understand it correctly.
A note attached to a labeled document explaining why a passage was marked as privileged.
Anything in a system's behavior or documents that doesn't match what you'd expect based on past, verified results.
Anomalies are early warning signs of bugs, fraud, or data problems, so catching them helps keep systems reliable.
A document set where one file's metadata date falls years outside every other file flags an anomaly worth checking.
The task of automatically spotting data points that don't fit the normal pattern.
It surfaces rare problems like fraud, errors, or tampering that would be impossible to find by hand at scale.
Flagging a single invoice among thousands whose amount and formatting look unlike all the others.
Changing data so it can no longer be traced back to a specific person, using techniques like removing, broadening, or scrambling identifying details.
Anonymization protects people's privacy when data is shared or analyzed, though it is hard to do perfectly without losing usefulness.
Replacing exact birthdates with age ranges and removing names before sharing a case dataset for analysis.
Treating an AI as if it has human feelings, intentions, or understanding — when really it is just predicting likely text.
When a chatbot sounds confident and warm, people tend to trust it like a person and forget it can be confidently wrong. Naming that reflex keeps you skeptical and checking the output.
A model writes 'I carefully reviewed every page of the contract.' It reviewed nothing in a human sense — it generated words that pattern-match to careful review, so the claim still needs a real human check.
A defined set of functions that lets one piece of software request services from another, acting as the agreed contract between them.
APIs let separate systems work together, so developers can plug in features like AI search without rebuilding everything from scratch.
A document tool calls an AI provider's API to turn text into embeddings without knowing how the model works internally.
A software program that runs on a computer system to perform tasks for users.
Applications are how people actually use technology day to day, turning underlying systems into usable tools.
A web application that lets reviewers upload, search, and tag case documents in a browser.
A single score summarizing how well a model balances precision and recall across all thresholds.
It is more informative than accuracy when the thing you're hunting for is rare, like a few key documents in a huge set.
Measuring a privilege-detection model with the PR curve area because privileged documents are a tiny fraction of the collection.
A hypothetical AI that could perform any intellectual task a human can, across any subject.
It is the long-term goal some researchers aim for, and the standard against which today's narrow, single-purpose AI is measured.
Today's tools answer document questions well but cannot, unlike imagined AGI, switch to diagnosing illness or arguing a case unprompted.
The process where an AI takes in a body of data, maps its meaning, and finds patterns and connections to make useful predictions or conclusions in a subject area.
This learning is what lets AI move beyond fixed rules to draw inferences and even, in advanced cases, form new hypotheses.
An AI ingests thousands of legal filings and learns to correlate phrasing patterns with case categories.
A machine-based system that, given some goals, produces outputs like predictions, recommendations, or decisions, and can operate with varying degrees of independence.
This is the standard definition used in governance and regulation, so knowing it helps you recognize what counts as 'AI' under the rules.
A tool that reads case documents and recommends which ones are most relevant is an AI system.
Applying specific written criteria to a piece of software to decide whether to accept or release it.
A consistent assessment step ensures software meets agreed standards before it reaches users.
Before release, a document tool is checked against a checklist for accuracy, security, and performance.
A way of training a model on many machines at once where each machine updates the shared model whenever it finishes, without waiting for the others.
It speeds up training on large datasets by keeping every machine busy instead of idling between rounds.
Ten servers each crunch different document batches and push their updates as soon as they're done.
A deliberate action aimed at a learning system to make it malfunction.
Knowing the kinds of attacks AI faces is the first step to defending systems that handle sensitive data and decisions.
Feeding a model carefully crafted inputs to trick it into misclassifying confidential documents.
A mechanism that lets a model focus on the most relevant words when interpreting or generating text.
Attention is the core idea behind modern language models, letting them connect related words even when far apart in a sentence.
When answering a question, the model pays extra attention to the clause in a contract that defines the disputed term.
A single measurable property of an item that a model uses as input.
Attributes are the raw signals a model learns from, so choosing good ones largely determines how well it works.
For predicting case outcomes, attributes might include the court, the filing date, and the type of claim.
A trick in decision forests where each tree is only allowed to consider a random subset of the available features.
It keeps the trees in a forest different from each other, which makes the combined prediction more accurate and robust.
When building a random forest, each split looks at only 5 of 20 features chosen at random.
A single score from 0 to 1 that measures how well a model separates two classes across all decision thresholds.
It summarizes a classifier's quality in one number that does not depend on picking a specific cutoff, making models easy to compare.
A document-relevance model with an AUC of 0.95 almost always ranks a relevant exhibit above an irrelevant one.
A systematic, independent, documented review that gathers evidence and judges it objectively to see how well requirements are met.
Audits provide credible, outside confirmation that a system actually does what it claims, which is essential for trust and compliance.
An independent reviewer audits an AI document tool's logs and outputs to confirm it meets fairness requirements.
A time-ordered record of what a system did, including who accessed it and what operations they performed over a period.
Audit logs let you reconstruct exactly what happened, which is vital for security investigations and proving accountability.
A log showing which user viewed each confidential document and when they did it.
Technology that overlays computer-generated images onto your view of the real world.
It blends digital information with physical surroundings, useful for visualization, training, and interactive displays.
Holding up a phone at a worksite and seeing AI-detected hazard labels appear over the real scene.
The property that something or someone is genuinely what it claims to be.
Verifying authenticity stops impostors and forged data from slipping into a system, protecting both security and trust in the evidence.
Confirming that an uploaded document truly came from the stated source and wasn't tampered with or faked.
A neural network that learns to compress data into a compact form and then rebuild it, capturing the data's essence.
It learns useful patterns without labels and can spot anomalies that don't rebuild well.
Compressing normal contract structures so well that a forged document fails to reconstruct cleanly, flagging it.
When a machine runs and coordinates the steps of a digital task on its own, without a person doing each step.
Automation removes repetitive manual work so people can focus on judgment calls instead of clicking through the same steps over and over.
A document system automatically converts every uploaded scan to text and files it in the right case folder without anyone touching it.
The tendency of people to trust a machine's recommendation even when their own judgment says otherwise.
It can lead reviewers to rubber-stamp wrong AI outputs, which is dangerous in high-stakes legal or medical decisions.
A reviewer accepts an AI's document classification without checking, even though the label is clearly wrong.
A self-managing system that senses its surroundings, makes decisions, and configures, optimizes, fixes, and protects itself with little human help.
Autonomic systems keep running and adapting on their own, which reduces downtime and the need for constant human babysitting.
A server platform that notices one machine is overloaded, shifts the work elsewhere, and repairs itself before users ever notice a slowdown.
A car, truck, boat, or other vehicle that can sense its surroundings and move safely with little or no human driving.
Self-driving vehicles are one of the most visible real-world tests of AI making safety-critical decisions in unpredictable environments.
A self-driving car detects a pedestrian stepping into the road and brakes on its own.
How independently an AI system can operate and make decisions without a human stepping in.
The level of autonomy decides how much human oversight a system needs, which is central to safety and accountability.
A document tool that only suggests answers (human-in-the-loop) has low autonomy, while one that files documents and acts on its own has high autonomy.
An AI model used to automatically grade the quality of another AI's outputs instead of a human grader.
It lets teams evaluate thousands of AI answers cheaply and quickly, though it can carry the grader model's own biases.
Using one model to score whether a chatbot's case summaries are accurate and well-sourced across 1,000 questions.
The practice of measuring AI output quality using an automated rater model rather than human reviewers.
It makes large-scale, repeatable evaluation affordable, which is essential for tuning and comparing models.
Running an autorater evaluation nightly to check whether a new model version answers document questions more faithfully.
Making sure information and systems can be reached and used reliably whenever they are needed.
If a tool goes down or data can't be reached when you need it, work stalls; availability is a core pillar of security alongside confidentiality and integrity.
A case file search tool stays up and responsive even during a busy filing deadline so the team can pull documents on demand.
A score that rewards a ranking system for putting correct results near the top of its list.
It captures how good a search or retrieval system is at ordering results, not just at finding them.
A case-document search engine that surfaces the most relevant exhibits first earns a high average precision.
A decision-tree test that splits data using just one feature at a time, like 'age greater than 40'.
These simple one-feature splits make decision trees easy to read and explain to non-experts.
A tree branches on the single question 'Is the contract value above $1 million?' to route a case.
Checking how good a model's predictions are by comparing what it forecasted against what actually happened.
Back-testing reveals whether a model would have worked on real past data before you trust it with new decisions.
A risk model's predicted outcomes are compared against last year's actual results to see how often it was right.
The core algorithm that teaches a neural network by tracing each mistake backward to figure out how much every internal setting should change.
Backpropagation is what makes modern AI learnable at all; without it, networks could not adjust millions of internal weights from their errors.
After a model wrongly tags a contract clause, backpropagation nudges the weights that contributed most to the error so the next guess is closer.
A simple way to represent text by counting which words appear, while throwing away their order.
It is an easy, fast baseline for text tasks, and understanding its blind spot (ignoring word order) explains why richer methods like embeddings exist.
A bag-of-words model can flag deposition transcripts that mention 'breach' often but cannot tell 'breach of duty' apart from 'duty to breach.'
A simple reference model or score you compare against to judge whether a fancier model is actually worth it.
Without a baseline you cannot tell if a complex AI is helping; a model that barely beats 'always guess the most common answer' may not be worth deploying.
Before trusting an AI to sort exhibits, you check it beats the baseline of simply filing everything under the largest category.
The group of training examples a model looks at together in one step before updating itself.
Batches let models learn efficiently from huge datasets in manageable chunks instead of one example at a time or all at once.
Instead of learning from one labeled page at a time, the model reviews 64 pages per batch and updates after each group.
Running a trained model on a large pile of inputs all at once, rather than one request at a time.
Batch inference is cheaper and faster for bulk jobs where you don't need an instant answer, like processing a whole case folder overnight.
Overnight, the system runs batch inference to summarize all 2,000 newly uploaded expert reports at once.
A technique that rescales the numbers flowing through a neural network during training to keep them in a steady, well-behaved range.
It makes training faster and more stable, letting deeper models learn without their internal numbers blowing up or fading away.
Like keeping the volume on every track at a consistent level so the whole mix stays balanced, batch normalization keeps each layer's outputs steady.
How many examples a model processes together before it updates itself once.
Batch size trades off speed, memory, and learning quality; too big can run out of memory, too small can make learning noisy.
Using a batch size of 32 means the model adjusts its settings after every 32 training documents.
Automating chunks of work that can be processed together as a group, separate from where they sit in a larger workflow.
Grouping similar tasks lets a system process many items at once efficiently instead of handling them one at a time.
Overnight, a system runs text extraction on all 2,000 documents uploaded that day in one batch instead of one by one as they arrive.
A neural network that not only makes a prediction but also reports how unsure it is about that prediction.
Knowing the model's confidence matters in high-stakes settings, so you can flag uncertain answers for a human to review.
A Bayesian model classifying a blurry exhibit might say 'likely a signature page, but I'm only 55% sure,' prompting manual check.
A smart search method for tuning settings that uses past results to decide which combination to try next, instead of testing blindly.
It finds good model settings with far fewer expensive trial runs, saving time and compute when each test is costly.
Rather than trying hundreds of random learning-rate values, Bayesian optimization learns from each run to zero in on the best one.
A formula that says the value of a situation equals the immediate reward plus the value of where you end up next.
It is the mathematical backbone of how AI agents learn to make good long-term decisions instead of just chasing the next reward.
An agent planning a chess move uses Bellman-style reasoning: this move's worth includes both its instant gain and the strength of the resulting position.
A standard test or yardstick used to measure and compare how well systems or models perform.
Benchmarks let you compare different tools fairly and track whether a model is actually improving.
Two document-search tools are run on the same set of test questions to see which returns more accurate, source-linked answers.
A standard, shared set of test data everyone uses so different models can be compared fairly on the same task.
Benchmarks let you compare AI tools apples-to-apples and track progress over time instead of trusting each vendor's own numbers.
Two document-extraction tools are both scored on the same benchmark dataset of 1,000 labeled contracts to see which is more accurate.
An influential language model that understands a word by looking at the words on both its left and right at the same time.
BERT made machines far better at understanding text meaning, powering search, classification, and question answering before today's chat models.
BERT can tell that 'bank' means a riverbank, not a financial bank, by reading the full sentence around it.
When an AI system systematically treats some groups of people unfairly, often because its training data reflected human prejudice.
Unfair AI can quietly discriminate in hiring, lending, or legal decisions, creating real harm and legal liability.
A resume screener trained on past hires keeps rejecting qualified women because the historical data favored men.
An extra adjustable number added inside a model that lets it shift its output up or down to fit the data better.
This purely mathematical 'bias' (unrelated to fairness) is a basic building block every neural network needs to learn well.
Like the constant in 'y = mx + b,' the bias term lets a model's line cross the axis where the data actually sits.
A method for reducing unfair bias that exists in training data or in a model's behavior.
Bias mitigation helps prevent AI from systematically disadvantaging certain groups, which matters for both fairness and legal compliance.
Before training a hiring model, an algorithm rebalances the data so one demographic group isn't underrepresented.
Using statistical tests to check whether a model's predictions differ unfairly across protected groups like race or sex.
Courts and regulators look at bias testing to judge whether a practice caused a discriminatory disparity rather than chance, making it key for legal defensibility.
A lender runs statistical significance tests to confirm its loan-approval model doesn't reject one group at a disproportionate rate.
A model that reads text using context from both before and after each word, not just what came earlier.
Reading in both directions gives a fuller understanding of meaning, since the right word often depends on what follows it.
To interpret 'served the papers,' a bidirectional model uses later words to decide whether 'served' means delivered legal documents or food.
A text model that predicts a word using the words on both sides of it rather than only the preceding ones.
It produces richer text understanding, which is why these models excel at filling in blanks and grasping context.
Given 'The witness ___ the testimony,' it weighs both the start and end of the sentence to fill the blank.
Very large, fast-growing, and varied datasets so big they need special scalable systems just to store and analyze them.
Big data is the raw fuel for modern AI, but its sheer volume and variety require tools built to handle scale.
Indexing millions of pages of case documents across thousands of matters so they can all be searched at once.
A pair of two words that appear next to each other, treated as a single unit when analyzing text.
Looking at word pairs captures simple phrases and order, giving more meaning than counting single words alone.
In 'expert report,' the bigram 'expert report' carries more meaning than 'expert' and 'report' counted separately.
A task where the model sorts each item into exactly one of two categories, like yes or no.
Many practical AI decisions are two-way splits, so this is one of the most common and useful types of model.
A model performs binary classification to label each email as 'privileged' or 'not privileged.'
A yes-or-no question inside a decision tree that splits the data into exactly two paths.
These simple two-way splits are the building blocks that make decision trees easy to follow and explain.
A tree sorting cases might ask the binary condition 'Is the claim over $1 million?' and branch yes or no.
Grouping a range of continuous numbers into a few labeled buckets, like turning exact ages into 'young,' 'middle,' and 'senior.'
Binning can simplify messy numeric data and help models spot patterns that exact values would hide.
Settlement amounts get binned into 'under $100K,' '$100K-$1M,' and 'over $1M' before analysis.
Personal data about a person's physical or behavioral traits, like a face scan or fingerprint, that can uniquely identify them.
Biometric data is highly sensitive and tightly regulated because it identifies a specific individual and can't be reset like a password.
A facial image used to unlock a phone or confirm someone's identity at a security checkpoint.
A score from 0 to 1 that measures how closely a machine's translation matches one written by a human.
BLEU gives a quick, automatic way to judge translation or text-generation quality without needing a human for every output.
A legal-document translator scoring high BLEU on a test set means its output closely mirrors the human reference translations.
A newer translation-quality score that judges meaning and fluency using a learned language model, not just word overlap.
It matches human judgment better than older overlap-based scores, catching good translations that use different but correct wording.
Two translations of a contract that share few exact words but mean the same thing can both score well on BLEURT.
A technique that builds many weak models one after another, each one focusing on the mistakes the previous ones made.
Boosting turns a crowd of simple, error-prone models into one highly accurate predictor and powers many top-performing tools.
Each new tree in a boosted model concentrates on the case files the earlier trees misclassified, steadily improving accuracy.
A rectangle drawn around an object in an image to mark exactly where it is.
Bounding boxes are how vision AI tells you not just what is in a picture but where, which is key for locating items on a page or scene.
An OCR system draws a bounding box around each signature block on a scanned contract before reading it.
The discipline of modeling, automating, running, measuring, and improving the flow of business activities to meet company goals.
BPM gives organizations a structured way to make their workflows faster, more consistent, and easier to optimize over time.
Mapping out every step of contract approval, then automating the routing and tracking how long each stage takes.
When personal information is accessed, exposed, or taken by someone who shouldn't have it, or used for a purpose it wasn't meant for.
A breach can expose sensitive data, trigger legal obligations, and damage trust, so preventing and detecting them is critical.
An unauthorized person gains access to a folder of confidential client records on a shared server.
A complex AI system that can describe, predict, recommend, and do limited reasoning across a wide set of related topics, sometimes beyond human expertise.
Broad AI sits between narrow single-task systems and the hypothetical general intelligence, marking a step toward more flexible, capable tools.
A system that can analyze, forecast, and advise across many connected areas of a business rather than doing just one narrow task.
A shortcut that lets math operations stretch a small array to match a bigger one so they can combine without manual copying.
It makes the numeric code behind AI shorter and faster, automatically handling differently shaped data.
Adding a single bonus value to every row of a spreadsheet at once, instead of writing it into each cell, is broadcasting.
Converting a continuous value into a category by assigning it to one of several ranges, or 'buckets.'
Bucketing helps models handle numbers that matter in ranges rather than exact amounts, and groups rare values together.
Document lengths get bucketed into 'short,' 'medium,' and 'long' so the model treats similar-sized files alike.
A reward paid to independent security researchers who find software vulnerabilities and report them to the system's owner.
Bug-bounty programs harness outside experts to find security holes before malicious attackers do.
A company pays a researcher who discovers and responsibly reports a flaw that could have let attackers read private files.
Test equipment or software embedded inside a system itself that checks whether the hardware or software is working properly.
Built-in tests let a system self-check and catch faults without needing separate external test gear, supporting reliability.
A device runs its own internal diagnostic at startup to confirm all components and data flows are functioning correctly.
A defined set of activities, steps, and resources arranged to accomplish a specific business goal.
Understanding a business process is the first step to improving or automating it, since you need to know the steps before you can streamline them.
The intake process for a new legal matter: collect documents, run conflict checks, open a file, and assign a team.
A defined condition or decision criterion that controls how a task runs or in what order steps happen within a process.
Business rules encode an organization's policies into systems so decisions are applied consistently and automatically.
A rule that says any contract over $50,000 must be routed to a senior partner for approval before it can be signed.
Comparing a device or measurement against a trusted standard to determine how far off it is.
Calibration ensures measurements and outputs are accurate and trustworthy by checking them against a known reference.
Checking a clock against an official time source to learn exactly how many milliseconds fast or slow it runs.
A post-processing adjustment that nudges a model's predicted probabilities so they match how often things actually happen.
A model that says '90% confident' should be right about 90% of the time; calibration makes those confidence scores trustworthy instead of just relative.
If a document-classifier marks 100 emails as '80% likely privileged' but only 50 truly are, a calibration layer corrects that gap.
The first stage of a recommender that quickly narrows millions of items down to a small shortlist worth scoring in detail.
You can't rank every item for every user, so candidate generation makes large-scale recommendation fast enough to be practical.
From a library of 500,000 case exhibits, the system first pulls a few hundred plausibly relevant ones before fine-ranking them.
A training shortcut that computes the loss using all the correct labels plus only a random sample of incorrect ones, instead of every possible label.
When there are millions of possible output classes, checking them all each step is too slow; sampling keeps training affordable.
Training a word-prediction model over a 100,000-word vocabulary by scoring the true word against a few hundred random others per step.
A measure of how much an entity, person, or organization can do and how well it can achieve its goals.
Knowing a system's capabilities and limits is essential for deciding what to trust it with and where humans still need to step in.
A document tool's capability might include accurate text search across PDFs but not legal advice.
A programming structure that picks one of several paths to follow based on the value of a single control expression, then continues on afterward.
Case statements let programs cleanly handle many possible options without long tangled chains of if/else checks.
Code that does one thing if a status is 'open', another if 'closed', and another if 'pending', based on a single value.
Information that falls into distinct named groups rather than being a measured number.
Models handle categories very differently from numbers, so knowing a field is categorical decides how you prepare it for training.
A 'document type' field with values like contract, deposition, or exhibit is categorical, unlike a page-count field.
A language model that predicts the next word using only the words that came before it, never peeking ahead.
This left-to-right design is what lets chatbots and writing assistants generate text one word at a time.
Given 'The witness testified that...', the model predicts the most likely next word based solely on what precedes it.
The center point of a cluster, found by averaging all the items grouped into that cluster.
Centroids are how clustering algorithms summarize and define each group of similar items.
After grouping similar contract clauses, each group's centroid represents its 'typical' clause.
A family of grouping methods that organize data around central points, assigning each item to its nearest center.
It's the most common and fastest way to discover natural groupings in unlabeled data.
Grouping thousands of emails into themes by repeatedly snapping each email to the nearest cluster center.
A prompting technique that asks the model to spell out its reasoning step by step before giving a final answer.
Working through intermediate steps makes the model far more accurate on math, logic, and multi-part questions.
Instead of asking 'Is this clause enforceable?', you ask the model to walk through each legal test and then conclude.
A back-and-forth conversation with an AI model where each new message can reference everything said earlier.
Chat is the most common way people interact with modern AI, letting them refine and follow up instead of asking one isolated question.
Asking a tool to summarize an exhibit, then following up with 'now list only the dates it mentions.'
A conversational program that talks back and forth with a user in natural language.
Chatbots are the most common way people interact with AI today, handling support, Q&A, and guided conversations at scale.
A support assistant that answers a user's typed questions about a product in plain language.
A saved snapshot of a model's learned settings at a particular point during training.
Checkpoints let you resume training after a crash, compare versions, or roll back to a better earlier state.
Saving the model every hour so a power outage doesn't waste a multi-day training run.
An ordered series of messages exchanged directly between systems, with no central controller running the show.
Choreography lets distributed systems coordinate by talking peer-to-peer, which can be more resilient than relying on one central coordinator.
Several microservices passing messages to complete an order, each reacting to events without a single boss service directing them.
Splitting a long document into smaller passages so it can be embedded, searched, and fed to a model.
Good chunking keeps related sentences together, which directly improves how accurately a tool can retrieve and cite the right passage.
A 300-page transcript is chunked into page- or paragraph-sized pieces before being indexed for search.
One of the possible categories a classification model can assign to an input.
Defining the classes is the foundation of any sorting task; everything the model learns is about telling them apart.
A document sorter might use the classes 'relevant,' 'privileged,' and 'irrelevant.'
A training set where one category vastly outnumbers the others.
Imbalance can trick a model into ignoring the rare category and still looking accurate, which is dangerous when the rare case is the one you care about.
Out of 100,000 reviewed documents only 200 are privileged, so a lazy model could mark everything 'not privileged' and still score 99.8%.
A learning task where the model's answer is one of a fixed set of categories, like 'yes/no' or 'sunny/cloudy/rainy'.
Classification powers many everyday AI tasks, from sorting spam to tagging documents by type.
Automatically labeling each incoming document as a contract, invoice, or letter.
A model that sorts each input into one of a fixed set of categories.
Classification powers a huge share of practical AI, from spam filters to document tagging to fraud detection.
A model that labels each incoming email as 'responsive' or 'non-responsive' to a discovery request.
The cutoff score above which a model's prediction counts as 'yes' and below which it counts as 'no.'
Moving the threshold trades off catching more true cases versus raising more false alarms, so it directly shapes a tool's behavior.
Setting the privilege-flagging cutoff at 0.7 instead of 0.5 to flag fewer documents but with higher confidence.
Another name for a model that assigns inputs to categories.
It's the everyday term people use for the most common kind of predictive model.
A spam classifier decides whether each message belongs in the inbox or the junk folder.
Capping values that fall outside a chosen range by forcing them back to the nearest allowed limit.
Clipping tames extreme outliers and runaway training updates that would otherwise destabilize a model.
Treating any contract value above $10 million as exactly $10 million so a few giant deals don't skew the analysis.
Google's specialized cloud hardware built to run machine learning math very fast.
TPUs make training and running large models dramatically quicker and cheaper than ordinary processors.
Training a large language model in hours on Cloud TPUs instead of weeks on regular servers.
Automatically grouping similar items together without being told in advance what the groups are.
Clustering reveals hidden structure in data, like natural topics or customer types, when you have no labels to start from.
Grouping tens of thousands of case documents into themes so reviewers see related material together.
When parts of a neural network become overly dependent on each other instead of each learning something useful on its own.
Co-adaptation is a form of overfitting that makes models brittle; techniques like dropout exist to prevent it.
Two neurons that only ever work as a pair fail when one is slightly perturbed by new data.
Automation that uses machine learning and reasoning to improve and expand its own analysis, even generating and testing new ideas on its own.
Cognitive automation goes beyond fixed scripts by adapting its own methods, allowing deeper and more flexible analysis over time.
An analytics system that, on its own, tries new approaches to dig deeper into data and proposes fresh hypotheses to test.
Complex systems designed to sense, understand, act on, and adapt to information in ways loosely comparable to the human brain.
Cognitive computing aims for systems that perceive and reason about the world, not just follow rigid programmed rules.
A system that reads incoming reports, understands their content, recommends an action, and gets better as it sees more examples.
Recommending items to a person based on what similar people liked, rather than on the items' contents.
It powers 'people who liked this also liked' suggestions and works even when you know little about the items themselves.
Suggesting a precedent case because attorneys with similar research patterns found it useful.
In a database table, a column is a set of values that are all the same type, like a single field running down the table.
Columns define the structure of tabular data, telling you what kind of information each field holds.
In a table of case records, one column holds every case's filing date.
The branch of AI that teaches computers to interpret and understand images and video the way people understand what they see.
It lets software read scanned documents, photos, and diagrams instead of only plain text, opening up huge amounts of visual evidence to automation.
A tool scans a photo of a damaged structure and automatically flags the cracked beams it detects.
When the real-world patterns a model learned slowly change over time, making its predictions less accurate.
A model that was great at launch can quietly degrade as the world shifts, so deployed models need monitoring and retraining.
A fraud detector trained on last year's scams misses new tactics that emerged this year.
A yes/no test at a branch point in a decision tree that decides which way an input goes next.
Conditions are the building blocks of decision trees and forests, the if/then questions that drive every prediction.
A node testing 'Is the document over 50 pages?' routes long and short documents down different branches.
Keeping data private so that only authorized people can see it and it is never disclosed to those who should not have it.
Sensitive case files and personal information must stay protected, and confidentiality is the rule that prevents unauthorized leaks.
A document tool restricts a privileged deposition transcript so only the assigned legal team can open it.
The tendency to look for, interpret, or favor information that confirms what you already believe.
It can creep into how data is collected, labeled, or evaluated, baking human prejudice into a model's behavior.
A reviewer who expects a model to perform well only spot-checks the cases where it agreed with them.
A simple table that shows how often a classifier was right and wrong, broken down by what it predicted versus the truth.
It reveals not just overall accuracy but exactly which kinds of mistakes a model makes, which a single score hides.
A table showing how many privileged documents were correctly flagged versus wrongly cleared, and vice versa.
A person's clear, freely given, and informed agreement to let their personal data be collected or used.
Using someone's data without proper consent can break privacy laws and erode trust, so genuine agreement is a legal and ethical foundation.
Before uploading a client's records into an AI tool, the firm confirms the client agreed to that processing.
Breaking a sentence into its grammatical building blocks, like noun phrases and verb phrases, and showing how they nest.
Understanding sentence structure helps AI grasp meaning in tasks like translation and question answering.
Splitting 'The defendant signed the contract' into its subject, verb, and object phrases.
An independent, fully working system that also serves as one piece inside a larger combined system of systems.
It explains how separate tools can each stand on their own yet team up to deliver a capability none could provide alone.
A query tool and a timeline tool each work by themselves but together act as parts of one unified case platform.
A rule that limits what content or combinations of values are allowed in a dataset.
Constraints keep data consistent and valid by blocking entries that do not fit the defined structure.
A case database rejects any record where the filing date is set before the case open date.
How well a test or measure actually captures the abstract idea it claims to be measuring.
If a metric does not truly reflect the concept it targets, conclusions drawn from it can be misleading.
Checking whether an AI 'reasoning quality' score really reflects sound reasoning rather than just answer length.
How completely and appropriately a measure covers all the skills or traits it is supposed to assess.
A test with weak content validity misses important parts of what it should measure, giving an incomplete picture.
A document-review benchmark that omits common file types would have poor content validity for real casework.
The circumstances, purpose, and point of view under which something is defined or used.
The same word or fact can mean different things depending on context, so understanding it is key to correct interpretation.
The word 'discovery' means something specific in a legal context versus a science context.
The maximum amount of text (in tokens) a model can consider at once.
If a case file is larger than the context window, the model cannot read it all in one go — which is exactly why retrieval (RAG) is used.
A model with a large context window can hold a full deposition transcript in memory while answering questions about it.
The real conditions a product is actually used in during normal day-to-day work.
A tool that performs well in testing can still fail in the messy real-world setting it is meant for, so its true context matters.
An OCR tool tested on clean scans may struggle in its real context of faded, handwritten documents.
A system knows enough about its own purpose to judge where its data comes from, how relevant it is, and how useful it is.
Systems that grasp context can use information more wisely instead of treating every input as equally important.
A research assistant recognizes that a recent ruling is more relevant to a current case than an outdated one.
A numeric representation of a word that changes depending on the surrounding words, capturing meaning in context.
It lets AI tell apart different senses of the same word, which fixed dictionaries of meaning cannot.
The word 'charge' gets a different embedding in 'criminal charge' than in 'charge the battery.'
The group in an experiment that does not get the treatment, used as a baseline to compare against the treated group.
Without a control group you cannot tell whether a change actually caused an effect or it would have happened anyway.
To test a new search algorithm, some users keep the old one as a control class for comparison.
The ability for a human or outside agent to step in and intervene in how a system operates.
Keeping a system controllable means people can correct or stop it, which is essential for safe and trustworthy AI.
A reviewer can pause an automated document classifier and override its decisions at any time.
The person or organization that decides why and how personal data gets processed.
The controller carries legal responsibility for protecting personal data, so knowing who it is matters for accountability.
A law firm that decides to load client records into an AI tool acts as the data controller for that data.
Collecting data from whatever is easiest to reach rather than from a representative slice of the real population.
Easy-to-grab data is often skewed, so a model trained on it can fail badly on the cases it never saw.
Training a document classifier only on cases from one law firm because those files were handy.
The point in training where the model stops improving meaningfully because its error has leveled off.
Knowing a model has converged tells you when to stop training, saving time and avoiding wasted computation.
After many training passes, the error barely changes from one pass to the next, signaling it's time to stop.
A curve shaped like a bowl, with a single lowest point and no misleading dips.
Convex problems are easy to optimize because any downhill path reaches the one true best answer.
Rolling a marble inside a smooth bowl, it always settles at the single bottom.
Finding the lowest point of a bowl-shaped problem using mathematical techniques that are guaranteed to reach the best answer.
When a problem is convex, you can trust the solution is truly optimal, not just a decent guess.
Tuning a simple linear model where there's exactly one best setting and the math always finds it.
A region where a straight line between any two points inside it stays entirely inside.
Convex sets are the well-behaved shapes that make optimization problems reliably solvable.
A filled circle is a convex set; a crescent moon is not, because a line across the gap leaves the shape.
A math operation that slides a small filter across an image to detect patterns like edges or textures.
Convolution is the core trick that lets AI recognize visual features no matter where they appear in a picture.
A filter sweeping across a scanned exhibit to highlight the edges of handwritten signatures.
The small grid of numbers slid across an image to pick out one specific visual pattern.
Each filter learns to spot a different feature, and stacking many of them is how vision models build up understanding.
One filter learns to detect horizontal lines, another detects corners, in a scanned document.
A layer of a neural network that applies many filters to its input to produce maps of detected features.
Stacking convolutional layers lets a network move from simple edges to complex objects like faces or letters.
An early layer finds strokes while a later layer combines them into recognizable characters on a page.
A neural network specialized for images that learns visual patterns by sliding filters across the picture.
CNNs power most image and document recognition, from reading scanned text to spotting objects in photos.
A CNN reads scanned deposition pages and converts the handwriting into searchable text.
The step-by-step process of multiplying a filter against each patch of an image and summing the results.
It's the exact arithmetic behind how vision models extract features, repeated across the whole image.
Computing one output number for every position as a filter slides across a scanned signature.
An AI-powered assistant built into an application that understands your requests and offers suggestions, summaries, and generated content.
Copilots speed up everyday work by helping right inside the tools people already use, instead of requiring a separate system.
A drafting copilot suggests the next clause as a paralegal types a contract.
A deliberately gathered collection of documents and data assembled because it likely holds information relevant to a topic.
AI tools draw their answers from a corpus, so the quality and coverage of that collection shapes what the tool can find.
All deposition transcripts and exhibits for a case form the corpus a query tool searches.
A measure of how two things tend to move or change together.
Spotting correlations helps reveal relationships in data, but a correlation alone does not prove one thing causes another.
Longer reports might correlate with more citations, without one necessarily causing the other.
Another word for loss, a number measuring how wrong a model's predictions are.
Training is essentially the process of pushing this number as low as possible.
A high cost means the model's privilege predictions are far off from the human-reviewed answers.
An explanation that says what would need to change in the inputs for a model to give a different result.
It makes AI decisions easier to understand and challenge by showing the smallest change that would flip the outcome.
A loan model explains: had income been ten thousand higher, the application would have been approved.
A fairness test asking: would the model's decision stay the same if we changed only a person's sensitive trait, like race or gender?
It checks whether protected characteristics secretly drive outcomes, a key safeguard against discrimination.
Verifying a loan model gives the same answer for two identical applicants who differ only in gender.
An action, device, or procedure that reduces a system's vulnerability to threats.
Countermeasures are the practical defenses that protect systems and data from attacks and failures.
Encrypting stored case files is a countermeasure against data theft if a server is breached.
When the data used to train a model leaves out part of the population it will later be used on.
If a group isn't represented in training, the model is likely to perform poorly or unfairly for that group.
A speech model trained only on adult voices struggles with children because they weren't covered.
A sentence whose wording is genuinely ambiguous, allowing two very different valid interpretations.
Crash blossoms show why language understanding is hard, since even correct grammar can mislead an AI.
A headline like 'Lawyer for defendant fails' can be read two completely different ways.
How well a measure's results line up with future outcomes or with another already-trusted measure.
It tells you whether a test can be relied on to predict or match the real thing it is meant to stand in for.
A document-relevance score has criterion validity if it agrees with how expert reviewers later rate the same documents.
The part of certain reinforcement learning systems that estimates how good a situation or action is, guiding the learner.
The critic's feedback helps the agent learn faster by judging moves rather than waiting for the final outcome.
In a game-playing agent, the critic scores each board position so the agent knows which moves look promising.
A way of measuring how far a model's predicted probabilities are from the true answers, used to score classification mistakes.
It's the standard yardstick that classification models try to minimize while training, punishing confident wrong answers heavily.
A model that confidently labels a privileged document 'not privileged' gets a large cross-entropy penalty.
A testing method that repeatedly trains on most of the data and checks on the held-out rest, rotating which part is held out.
It gives a more reliable estimate of how a model will perform on new data than a single train/test split.
Splitting 1,000 labeled documents into five rotating folds to see how consistently the model performs.
Inviting a large, varied group of people through an open call to voluntarily take on a task.
Crowdsourcing can label huge amounts of data or solve problems faster than a small team by pooling many contributors.
Many volunteers each tag a few scanned pages, together building a large labeled dataset.
A function that tells you the probability a value falls at or below a given level.
It's a basic statistics tool for understanding how data is spread and answering 'what fraction is below X?'
Reading off that 80% of reviewed documents are 20 pages or shorter.
The person who benefits from an automated task, process, or service being carried out.
Knowing who the customer is keeps a tool focused on delivering real value to the people it serves.
An attorney who receives a finished case summary from an AI tool is the customer of that service.
Protecting computers, networks, and the information they hold from damage, attack, or unauthorized access.
It keeps sensitive data and critical systems safe, available, and trustworthy against a constant stream of threats.
Multi-factor login on a case platform is a cybersecurity measure that blocks stolen passwords from granting access.
A user interface deliberately designed to trick or pressure people into choices they would not otherwise make.
Dark patterns undermine user autonomy and trust, and many are now restricted by regulation.
A signup flow that hides the cancel button while making 'subscribe' huge and bright is a dark pattern.
Facts or information, often numbers, gathered through observation.
Data is the raw material every AI system learns from and reasons over, so everything else builds on it.
The dates, names, and amounts pulled from a contract are the data an analysis tool works with.
Looking carefully at a dataset to understand its patterns, quality, and quirks before building anything with it.
You can't trust a model trained on data you never inspected; analysis catches gaps, errors, and bias up front.
Before training, you chart how many depositions per case exist and spot that one expert has almost no labeled examples.
Creating extra training examples by slightly altering existing ones, so the model sees more variety.
It cheaply expands a small dataset and helps a model generalize instead of memorizing the few examples it has.
Rotating and cropping the same scanned exhibit image many ways teaches a model to read it at any angle.
The process of finding and fixing or removing inaccurate, broken, or duplicate data records.
Clean data leads to reliable results, while messy data can quietly corrupt analysis and model outputs.
Removing duplicate exhibits and fixing garbled OCR text before loading documents into a search tool.
Management oversight of an organization's information policies, including watching how data processes run and handling issues.
Strong data control keeps information handled responsibly and helps catch problems before they cause harm.
A firm monitors who accesses client files and reports any unusual activity as part of its data control.
Testing so many hypotheses against a dataset that some appear statistically significant purely by chance.
It produces false 'findings' that look real but do not hold up, leading to misleading conclusions.
Running hundreds of comparisons on case outcomes until one random pattern looks meaningful is data dredging.
When the kind of data feeding a model gradually changes over time, causing its performance to slip.
A model trained on old data can quietly grow less accurate as the real world shifts, so drift must be watched.
A classifier trained on older document formats starts misreading newer e-filed layouts.
A connected collection of data that has been organized and tagged with meaning so AI can draw useful insights from it.
A well-built data fabric lets AI tools reason across many sources instead of treating each file as an isolated island.
Linking deposition transcripts, exhibits, and timelines so a tool can surface connections across the whole case.
Combining data from several different sources into one unified, more useful picture.
Merging multiple inputs often reveals patterns no single source could show on its own.
Blending witness statements, photos, and sensor logs into a single coherent account of an incident.
The rules and processes an organization uses to formally manage who controls its data and how it is used.
Clear governance keeps sensitive data secure, traceable, and compliant, which is essential when handling confidential case material.
Defining who may access privileged client documents and how long they are retained before deletion.
Using computers to sift through large amounts of data and uncover hidden patterns and relationships.
It turns raw piles of records into useful patterns that humans would never spot by reading manually.
Scanning thousands of contract clauses to find recurring language that signals a common risk.
Speeding up training by copying the model across many machines and giving each a different slice of the data.
It lets huge models train in hours instead of weeks by splitting the workload across many chips.
Eight GPUs each process a different batch of documents, then combine what they learned each step.
A single, discrete piece of information.
Data points are the building blocks of every dataset, chart, and AI prediction.
One date extracted from a contract is a single data point on a case timeline.
The cleanup and reformatting steps done early on so raw data is ready for analysis.
Models are only as good as their inputs, so careful preparation prevents garbage-in, garbage-out results.
Stripping headers and fixing OCR errors in scanned PDFs before feeding them to a search tool.
A stand-in measurement used when the thing you actually want to measure can't be observed directly.
Proxies let analysis proceed when direct data is missing, but a poorly chosen proxy can quietly distort conclusions.
Using a document's word count as a rough proxy for how complex it is to review.
How well data meets the needs it's being used for under the conditions it's used in.
Low-quality data leads to wrong answers, so quality directly determines how much you can trust an AI tool's output.
Checking that scanned exhibits are complete and legible before relying on them for analysis.
The discipline of pulling useful knowledge out of data through discovery and testing of ideas.
It's the field that turns raw data into the insights and models that power AI tools.
Studying years of case outcomes to learn which document features predict a settlement.
A professional who combines business sense, domain knowledge, and technical skill to manage data and build models from start to finish.
Data scientists translate messy real-world problems into models that actually answer the right questions.
Someone who builds and tunes the model that flags inconsistencies across expert reports.
Deliberately adding starting examples and known outcomes to data so a machine learning system has clear signals to learn from.
Good seed data helps a model learn the right relationships early instead of guessing from scratch.
Feeding a few labeled good and bad contract clauses so a tool can begin recognizing problem language.
A collection of examples gathered together to train or test a model.
The dataset is the raw material of machine learning; its size and quality cap how good the model can be.
A dataset of 5,000 labeled contract clauses used to teach a model to flag indemnity language.
The hands-on work of finding, extracting, cleaning, and combining data into a tidy set ready for analysis.
Most analysis time is spent wrangling, and skipping it leaves errors that corrupt every later step.
Pulling dates, names, and amounts out of inconsistent invoices into one clean spreadsheet.
Making choices based on what the data actually shows rather than on gut feeling or intuition.
Data-driven decisions are easier to defend and audit because they trace back to evidence instead of opinion.
Deciding which expert reports to review first based on flagged inconsistency counts rather than a hunch.
A table of data in memory, with named columns and rows, that code can easily filter and transform.
It is the everyday container teams use to clean and prepare data before feeding it to a model.
You load 10,000 exhibit records into a DataFrame and drop the rows missing a document date.
A TensorFlow toolkit for loading and feeding data into a model efficiently during training.
It keeps the expensive chips fed with data so training does not stall waiting on the disk.
You use tf.data to stream millions of document pages from storage in shuffled batches.
A conclusion reached after weighing the relevant rules and data in a given situation.
Decisions are the output that data and AI tools ultimately exist to support.
Concluding that a flagged document needs a closer manual review after seeing the tool's findings.
The dividing line a classifier draws to separate one category from another.
It shows exactly where the model flips its answer, which helps explain borderline or wrong predictions.
On a spam filter, the decision boundary is the score above which an email is called spam.
A model that combines the votes of many decision trees to make a more reliable prediction.
Forests are accurate, fast, and easy to explain, making them a workhorse for everyday tabular data.
A decision forest predicts which cases will settle by pooling hundreds of trees' opinions.
A moment in a process where the flow can branch into one of several possible paths.
Mapping decision points shows where human or automated judgment changes the course of a workflow.
A review step that routes a document to escalation only if a risk score is high.
A software tool that analyzes large amounts of data and presents the best options to help people decide.
It brings together scattered data and knowledge so people can make informed choices instead of relying on basic reports.
A dashboard that combines case documents and risk scores to help a team prioritize what to review.
The cutoff score at which a model's prediction switches from one class to the other.
Tuning the threshold trades false alarms against missed cases, letting you match the model to real-world stakes.
Lowering the threshold makes a document classifier flag more pages as privileged, catching more but with more false hits.
A model that makes a prediction by following a branching series of yes/no questions.
Its logic reads like a flowchart, so it is one of the easiest models to inspect and trust.
A tree decides whether a clause is risky by asking, in turn, about its category, length, and key words.
The thinking process of choosing one belief or course of action from several options.
Understanding how decisions get made helps you place AI as a support rather than a replacement for human judgment.
A reviewer weighing a tool's suggestions against case facts before choosing how to proceed.
The part of a model that turns an internal representation into output, like generating text word by word.
The decoder is what actually produces a chatbot's or summarizer's written answer.
After reading a deposition, the decoder writes out the summary one token at a time.
A language model built only from the text-generating half of the architecture, used by most modern chatbots.
This is the design behind most popular LLMs, so understanding it explains how today's chat assistants work.
A decoder-only model reads your prompt and continues it, generating an answer word by word.
Safely removing a system or component from active use while controlling risk and meeting policy rules.
Properly retiring old systems avoids security gaps, compliance failures, and wasted cost.
Taking down an outdated document-processing service after migrating its data to a newer one.
Analysis that explores what would likely happen if a particular future event occurred.
It helps teams anticipate the consequences of choices before committing to them.
Estimating how a case timeline would shift if a key deposition were delayed by a month.
Machine learning that uses neural networks with many layers to learn complex patterns.
Deep learning powers today's language and vision models. It is what makes reading messy scanned pages and natural language possible.
The model that turns a blurry scanned exhibit into searchable text is built on deep learning.
A neural network with many stacked layers, letting it learn complex patterns.
Depth is what lets these models handle hard tasks like language and images that simpler models can't.
A deep model with dozens of layers reads scanned reports and understands their meaning, not just their words.
Another name for a neural network with many layers between input and output.
It is the engine behind modern AI, from speech recognition to document understanding.
A deep neural network converts a photographed signature page into searchable text.
A method that uses a neural network to learn the best action to take in each situation through trial and error.
It powered breakthroughs in game-playing AI and is a foundation of modern reinforcement learning.
A DQN learns to play a video game well just by trying moves and tracking which led to higher scores.
AI-generated or altered image, audio, or video that looks real but is actually fabricated.
Deepfakes can be used to forge evidence or impersonate people, making authenticity verification critical.
A faked audio clip that seems to capture a witness saying something they never actually said.
The act of destroying a specific piece of data or a created item so it no longer exists.
Reliable deletion is key to privacy, retention limits, and honoring requests to remove personal data.
Permanently removing a client's uploaded documents once a case is closed and retention expires.
A fairness goal where a model approves the same share of people from each group, regardless of group.
It is one common way to check that an automated decision is not systematically favoring one group over another.
A screening model meets demographic parity if it flags the same percentage of applicants in every demographic.
An attack or failure that blocks legitimate users from accessing a resource or delays time-critical operations.
A successful denial-of-service can take a tool offline at a critical moment, so defenses are part of keeping systems available.
Flooding a document-processing service with bogus requests so real users can't upload their files.
A training trick where a model learns by removing deliberately added noise from corrupted examples.
It lets a model learn useful structure from unlabeled data, and underpins how image generators work.
A model trained to clean up blurred scans of old documents learns what clean text should look like.
An input value that is usually present and non-zero, like a price or an age.
Most real measurements are dense, and models handle them differently from mostly-empty sparse features.
A document's page count is a dense feature because every document has one.
A neural network layer where every input connects to every output node.
It is the most common building block of neural networks, mixing information from all inputs together.
A dense layer combines all extracted clues from a contract to score how risky it is.
A system's overall ability to perform as needed, covering availability, reliability, and maintainability.
Dependable systems can be trusted to work when it counts, which matters for tools relied on under deadline pressure.
A case tool that stays available and produces consistent results every time it's needed.
The phase where a finished system is put into real operation and any switchover issues are resolved.
Deployment is when a model or tool finally delivers value to users, and it's where many practical problems surface.
Releasing a tested document-search tool so the whole case team can start using it live.
The number of layers a neural network has stacked between its input and output.
More depth lets a model learn richer patterns, but also makes it slower and harder to train.
Increasing a model's depth from 4 to 40 layers helps it understand legal phrasing more reliably.
A faster, lighter way of doing the image-scanning math inside vision models by splitting it into two steps.
It lets accurate image models run on phones and other small devices without heavy hardware.
A mobile app uses this technique to read a document photo quickly on the phone itself.
Analysis that explains why something happened by examining its causes.
Understanding the cause of past events is the foundation for predicting and improving future ones.
Looking back at why a batch of documents failed OCR to find the underlying scan-quality issue.
Behaving in a fixed way so the same inputs always produce the same outcome, with no randomness involved.
Deterministic behavior makes results repeatable and easy to verify, which is reassuring when accuracy matters.
A formula that totals invoice amounts gives the identical sum every time it's run on the same data.
A step-by-step procedure that always returns the same output for the same input.
Because its results never vary, you can test it once and trust it to behave the same way every time.
A sorting routine that arranges the same list of names in the exact same order on every run.
Anyone who builds systems, components, or services, whether in-house or as an outside vendor.
Knowing who developed a system helps trace responsibility, support, and accountability for how it works.
The team that built and maintains a case management tool used across a firm.
The piece of hardware, like a CPU, GPU, or TPU, that actually runs the model's calculations.
Choosing the right device controls how fast and how cheaply a model trains or answers.
Moving training from a CPU to a GPU device cuts the time to fine-tune a model from days to hours.
Analysis that digs into data to explain why something happened.
Knowing the cause behind a trend or failure lets teams fix the real problem instead of guessing.
After a contract review tool flags more errors one month, diagnostic analytics traces it to a batch of poorly scanned documents.
Tools and checks that detect and pinpoint faults or failures in a system.
Good diagnostics tell you exactly what broke and where, so problems get fixed faster.
When a document-processing pipeline stalls, diagnostics reveal that the OCR step timed out on a corrupted PDF.
A technique that adds carefully measured noise so a model can't leak details about any single person in its training data.
It lets organizations learn from sensitive records while provably protecting individuals' confidentiality.
A model trained with differential privacy on case files cannot be probed to reveal any one client's information.
When a test or model predicts outcomes accurately for one group of people but not equally well for another.
If a model's accuracy varies by group, decisions based on it can be unfair to the group it serves worse.
A resume-screening model that predicts job success well for one applicant population but poorly for another shows differential validity.
A type of generative model that learns to create content by starting from random noise and gradually cleaning it into a finished result.
Diffusion models power many of today's high-quality image and media generators and are a major branch of generative AI.
A diffusion model can turn a text prompt into a detailed illustration by refining random noise step by step.
Software and automation doing work that a person would otherwise do.
Digital labor frees people from repetitive tasks so they can focus on judgment-heavy work.
A tool that reads incoming case documents and sorts them by type is performing digital labor.
The whole set of automation tools a business uses to do work alongside or instead of people.
Treating bots and automations as a managed workforce helps organizations scale work without hiring for every task.
A firm's digital workforce includes the bots that extract dates, summarize reports, and route filings.
Shrinking data with many features down to fewer, while keeping the important information.
It speeds up models, saves storage, and makes high-dimensional data easier to visualize and search.
Compressing each document's 1,000-number fingerprint down to 50 still keeps similar documents close together.
The count of numbers used to describe each example, such as the length of an embedding.
More dimensions can capture more nuance but cost more memory and compute to store and compare.
An embedding with 768 dimensions represents each paragraph as a list of 768 numbers.
An input that can only take a limited set of distinct values, not a smooth range.
Models treat discrete inputs like categories differently from continuous numbers, so labeling them right matters.
A document's type, such as motion, exhibit, or transcript, is a discrete feature.
A model that learns to tell categories apart rather than to create new examples.
Most classifiers are discriminative; knowing this distinguishes them from generative models that produce content.
A discriminative model decides whether a clause is privileged, but cannot write a new clause.
In a paired training setup, the model whose job is to tell real examples from fake generated ones.
Its feedback pushes a generator to produce ever more realistic output, the core idea behind GANs.
A discriminator learns to spot AI-forged signatures, forcing the generator to make them more convincing.
When a decision process unintentionally harms one group more than others, even without targeting them.
It is a key legal and ethical fairness concern because harm can occur without any intent to discriminate.
A hiring model trained on past data shows disparate impact if it rejects far more of one group.
When a model uses a sensitive attribute, like race or gender, directly in making a decision.
Unlike accidental harm, this is intentional use of a protected trait, which is usually prohibited outright.
A model commits disparate treatment if it factors an applicant's gender into a loan decision.
Training a small, fast model to copy the behavior of a large, slow one.
It captures most of a big model's skill in a cheaper package that is practical to deploy.
A compact model distilled from a large LLM summarizes documents almost as well but runs far cheaper.
The pattern of how often each value or outcome shows up across a dataset.
Knowing the distribution reveals what is typical versus rare, which shapes how a model behaves.
The distribution of document lengths shows most are short, with a few very long transcripts.
Building a model so it stays accurate across many kinds of input data, not just one expected kind.
Real-world data rarely matches training data exactly, so robust models avoid failing on unfamiliar inputs.
A document classifier trained for clean scans that still works on faxed or photographed pages has distributional robustness.
When the data a model sees in the real world differs from the data it was trained on.
It quietly degrades accuracy after deployment, so it is a top reason models stop working over time.
A model trained on 2010s contracts performs worse on today's documents written in a newer style.
Including the wide range of communities, identities, backgrounds, and abilities of people.
Building AI with diverse perspectives and data helps it serve everyone fairly rather than just a narrow group.
Gathering training documents from many regions and writing styles improves how a tool handles diverse users.
Grouping data by starting with one big cluster and repeatedly splitting it into smaller ones.
It organizes unlabeled data into a tree of groups, useful for exploring how items relate.
Divisive clustering splits all case documents into broad themes, then into finer subtopics.
Written or visual records that describe, specify, or report how something works or what was done.
Clear documentation lets others understand, audit, and reproduce a system or process.
A model card documenting an AI's training data and known limits is a key piece of documentation.
A defined subject area or scope where common rules and characteristics apply.
Knowing a system's domain sets expectations for what it can handle well and where it may fail.
A model trained on the legal domain understands court filings but may struggle with medical records.
Deep knowledge of the essential details of a specific field.
Domain expertise helps people judge whether an AI's output is correct and meaningful in a real field.
A forensic engineer's domain expertise lets them spot when an AI summary of an inspection report misses a critical defect.
When the data a model sees in use differs from the data it was trained on.
Domain shift quietly degrades accuracy because the model meets patterns it never learned.
A tool trained on typed reports loses accuracy when fed handwritten field notes, a classic domain shift.
Reducing data, either by shrinking images or by keeping fewer examples of an overrepresented class.
It balances lopsided datasets and lowers compute cost, helping models train fairly and faster.
With far more ordinary documents than privileged ones, you downsample the ordinary ones for balance.
When a company's own staff use the product they build to find what works and what needs fixing.
Using your own product surfaces real flaws early, before customers hit them.
A team that runs its own case files through its document-search tool catches bugs before clients do.
A training method that randomly ignores some neurons each step so the model can't over-rely on any one.
It is a simple, powerful way to stop a model from memorizing its training data instead of learning general rules.
Adding dropout keeps a document classifier from fixating on one rare word and missing the big picture.
Describes anything that updates frequently or happens live, as opposed to being fixed ahead of time.
Telling dynamic from static behavior clarifies whether a system reacts in real time or runs on a fixed snapshot.
A dynamic pipeline re-scores documents the moment they are uploaded rather than in a nightly batch.
A model that keeps learning and updating continuously as new data arrives.
It stays current with changing conditions instead of going stale like a model trained once and frozen.
A dynamic model updates nightly as new filings come in, so it adapts to evolving legal language.
A process that can follow different paths depending on the conditions at the moment it runs.
Dynamic processes adapt to each situation instead of forcing every case down one fixed route.
A document intake workflow that routes contracts and depositions to different reviewers is a dynamic process.
A programming mode where each operation runs immediately as it is written, instead of being collected into a plan to run later.
It makes machine learning code easier to write and debug because you see results step by step, like normal programming.
You add two numbers and the answer appears right away, rather than the system waiting until the whole program is assembled.
A technique that halts model training as soon as the model stops improving on held-out data, before it starts memorizing.
It prevents wasted time and overfitting by stopping at the point where the model is best at handling new, unseen data.
Training a contract-classifier stops automatically once accuracy on a validation set of unseen contracts plateaus.
A way to measure how different two distributions are by calculating the least effort needed to reshape one into the other.
It gives a meaningful distance even when two distributions barely overlap, which simpler measures fail to do.
Comparing how the spread of document lengths in one case file differs from another by computing the minimum work to match them.
A rare or extreme situation that only happens at the limits of what a system normally handles.
Edge cases often expose bugs or failures that ordinary testing misses.
A date-extraction tool that handles normal dates but breaks on "the 31st of February" hit an edge case.
A count of the smallest number of single-character changes (insert, delete, or substitute) needed to turn one string into another.
It measures how similar two pieces of text are, useful for spell-checking, fuzzy matching, and catching typos in names.
Matching a witness name spelled "Jon Smyth" in one exhibit to "John Smith" in another despite the small differences.
Special numbers that describe how much a mathematical transformation stretches or shrinks along its key directions.
They underpin dimensionality reduction and help reveal the most important patterns hidden in large datasets.
Finding which directions in a customer dataset carry the most variation so you can summarize it with fewer numbers.
A direction that stays pointing the same way (only stretched or shrunk) when a transformation is applied to it.
Eigenvectors identify the natural axes of variation in data, which is the basis of techniques like PCA.
Identifying the main axis along which a cloud of data points is spread out so it can be described compactly.
A list of numbers that captures the meaning of a piece of text so a computer can compare meanings.
Embeddings let search find passages by meaning, not just exact keywords — so 'who paid the invoice' finds 'remittance was made by...'.
Two paragraphs about the same event get similar embeddings even if they share no words.
The multi-dimensional space where each item is placed as a point of numbers so that similar items sit close together.
It is what lets search and AI tools group things by meaning, so related documents cluster near each other.
Deposition passages about the same incident land near one another in the embedding space even with different wording.
The actual list of numbers that represents one item's meaning inside the embedding space.
This number list is what the computer compares to decide whether two pieces of text mean the same thing.
Each exhibit gets its own embedding vector so the system can find the closest matching passages to a question.
A skill a model was never directly trained for that suddenly appears once the model gets big enough.
It explains why larger models can do new things — like multi-step reasoning or translation — without anyone adding that feature. It also means abilities are hard to predict in advance.
A small model can only echo phrases, but a much larger version of the same model starts correctly extracting dates and parties from a contract clause, even though no one wrote a rule for it.
Choosing the model that makes the fewest mistakes on the training examples you actually have.
It is the core idea behind how most models learn, but relying only on it can lead to memorizing rather than generalizing.
A model tuned purely to label every training contract correctly, which may still stumble on new contracts it has not seen.
Using one system to imitate another so it accepts the same inputs and produces the same results.
Emulation lets you test or run software in an environment that copies the real one without needing the original system.
Developers emulate a production server on a laptop to test a document pipeline before deploying it.
The part of a model that reads input and compresses it into a numeric summary capturing its meaning.
Encoders turn raw text or images into a form the rest of the system can reason over, powering translation and search.
An encoder reads a full deposition answer and produces a compact representation the model uses to retrieve related testimony.
The activity or output that marks the conclusion of a process.
Clearly defining where a process ends keeps workflows from running open-ended or stalling.
In a review workflow, delivering the finished report to the client is the end event.
A person trained to design, build, or manage technical systems and machines.
Engineers turn ideas and requirements into working systems that people can actually use.
A software engineer builds the pipeline that turns scanned documents into searchable text.
A group of separate models whose predictions are combined to produce a single, usually more accurate, answer.
Combining several models smooths out individual mistakes, often beating any single model on its own.
Three different document-classifiers vote on whether a filing is privileged, and the majority decision is used.
In reinforcement learning, the world the AI acts in, which responds to each action with a new situation and a reward.
The environment defines the rules of the game the agent is trying to master, shaping everything it can learn.
A chess board and its rules form the environment in which a game-playing agent makes moves and gets feedback.
One complete run of an agent interacting with its environment, from the starting point until it reaches an end.
Episodes are the repeated trials through which a reinforcement learning agent practices and improves.
A single full game of chess, start to checkmate, counts as one episode of training.
One full pass of the model over the entire training dataset during learning.
The number of epochs controls how many times a model studies the data; too few underfits, too many overfits.
Training a classifier for 10 epochs means it reviews the whole labeled document set 10 times.
A strategy where an agent mostly picks the action it currently thinks is best but occasionally tries a random one.
That occasional randomness lets the agent discover better options instead of getting stuck on a so-so choice.
A recommendation agent usually suggests the top-rated item but 5 percent of the time tries something new to learn more.
A fairness goal met when a model is equally good at correctly approving qualified people across different groups.
It checks that people who deserve a positive outcome get it at the same rate no matter their group.
A loan model approves equally qualified applicants at the same rate whether they are in one demographic group or another.
A fairness goal met when a model has the same true-positive and false-positive rates across different groups.
It is a stricter fairness check than equality of opportunity, balancing both correct approvals and wrongful approvals.
A screening model wrongly flags and correctly clears applicants at equal rates across all demographic groups.
The difference between a measured value and its true value, which can be random or systematic.
Understanding error helps you know how much to trust a measurement or prediction.
If an OCR tool reads a page count as 48 when it is really 50, that gap is the error.
How uncertainty in the inputs spreads into uncertainty in the final calculated result.
Small input errors can compound into large output errors, so tracking propagation reveals how reliable a result really is.
If each extracted figure in a report has a slight error, error propagation tells you how shaky the final total is.
A TensorFlow component that bundles training, evaluation, and prediction behind a simple, consistent interface.
It once gave developers a standard, reusable way to build models without wiring every step by hand.
Using a prebuilt estimator to train a classifier with a few lines instead of coding the full training loop.
The practice of designing and using AI in ways that respect moral principles like fairness, accountability, and avoiding harm.
Ethical AI keeps powerful systems aligned with human values so they help people without causing unfair or harmful outcomes.
An ethical approach to a document-AI tool ensures it protects confidential data and does not produce biased results.
A set of moral principles about what is right and wrong that guides how people and groups act.
Ethics shapes whether AI is built and used in ways that are fair, honest, and respectful of people.
Deciding not to deploy a tool that could expose private case details reflects ethics in practice.
Building ethical values into a technology right from its design and development stage.
Baking ethics in early prevents harmful designs instead of trying to patch problems after launch.
Designing a document tool to redact personal data by default is ethics by design.
The ordinary straight-line distance between two points, measured the way you would with a ruler.
It is a common way to judge how far apart two data points or embeddings are, driving clustering and nearest-neighbor search.
Measuring how close two document embeddings are by the straight-line gap between their positions.
The process of measuring how well a trained model performs, usually on data it did not learn from.
Evaluation tells you whether a model is actually good enough to trust before you rely on it for real decisions.
Testing a contract-review model on 200 unseen contracts to see how often it flags the right clauses.
One row of data the model learns from, made up of its input features and (in supervised learning) the correct answer.
Examples are the raw material of machine learning; the quality and quantity of them shape what a model can learn.
One labeled deposition excerpt marked "relevant" serves as a single training example for a relevance classifier.
An unexpected event during a process that diverts it from its normal flow.
Handling exceptions gracefully keeps a system from crashing or producing wrong results when something unusual happens.
When a document upload contains an unsupported file type, the tool raises an exception instead of silently failing.
To carry out a plan, command, or instruction.
Executing instructions is the basic action that turns code or a plan into actual results.
When you click Run, the tool executes the steps that extract every date from a case file.
A technique where an agent stores its past experiences and re-learns from random samples of them later.
Reusing old experiences makes reinforcement learning more stable and data-efficient instead of learning once and forgetting.
A game-playing agent saves thousands of past moves and replays random batches to keep improving from them.
A controlled test that changes one factor and measures the effect to learn what causes what.
Experiments give evidence about whether a change actually helps, instead of relying on hunches.
Running an A/B test to see whether a new prompt improves answer quality is an experiment.
An older form of AI that mimics a human expert by combining stored knowledge with hand-written rules.
Expert systems showed how AI could automate specialized decisions, and they paved the way for today's learning-based models.
A rule-based system that diagnoses equipment faults from a fixed checklist is an expert system.
Specialized knowledge built up through experience or deliberate mastery of a field.
Expertise is what lets a person reliably judge quality and make sound decisions in their field.
An engineer's expertise lets them confirm whether an AI-generated inspection summary is technically sound.
The ability to give a human-understandable reason for why a model made a particular prediction.
Explainability lets people trust, check, and challenge AI decisions instead of accepting them blindly.
A tool that shows which passages led it to flag a document gives users explainability.
AI built so that people can understand the reasons behind its decisions instead of treating it as a black box.
When users can see why an AI reached a conclusion, they can trust it, challenge it, and meet accountability requirements.
An explainable document-AI tool shows which passages led to its answer so a reviewer can judge whether to rely on it.
A feature that gives details about, or reasons behind, the results of a fairness metric.
An explainer helps people understand why a fairness score came out as it did, so they can act on it.
After a fairness check flags a model, the explainer shows which feature drove the disparity.
When a system provides supporting evidence or the reasoning behind every answer it gives.
Explanations let people check whether an AI's output is trustworthy instead of accepting it blindly.
A document tool that not only answers a question but shows the exact passages and reasoning it used to get there.
A training failure where the numbers guiding learning grow uncontrollably large, making the model unstable.
It can wreck training of deep or sequence models, which is why techniques like gradient clipping exist to tame it.
A long-sequence model's update values blow up to huge numbers and training produces meaningless results until clipped.
The balance between trying new actions to learn more (explore) versus repeating known good actions to win now (exploit).
Lean too far either way and an agent either never finds better options or never cashes in on what it knows.
A recommendation agent must decide whether to suggest a proven favorite or test an unfamiliar item to learn its appeal.
An early, mostly visual look at a dataset to understand its structure, spot odd values, and form ideas before building a model.
It surfaces problems and patterns up front so you do not waste effort modeling data you do not actually understand.
Charting the dates across thousands of case documents to spot gaps, duplicates, or outliers before any analysis.
How well results from one test sample are likely to hold true for other people, settings, or time periods.
If a model was only tested on a narrow group, its results may not transfer to the real-world cases you care about.
A tool tuned only on construction-defect reports may not perform as well on medical-injury reports.
A single score that balances how often a model is right when it says 'yes' (precision) against how many of the real 'yes' cases it catches (recall).
F1 gives one fair number for lopsided problems where simply counting correct answers would be misleading.
A tool that flags privileged documents reports an F1 of 0.88, telling you it both catches most privileged files and rarely false-flags ordinary ones.
Technology that turns a face in a photo into a numeric fingerprint and compares it against another to judge if they are the same person.
It powers identity checks and surveillance, but its accuracy and fairness vary, which raises real privacy and bias concerns.
Comparing a surveillance still against an ID photo to estimate whether both show the same individual.
How well an AI's statements match real, verifiable truth rather than made-up content.
Low factuality means the system sounds confident while being wrong, which is dangerous in legal or business decisions.
An assistant scoring high on factuality cites the actual deposition date instead of inventing one.
A rule added during training that forces a model to treat different groups of people more evenly.
It is a concrete lever to stop a model from quietly favoring or penalizing one group.
A hiring screen is trained with a fairness constraint so that qualified candidates are passed at similar rates across genders.
A measurable definition of what counts as 'fair' so you can check whether a model treats groups equitably.
You cannot prove a model is fair without first agreeing on a number that defines fairness.
Auditors pick a fairness metric, then test whether a loan model approves applicants at comparable rates across neighborhoods.
A case the model wrongly labels as 'no' when the real answer was 'yes'.
False negatives are the things you missed, which can be the most costly errors when overlooking something matters.
A document-review model marks a key responsive email as 'not relevant,' so it never reaches the attorney.
Of all the cases that were truly 'yes', the share the model wrongly called 'no'.
It tells you how much real, important material your model is letting slip through.
A privilege detector with a 5% false negative rate misses about 1 in 20 truly privileged documents.
A case the model wrongly labels as 'yes' when the real answer was 'no'.
False positives are false alarms that waste reviewer time and erode trust in the tool.
A relevance model flags hundreds of ordinary calendar invites as key evidence, burying reviewers in noise.
Of all the cases that were truly 'no', the share the model wrongly called 'yes'.
It measures the false-alarm load, which is the x-axis of the ROC curve used to judge classifiers.
A high false positive rate means an exhibit flagger keeps marking irrelevant files, forcing extra manual review.
A system's ability to keep working normally even when some hardware or software parts fail.
Fault tolerance keeps a service running through glitches so users are not interrupted by every small failure.
A document platform that stays available even if one of its servers crashes mid-upload.
A predicted outcome that benefits the person receiving it, as opposed to an unfavorable one that disadvantages them.
Fairness checks often measure whether favorable outcomes are spread evenly across different groups.
In a loan model, an 'approved' label is favorable while a 'denied' label is unfavorable.
A single input variable the model uses to make a prediction, like one column in a spreadsheet.
Features are the raw ingredients of any model; better features usually beat fancier algorithms.
To predict case duration, useful features include claim type, court, and number of parties.
A new combined input made by multiplying or pairing two existing inputs to capture how they interact.
It lets simple models learn relationships that no single feature could express on its own.
Crossing 'document type' with 'sender role' helps a model notice that contracts from outside counsel matter most.
The work of turning raw data into clean, useful inputs that help a model learn.
Most real-world model quality comes from good feature engineering, not from the choice of algorithm.
Extracting the signing date and dollar amount from a contract PDF turns messy text into features a model can use.
Automatically pulling meaningful signals out of raw data such as text, images, or audio.
It converts unstructured material into numbers a model can actually work with.
Feature extraction reads a scanned exhibit and produces the key entities and dates it contains.
A ranking that shows which inputs influenced a model's predictions the most.
It explains what drives a model's decisions, which is essential for trust and for spotting bad signals.
Feature importances reveal that 'document age' weighed far more than 'file size' when ranking relevance.
The full group of input variables a model is given to learn from.
Choosing the right feature set determines what patterns a model can possibly discover.
A case-outcome model uses a feature set of jurisdiction, claim amount, and filing year.
When the behavior of one specific input variable changes over time, even if the overall data looks unchanged.
Pinpointing which feature drifted helps you find the real cause of degrading model performance instead of guessing.
A fraud model where one sensor's readings quietly change while all other inputs look normal.
A definition that describes each input's name and data type so a pipeline can read raw data correctly.
It keeps training and live prediction reading data the same way, preventing silent mismatches.
A feature spec declares that 'filing_date' is a date and 'amount' is a number before the model ingests new records.
The list of numbers representing all of one example's features, fed into the model at once.
Models do math on numbers, so every example must be expressed as a feature vector first.
One contract becomes a feature vector like [2021, 50000, 1] capturing its year, value, and type.
A way to train one shared model across many devices or sites without moving the raw data off each one.
It lets organizations learn from sensitive data while keeping that data private and on-premises.
Several law firms jointly improve a clause-detection model while each firm's confidential contracts never leave its servers.
When a model's own predictions influence the future data it later learns from.
Unnoticed feedback loops can amplify mistakes or bias over time as the model trains on its own outputs.
If reviewers only check documents the model flags, the model never learns from the files it wrongly ignored.
A basic neural network where information flows straight from input to output with no loops back.
It is the foundational building block inside larger systems, including the layers within transformers.
A feedforward network takes a document's numeric features and outputs a single relevance score.
Teaching a model to handle a task after showing it only a handful of examples.
It lets you adapt powerful models to a new task quickly without gathering huge labeled datasets.
Show the model three sample privilege-log entries and it formats the rest of the case the same way.
Including a few worked examples directly inside your prompt to show the model exactly what you want.
A few good examples in the prompt often steer output far more reliably than a long instruction alone.
Paste two correctly summarized depositions before asking the model to summarize a third in the same style.
Further training a pre-built model on your own examples so it adapts to a specific task or style.
Fine-tuning can specialize a model, but for most document Q&A, retrieval (RAG) is cheaper and easier to keep up to date.
Fine-tune a model on past expert reports so its drafts match your firm's preferred structure and tone.
Software permanently stored on a hardware chip that cannot be changed while the device is running.
Firmware controls the basic behavior of devices, so its security and reliability underpin everything running on top.
The built-in code on a scanner that controls how it captures and feeds document pages.
Checking whether a data value falls inside a previously defined acceptable range.
It is a basic quality check that flags values which fall outside expected limits before they cause problems downstream.
Verifying that an extracted date from a document falls within the case's valid time window.
Estimating future conditions or events using what is known about past and expected performance.
Good forecasts let teams plan timelines and resources around likely outcomes rather than surprises.
Projecting how long document review will take based on current throughput and remaining volume.
The part of an LSTM memory cell that decides how much old information to throw away at each step.
It lets a sequence model keep what still matters and drop what no longer does, handling long text better.
While reading a long deposition, the forget gate lets the model discard an earlier irrelevant tangent.
A very large model trained on broad data that can be adapted to many different downstream tasks.
Foundation models are the shared base that powers most modern AI tools instead of building each one from zero.
One foundation model is adapted to draft contracts, summarize depositions, and answer case questions.
Watching how groups of users behave in order to spot or prevent activity that is dishonest or harmful.
It protects organizations and customers by catching suspicious patterns before they cause real losses.
Flagging an account whose transaction pattern suddenly looks nothing like its normal history.
A calculation that turns a model's raw scores into probabilities across every possible class, considering all of them.
It is accurate but slow when there are millions of classes, which is why faster approximations exist.
Picking the next word from a 50,000-word vocabulary with full softmax computes a probability for every single word.
A system that completes its assigned task within set limits without any human stepping in, adjusting to conditions on its own.
Full autonomy can scale work without supervision, but it also raises the stakes for safety and oversight when things go wrong.
An agent that gathers, sorts, and files incoming case documents end to end without a person guiding each step.
A neural network layer where every input connects to every output, mixing all the information together.
It is the workhorse layer that combines learned signals into a final decision in most networks.
After spotting features in a scanned page, a fully connected layer combines them into a single document-type label.
A feature that lets a language model ask to run an outside tool or API — like a lookup, a calculation, or a database query — and then use the result in its answer.
It lets a model reach real, current information and take actions instead of relying only on what it memorized during training, which is what turns a chatbot into a useful assistant or agent.
Asked for today's filing deadline, the model could call a calendar function to fetch the real date instead of guessing from old training data.
A setup where two AI models compete — one creates fake data and the other tries to spot the fakes — until the fakes become convincing.
GANs were a breakthrough for generating realistic images, audio, and synthetic data, and they explain how AI can fabricate lifelike content.
A GAN can generate a photorealistic face of a person who does not exist, which is also why deepfakes are a concern in evidence review.
Google's family of advanced AI models that can work with text, images, audio, and video together.
Gemini is one of the major frontier models businesses use to power chatbots, document analysis, and AI assistants.
A firm might use Gemini to summarize a long contract and answer follow-up questions about its clauses.
Google's family of lightweight, openly available AI models that anyone can download and run themselves.
Open models like Gemma let organizations run AI on their own servers for privacy and cost control instead of sending data to an outside service.
A legal team concerned about confidentiality could run Gemma on internal hardware to draft document summaries without data leaving the building.
A model's ability to perform well on brand-new data it has never seen before, not just the examples it was trained on.
Generalization is the whole point of machine learning — a model that only works on its training data is useless in the real world.
A document classifier that was trained on last year's contracts should still correctly sort this year's contracts if it generalizes well.
A chart showing how a model's error on training data and on new test data changes as training continues.
It is the main tool for spotting overfitting — when training error keeps dropping but test error starts rising, the model is memorizing instead of learning.
Watching a generalization curve helps you decide when to stop training before the model starts memorizing quirks of the training documents.
A flexible extension of basic linear models that can handle outcomes that are not just simple straight-line numbers, such as yes/no or counts.
It is a foundational, interpretable modeling tool — many classic predictions (like probability of an event) are built on it.
A generalized linear model could estimate the probability that a case settles based on factors like claim size and jurisdiction.
Two AI models that train against each other — one inventing fake data and one judging it — until the fakes look real.
This adversarial approach produced some of the first highly realistic AI-generated images and synthetic datasets.
A generative adversarial network can produce convincing synthetic photos, which is part of why deepfake authenticity is hard to verify.
AI that creates new content — text, images, audio, code, or video — rather than just classifying or scoring existing data.
Generative AI is the technology behind chatbots and writing assistants, and it is reshaping how knowledge work gets done.
Generative AI can draft a deposition outline from your case notes, which you then edit and verify.
A model that learns the patterns of its training data well enough to produce brand-new examples that resemble it.
Generative models underpin all content-creating AI, in contrast to discriminative models that only sort or score existing data.
A generative model trained on contract language can draft a plausible new clause in the same style.
The full name of the GPT family — transformer-based models pre-trained on vast text that then generate new text.
This architecture is the backbone of today's most widely used chatbots and writing tools.
Generative pre-trained transformers power the assistant that drafts a summary of a 200-page expert report.
The half of a generative adversarial network whose job is to create fake data convincing enough to fool the judge model.
Understanding the generator clarifies how AI learns to fabricate realistic images, audio, or synthetic records.
In a face-generating GAN, the generator is the part that actually produces each fake portrait.
A score that measures how mixed-up the categories are in a group of data — low means the group is nearly all one class.
Decision trees use Gini impurity to decide the best questions to split data on, which controls how accurately the tree sorts things.
A decision tree sorting documents into privileged vs. non-privileged uses Gini impurity to pick the split that best separates the two.
An explanation that describes a model's overall behavior with a simpler, understandable approximation of the complex original.
It helps people grasp how a model generally makes decisions, not just why it made one specific call.
Summarizing which factors a risk model relies on most across all its predictions.
Answers that are treated as the trusted, correct ground truth used to judge how well a model performs.
Without reliable gold labels you cannot tell whether a model is actually accurate — they are the measuring stick for evaluation.
Lawyers hand-label 500 documents as relevant or not; those gold labels become the benchmark for testing an AI review tool.
A carefully verified, high-quality set of examples with trusted correct answers, used as the authoritative benchmark for testing.
It is the reference standard a model is measured against, so its quality directly determines how meaningful your evaluation results are.
A gold standard dataset of expertly coded case documents lets you fairly compare two AI review tools head to head.
A free web tool from Google for experimenting with its AI models, writing prompts, and getting code to use them in an app.
It is the easiest on-ramp for trying Google's AI models without setup, useful for prototyping before committing to a build.
A developer uses Google AI Studio to test how Gemini summarizes a sample contract before wiring it into their product.
The process of setting direction, making decisions, and monitoring an AI system so it serves agreed-upon goals and rules.
Governance keeps AI use aligned with policy, ethics, and accountability instead of drifting unchecked.
An organization's policy deciding which AI tools may be deployed on sensitive case files and how they are monitored.
A family of AI language models, made by OpenAI, that generate human-like text after being trained on huge amounts of writing.
GPT models popularized modern chatbots and are widely used for drafting, summarizing, and answering questions.
A GPT model can draft a first-pass summary of a deposition transcript that a paralegal then reviews.
A specialized computer chip, originally built for graphics, that does the massive parallel math needed to train and run AI fast.
GPUs are the workhorses of modern AI — their availability and cost largely determine how expensive and fast AI projects are.
Training a custom document-classification model on thousands of files takes hours on GPUs instead of weeks on a normal computer.
A mathematical signal that tells a model which direction to adjust its settings to reduce its errors.
The gradient is the compass that guides nearly all model training — it is how a model knows how to improve.
During training, the gradient points the model toward settings that make its predictions slightly more accurate each step.
A technique that adds up the learning signals from several small batches of data before updating the model, mimicking one big batch.
It lets you train large models on limited memory hardware that could not otherwise fit a big batch of data at once.
On a single modest GPU, gradient accumulation lets a team fine-tune a model as if they had far more memory.
A powerful model made of many decision trees built in sequence, each correcting the errors of the ones before it.
GBDTs are a go-to choice for accurate predictions on structured, tabular data and often beat fancier methods there.
A gradient boosted decision trees model could rank thousands of documents by likelihood of being relevant to a case.
A method that builds many small models one after another, where each new model focuses on fixing the previous ones' mistakes.
Gradient boosting is one of the most accurate and widely used techniques for predictions on spreadsheet-style data.
A gradient boosting model can predict which incoming case files are likely to need urgent review based on past patterns.
A safety limit that caps how large a single learning adjustment can be so training does not spiral out of control.
It keeps training stable, preventing sudden huge updates that would wreck a model's progress — common in sequence models.
Gradient clipping stops a wildly large update from blowing up training partway through, like a governor on an engine.
The core training method where a model takes repeated small steps that reduce its errors, like walking downhill to the lowest point.
Gradient descent is how almost every modern AI model actually learns — it is the engine of training.
Through gradient descent, a model gradually tunes itself until its document predictions closely match the correct answers.
In TensorFlow, a map of all the computation steps a model performs, laid out so they can be run efficiently.
Representing computations as a graph lets AI frameworks optimize, distribute, and speed up the heavy math behind models.
TensorFlow turns a model's operations into a graph so the same model can run fast on different hardware.
A mode where TensorFlow first builds the full plan of computations as a graph, then runs it all at once for speed.
Graph execution makes models run faster and more efficiently in production than running each step one at a time.
Switching a trained model to graph execution speeds up how quickly it can classify a large batch of incoming documents.
A type of AI model designed to learn from data shaped like networks of connected things, such as people, places, or documents linked together.
Many real-world relationships are networks, and graph neural networks can find patterns in those connections that ordinary models miss.
A graph neural network could analyze a web of emails between people to surface key communication links in a case.
A decision strategy where an AI agent always picks the action it currently believes gives the best immediate reward.
It illustrates the core tension in learning agents: always grabbing the known-best option can miss better choices you never explored.
A greedy policy in a routing agent always takes the fastest known path, even if an untried route might be quicker.
The verified correct answer for an example, used as the standard to check whether a model's prediction is right.
Every measure of model accuracy depends on ground truth — without it you cannot know if the AI is correct.
If experts confirm a document truly is privileged, that confirmation is the ground truth the AI's guess is graded against.
Tying an AI's answers to verifiable real sources or data so it responds from facts rather than its own guesses.
Grounding is a primary defense against made-up answers — it forces the AI to back claims with actual documents.
A grounded legal assistant answers a question by quoting and linking the exact exhibit, not by recalling something vaguely.
The mistake of assuming what is true for an individual is true for their whole group, or vice versa.
This bias can creep into AI decisions and produce unfair outcomes by judging people based on group stereotypes rather than their own facts.
Assuming an applicant must behave like others from their region is group attribution bias — a fairness risk in automated screening.
The aim of having groups defined by protected traits receive similar treatment or outcomes from a model.
It is a core way to check that an AI system is not systematically advantaging or harming whole groups of people.
Confirming an approval model grants similar acceptance rates across different demographic groups.
Rules and checks that constrain what an AI system is allowed to do or say.
Guardrails keep tools safe and on-task — e.g. refusing to answer without a source, or blocking sensitive data from leaving your environment.
A guardrail requires every answer to include a document citation, so unsupported claims are blocked.
Someone who tries to or does break into a system without permission.
Understanding who attacks systems and how is the starting point for defending sensitive data against them.
An intruder attempting to access a confidential document repository they have no rights to.
When an AI model states something that sounds plausible but is factually wrong or made up.
Hallucinations are the central risk in legal use. Tools reduce them by forcing answers to cite real source documents.
An ungrounded model might invent a case citation; a grounded tool answers only from your uploaded files and links the source.
Techniques used to reduce how often a generative AI confidently makes up facts that aren't true.
For any tool people rely on for accuracy, reducing made-up answers is essential to trust; this is where grounding and source-linking come in.
Forcing the assistant to answer only from retrieved case documents, and to cite them, sharply cuts invented citations.
The physical machines used to process, store, or move computer programs and data.
Hardware sets the limits on how fast and how much AI work can be done, from training models to running tools.
The servers and GPUs that power a document-analysis platform behind the scenes.
An undesired outcome whose cost is high enough, in human terms, to count as genuinely damaging.
Defining harm is the basis of safety work, which aims to lower both the chance of expected harms and the risk of unexpected ones.
An AI tool that misclassifies a key document and causes a wrong decision in a case.
A conscious or unconscious tilt that leads to unfair, negative treatment of people based on traits like race, gender, or ability.
Harmful bias in AI can quietly produce discriminatory outcomes at scale, making detection and mitigation essential.
A screening model that systematically rates applicants from one group lower for no valid reason.
Turning categories with huge numbers of possible values into a smaller fixed set of buckets by running each value through a math function.
It lets a model handle features like ZIP codes or product IDs that have millions of possible values without needing millions of columns, keeping memory and training manageable.
Millions of distinct case-file IDs get hashed into 1,000 buckets so a model can use them as a feature without exploding in size.
The study and design of how people use and interact with computer interfaces and technology.
Good HCI makes AI tools usable and trustworthy instead of confusing, which decides whether people actually adopt them.
Designing a clear search-and-results screen so users can quickly find passages in a large document set.
An AI system that runs purely as a backend service through an API, with no built-in visual interface of its own.
Headless design lets the same AI capability be plugged into many different apps and screens, instead of being locked to one fixed interface.
A headless extraction service could feed pulled-out dates and names into both a dashboard and a report tool from one shared backend.
A simple, practical rule of thumb that gives a good-enough answer quickly without guaranteeing the perfect one.
Heuristics give you a cheap baseline to beat; if a fancy model can't outperform a one-line rule, the model isn't earning its keep.
Before building any model, you flag every document longer than 50 pages as 'likely an expert report' as a quick heuristic baseline.
A way of grouping similar items that builds a tree of clusters, merging or splitting groups step by step.
It reveals how data groups together at different levels of detail, so you can choose how finely to split things after seeing the structure.
Thousands of exhibits get organized into a tree where closely related documents merge first, letting you zoom from broad themes down to tight sub-groups.
A way of measuring a classifier's mistakes that pushes predictions to be not just correct but confidently correct by a margin.
It is the loss behind support vector machines and encourages a clear gap between categories, which can make decisions more robust.
A spam filter using hinge loss is penalized even when it labels an email correctly if it was only barely confident about the call.
A bar chart that shows how often values fall into each range, giving a picture of how a set of numbers is spread out.
Looking at a histogram is one of the fastest ways to spot odd data, outliers, or skew before you trust it for analysis or training.
A histogram of page counts across all filed exhibits shows most are under 10 pages, with a few enormous outliers worth checking.
Humans and AI systems working together on complex, changing tasks with smooth handoffs in both directions.
It aims to combine human judgment with machine speed so the team performs better than either could alone.
A reviewer and an AI tool splitting document analysis, each handing work back to the other as needed.
Data deliberately set aside and never used during training, so you can fairly check how well the model does on examples it has never seen.
Without holdout data you can't tell if a model truly learned or just memorized; it is the honest test of real-world performance.
You train a document classifier on 800 contracts and reserve 200 as holdout data to measure its accuracy on unseen contracts.
Methods for testing and optimizing all the human-related factors of a system viewed as one whole.
It ensures the people, tools, and processes work together well rather than optimizing technology in isolation.
Evaluating how reviewers, software, and workflow combine to handle a full case end to end.
Judging an AI system's output quality by having real people rate or compare the results, rather than relying only on automatic scores.
For tasks like writing or summarizing, human judgment is still the gold standard because automatic metrics miss tone, accuracy, and usefulness.
Three reviewers rate which of two AI-written case summaries is clearer and more faithful to the source.
Living people whom a researcher studies by interacting with them or by using their identifiable private information.
When research involves human subjects, consent and privacy protections become legal and ethical requirements.
A study analyzing identifiable witness records must follow rules protecting those individuals.
A way of working where a person steps in during a task's execution, at the level of its detailed plan.
It keeps a human involved at key moments so automation handles routine work while people guide the tricky parts.
An extraction tool that pauses for a reviewer to confirm an ambiguous field before continuing.
An approach to AI that keeps human needs, abilities, and oversight at the center of how systems are designed and used.
Designing AI around people, rather than replacing them, keeps humans in control and makes the technology genuinely useful.
A human-centered document tool drafts a summary but leaves the reviewer in charge of the final judgment.
The mental shortcuts and tendencies that shape how people interpret AI information or fill in gaps when making decisions.
These biases affect every stage of an AI system's life, so ignoring them can quietly distort how AI is built and used.
A reviewer over-trusting an AI's suggested answer simply because it sounds confident.
Learning built by watching how people operate software, capturing their successes and failures to power useful predictions.
It turns everyday human actions into training signal, letting tools learn to anticipate what users need.
A tool that learns which documents reviewers usually open next and surfaces them automatically.
A setup where a person reviews, approves, or corrects what an AI system produces instead of letting it act fully on its own.
Keeping a human in the loop catches AI mistakes before they cause harm and is often required for high-stakes or regulated decisions.
An AI drafts deposition questions, but an attorney reviews and edits every one before it is used.
The moment when a normally self-running system needs a person to step in because of unusual circumstances.
It is the safety valve that lets people take control when an autonomous system hits a situation it cannot handle.
An automated filing agent pausing for a person when it encounters a document type it does not recognize.
A setting you choose before training that controls how the model learns, rather than something the model figures out on its own.
Picking good hyperparameters can be the difference between a model that works and one that fails; they are the dials a practitioner tunes.
How fast the model adjusts during training (the learning rate) is a hyperparameter you set by hand before learning begins.
The process of trying different training settings to find the combination that makes the model perform best.
Most of the gain after a working model often comes from tuning; doing it systematically beats guessing and saves wasted training runs.
You test several learning rates and batch sizes, keeping the combination that scores highest on the validation set.
A flat dividing boundary that a model draws through data to separate one group from another, generalized to many dimensions.
Many classifiers work by finding the best hyperplane to split categories apart, so it's the geometric idea behind how a lot of decisions get made.
A classifier separates 'relevant' from 'irrelevant' documents by placing a hyperplane between the two clusters of data points.
A statistical method for checking whether observed results are real effects or could just be due to chance.
It gives a disciplined way to decide if a difference, like one model beating another, is meaningful or random.
Testing whether a new extraction model's higher accuracy is a true improvement or just noise.
The ability of a computer to identify objects, people, text, or actions inside a picture.
It turns raw scanned images into machine-understandable labels, which is the first step in reading or sorting visual documents and exhibits.
A system scans a folder of accident-scene photos and tags which ones show a damaged vehicle versus a road sign.
A training set where one outcome appears far more often than the other, like 99 normal cases for every 1 fraud case.
A model trained on lopsided data can score high just by always guessing the common answer while completely missing the rare cases that matter most.
Out of 10,000 reviewed contracts only 80 contain a problem clause, so the model needs special handling to learn to flag them.
A structured review that weighs an AI system's benefits against its costs and asks who it might harm before it is put into use.
It forces teams to think through failure modes and affected groups up front, rather than discovering harms after deployment.
Before rolling out a document-classification tool across thousands of case files, a firm runs an impact assessment to check who could be hurt if the tool mislabels sensitive records.
An attacker pretends to be a legitimate person in order to pass identity checks and trick a system into handing over that person's data.
It is the core threat that authentication is meant to stop, because a successful impersonation leads directly to a data breach.
Someone forges a valid-looking access request to a document system, posing as a client so they can pull confidential files that aren't theirs.
An automatic association or stereotype, held by people or baked into data, that influences decisions without anyone intending it.
Implicit bias in the data used to build a model can quietly carry unfair assumptions into the model's predictions.
If past hiring records implicitly favored one group, a model trained on them may repeat that pattern even though no one programmed it to.
Running an experiment through a computer simulation rather than in the physical world.
Simulating experiments lets researchers test ideas quickly and safely without the cost or risk of real-world trials.
A team tests how a model would behave on thousands of synthetic documents in silico before letting it touch real case files.
When a large language model learns to do a task purely from examples you put in the prompt, without retraining it.
It lets you steer an AI's behavior on the fly just by showing it a few worked examples, no engineering or retraining required.
Show the model three sample deposition summaries in the prompt and it produces a fourth in the same style.
The tendency to favor your own group or the categories you belong to, which can slip into data and labels.
When the people labeling training data show in-group bias, the model inherits a skewed, less fair view of the world.
Reviewers from one region rating documents from their own region more favorably can teach a model that same lopsided preference.
Reducing bias by changing how a model learns during training, such as adjusting its cost function or adding fairness constraints.
Tackling bias while the model trains can produce fairer results than only cleaning the data beforehand or correcting outputs afterward.
While training a model that flags relevant documents, engineers add a constraint so it doesn't systematically miss material from one group of authors.
A bias-reducing technique applied to a model while it is being trained.
It builds fairness directly into the model's learning step instead of relying only on fixing inputs or outputs.
A fairness-aware training routine penalizes the model whenever its predictions skew unfairly against a protected group.
A decision-tree test that checks whether a value belongs to a specific list of allowed categories.
It lets trees split on categories (like document type) rather than only on numbers, making them flexible for messy real-world data.
A node asks 'is the document type in {motion, brief, order}?' and routes matching files down one branch.
A situation where an AI system caused, or came close to causing, real harm in the world.
Tracking incidents helps organizations learn from near-misses and actual failures so they can prevent the next one.
A document-review AI quietly drops a category of files from its results, nearly causing a key exhibit to be overlooked, which is logged as an incident.
The formal reaction by an official body to an AI incident, directed at whoever built or deployed the system involved.
A clear response process means harms get investigated and the responsible party is held accountable instead of the problem being ignored.
After an AI tool produces a harmful error, a regulator opens an incident response asking the vendor to explain what went wrong and how it will be fixed.
In software quality assurance, the condition where the people checking quality are free from technical, managerial, or financial pressures that could sway their judgment.
Independent review keeps quality checks honest, since reviewers who answer to the team they audit may overlook problems.
An outside group with no stake in the project tests a legal AI tool so its quality findings can't be quietly shaped by the developers.
Independently and Identically Distributed
An assumption that each data point is drawn at random the same way and does not depend on the others.
Most statistics and machine-learning math assume i.i.d. data, so when real data violates it, models can be quietly unreliable.
Pulling random pages from a large archive is roughly i.i.d.; reading consecutive pages of one document is not, because nearby pages are related.
A fairness goal where any two people who are similar in the ways that matter for a decision get treated similarly by the model.
It checks fairness at the level of single people rather than only at the group level, catching cases where lookalikes get different outcomes.
Two applicants with nearly identical qualifications should receive nearly identical scores from a screening model.
The step where a trained model is actually used to make a prediction on new, real-world input.
Training builds the model once, but inference is what happens every time a user asks a question, so its speed and cost shape the live experience.
After training, the model runs inference each time it labels a newly uploaded exhibit as relevant or not.
The exact sequence of yes/no questions a decision tree walks through to reach a single prediction.
Because you can read the path step by step, decision trees offer a clear explanation of why a particular prediction was made.
For one case file the path might be: dollar amount over $1M? yes -> out of state? no -> route to senior review.
A measure of how much a yes/no split cleans up the data, used to decide which question a decision tree should ask.
It is the math that lets a tree pick the most useful questions first, which keeps the tree small and its decisions sharp.
Splitting documents by 'mentions a deadline?' yields high information gain if it cleanly separates time-sensitive files from the rest.
The part of a model that feeds it the assumptions and data it works from.
What goes into a model shapes everything it produces, so understanding the input component is key to trusting its output.
In a document-analysis model, the information input component is what supplies the case files and the rules the model should apply to them.
Keeping information confidential, accurate, and available, while also protecting qualities like authenticity and accountability.
Without it, sensitive data can be leaked, tampered with, or made unavailable, undermining trust in any system that handles it.
A document platform encrypts files, restricts who can open them, and logs every access to uphold information security for confidential case records.
Data that a system receives from an outside source.
Inputs are the raw material a model acts on, so their quality and content directly affect the results.
The PDFs and search queries you hand to a document tool are its input.
The first layer of a neural network, which receives the raw data before any processing happens.
It defines exactly what kind and shape of data the model expects, so getting it right is the foundation of any working model.
For a text classifier, the input layer takes the numeric representation of each word before deeper layers interpret it.
Harm caused by someone inside an organization who uses their authorized access and inside knowledge to act against it.
Insiders already have legitimate access and know where the sensitive data lives, so they can do serious damage that outside defenses won't catch.
An employee with access to a confidential document repository copies sensitive client files to sell them.
A single data point or example, such as one row in a spreadsheet or one document the model looks at.
Models learn from and make predictions about instances one at a time, so it is the basic unit everything else is measured against.
Each deposition transcript in your dataset is one instance the model can classify.
A number that scales how much a single data point counts toward a model's learning.
Weighting lets a model pay more attention to important or underrepresented examples and less to noisy or redundant ones.
Rare but critical document types are given higher instance weights so the model doesn't ignore them during training.
Extra training that teaches a language model to follow plain-English instructions and requests instead of just predicting text.
It is the difference between a model that rambles and one that actually does what you ask, making everyday prompts work as expected.
After instruction tuning, asking 'summarize this report in three bullet points' reliably produces three bullet points.
How well a system stops unauthorized people from accessing or changing its programs or data.
If data can be secretly altered, you can no longer trust that what you're reading is the original, accurate record.
A case-file system enforces integrity so no one can quietly edit an uploaded document after it has been stored.
Software that combines business rules, context, and decision logic to run multiple connected tasks across systems with little or no human help.
It can handle complex, multi-step workflows end to end, freeing people from repetitive coordination work.
A system intakes a new document, classifies it, routes it to the right reviewer, and updates the case record automatically.
A measure of how often different human reviewers give the same rating to the same items.
If human experts can't agree on the right answer, you can't fairly judge a model against their labels, so it sets the ceiling for trustworthy evaluation.
Three attorneys independently label the same 100 clauses as risky or safe, and you check how often their labels match.
An action that happens with the involvement of the surrounding environment of an object.
Many systems only do useful work through interaction, so understanding how a component engages its environment explains its behavior.
A chatbot's interaction with a user shapes which documents it retrieves and how it phrases its answer.
How free a study is from internal flaws, so its results truly reflect the cause-and-effect relationship being studied.
Without strong internal validity, you can't trust that an observed effect was actually caused by what the study claims.
A test claiming a new model improves search accuracy has high internal validity only if other factors, like a changed dataset, didn't secretly cause the gain.
The ability of different software or hardware systems to work together smoothly without much effort from the user.
Interoperable tools can share data and pass tasks between each other, avoiding manual copying and brittle workarounds.
A document tool exports results in a standard format so a case-management system can import them directly.
How easily a person can understand the reasoning behind a model's predictions in human terms.
In high-stakes settings like law or medicine, you often need to explain and defend why the AI reached a conclusion, not just trust it blindly.
A decision tree is highly interpretable because you can trace the exact questions that led it to flag a contract.
A machine learning model built to follow domain-specific rules so people can more easily understand how it reaches its results.
When a model's reasoning is understandable, people can check, trust, and challenge its decisions instead of taking them on faith.
A model that flags high-risk documents using a small set of readable rules lets a reviewer see exactly why each file was flagged.
A score for how well a predicted box around an object overlaps the correct box, from 0 (no overlap) to 1 (perfect).
It is the standard yardstick for grading object-detection systems, telling you whether the AI located something in the right place.
When a model draws a box around a signature on a scanned page, IoU measures how closely that box matches the true signature region.
The property that people can step in to control how their personal data is being processed, by whom, and at any stage.
It gives data subjects real control over their information, which is central to privacy rights and lawful data handling.
A platform lets a person pause or stop the processing of their submitted documents whenever they choose.
A data structure that holds rows of different lengths, so each example can have a different number of values instead of being forced into one fixed size.
Real data is rarely uniform, sentences and documents vary in length, so jagged tensors let models handle variable-length inputs without padding everything to the same size.
Storing 1,000 depositions where each has a different number of pages, without wastefully padding short ones to match the longest.
A high-performance numerical computing library from Google that combines automatic differentiation with hardware acceleration to run machine-learning code very fast on GPUs and TPUs.
JAX lets researchers and engineers train large models faster by automatically handling the math (gradients) and compiling code to run efficiently on specialized chips.
A team uses JAX to train a document-classification model across many TPUs, getting results in hours instead of days.
A method that automatically sorts data into a chosen number of groups by clustering items that are similar to each other.
It finds natural groupings in data with no labels, useful for organizing large piles of documents or records before anyone reviews them.
Run k-means on thousands of case files to auto-group them into themes like medical, financial, and correspondence.
A grouping method similar to k-means, but it measures distance differently, which makes it less thrown off by extreme outliers.
When a few unusual data points would distort normal clustering, k-median produces steadier, more representative groups.
Clustering expert fee records with k-median keeps a single huge invoice from skewing the groups.
A popular, beginner-friendly software library for building and training machine learning models without writing low-level math code.
Keras lets developers assemble a working AI model in a few lines, which is why many tools and prototypes get built fast.
An engineer uses Keras to stand up a model that flags which contract clauses look unusual, instead of coding the math from scratch.
A tool that automatically tries many setting combinations for a Keras model to find the ones that work best.
Choosing model settings by hand is slow and easy to get wrong; automating the search saves time and usually gives a better model.
Instead of manually testing dozens of configurations, a team lets Keras Tuner search for the settings that most accurately sort case documents by topic.
A small grid of numbers that slides across an image to detect a specific pattern, like an edge or a corner.
Kernels are the building blocks that let vision models recognize shapes and features in pictures and scanned pages.
A kernel passing over a scanned exhibit highlights the sharp horizontal lines of a printed table.
A classic classification method that draws a dividing line between categories, using a math trick (a kernel) to separate data that is tangled together.
KSVMs were a go-to tool for accurate classification before deep learning, and they still work well on smaller, well-structured datasets.
A KSVM sorts emails into 'privileged' versus 'not privileged' based on patterns in the text.
Specific notable spots in an image, like corners or joints, that a vision model uses as landmarks.
Keypoints let software line up, compare, or track objects across different photos, even when angle or lighting changes.
A model uses keypoints on a signature to compare it against signatures on other exhibits.
The total of all insight a cognitive computing system draws from descriptive, diagnostic, predictive, and prescriptive analytics.
Knowledge is what turns raw data and analysis into something a system can actually reason and act on.
By combining what happened, why, what's likely next, and what to do, a system builds the knowledge needed to recommend a document review strategy.
A penalty that pushes a model to use as few input features as possible by counting how many of its weights are non-zero.
It forces a model to rely on only a handful of signals, which keeps it simple and easier to explain.
A risk model trained with this penalty might keep just 6 of 400 possible factors, ignoring the rest entirely.
A way of measuring error by adding up the plain (absolute) sizes of the gaps between predictions and the true values.
Because it does not square the errors, one wild outlier hurts the score less, so the model stays steadier when data has odd extremes.
Scoring a price-prediction model by summing how many dollars off each guess was, ignoring direction.
A penalty that discourages large weights by adding up their absolute sizes, which tends to drive useless weights all the way to zero.
It automatically drops irrelevant features, producing a leaner model that is faster and easier to interpret.
Training a document classifier with this penalty so it keeps only the few words that actually predict the category.
A way of measuring error by squaring each gap between a prediction and the true value, then adding them up.
Squaring makes big mistakes count much more, so the model works hard to avoid large errors.
Penalizing a forecasting model heavily whenever a single prediction is far off, while small misses barely matter.
A penalty that keeps a model's weights small by adding up their squared sizes, nudging all weights toward (but not exactly to) zero.
It spreads influence across many features and curbs overfitting, helping the model generalize to new data.
Adding this penalty so a model trained on past cases does not latch onto quirks of the training set.
The correct answer attached to a training example that the model is trying to learn to predict.
Labels are the ground truth a supervised model learns from; without them there is nothing to teach the model what right looks like.
Each email in a training set tagged 'spam' or 'not spam' carries the label the model learns to reproduce.
A flaw where the model accidentally sees information during training that gives away the answer, information it would not have in real use.
It makes a model look brilliant in testing but fail in the real world, because the cheat-sheet clue disappears at deployment.
A model predicting whether a case settles 'learns' from a column that only gets filled in after settlement.
A change where how often each category appears shifts, even though the typical features of each category stay the same.
If the mix of categories in real use differs from training, a model's accuracy can drop even when nothing about the categories themselves changed.
A document classifier trained when most files were contracts performs worse once the incoming mix shifts toward emails, even though contracts and emails still look the same.
A training item that includes both the input features and the correct answer (the label).
Labeled examples are the raw material of supervised learning; gathering enough of them is often the hardest, costliest part of a project.
A contract paragraph paired with the human-assigned tag 'indemnification clause' is one labeled example.
A dial that sets how strongly a regularization penalty pushes a model toward simplicity.
Turn it too low and the model overfits; too high and it underfits, so tuning lambda is key to a balanced model.
Raising lambda until the model stops memorizing training quirks but still predicts new documents well.
A large conversational language model developed by Google and trained on dialogue to hold natural back-and-forth chats.
It was an early demonstration of how big language models could power fluent assistants, a forerunner to today's chatbots.
Asking a LaMDA-style assistant a follow-up question and getting an answer that stays on topic from the prior turn.
Specific marked points of interest in an image, such as the corners of an eye or the tip of the nose on a face.
Landmarks let vision systems measure shape and pose precisely, powering face alignment, gesture tracking, and document-region detection.
Detecting the four corner landmarks of a scanned form so the page can be straightened before reading the text.
A model that estimates how likely a sequence of words is, letting it predict or generate the next word in text.
Language models are the engine behind autocomplete, chatbots, and summarizers; understanding them clarifies what these tools can and cannot do.
Typing 'the witness testified that' and the model predicting the most likely words that follow.
A language model with a very large number of internal settings, trained on huge amounts of text so it can answer questions and write fluently.
These are the systems behind modern AI assistants; their scale is what makes them flexible but also prone to confident mistakes.
Pasting a 40-page expert report into an assistant and asking it to summarize the key opinions.
The time it takes between asking a model a question and getting its answer back.
High latency makes an AI tool feel sluggish; for live or interactive use, low latency is as important as accuracy.
A document-search tool that returns a source-linked answer in under two seconds has low latency.
An abstract numerical map where a model places items so that things with similar meaning sit close together.
It is what lets AI compare ideas by meaning rather than exact words, the foundation of semantic search and recommendations.
In latent space, two depositions describing the same accident land near each other even with different wording.
A set of neurons in a neural network that processes its inputs and passes results to the next set.
Stacking layers lets a network learn increasingly complex patterns; the number of layers is a core design choice.
An image network whose first layer detects edges and later layers combine them into shapes and objects.
A TensorFlow programming interface for building neural networks by stacking ready-made layers like building blocks.
It lets developers assemble models quickly without coding the math of each layer from scratch.
A developer adds a dense layer and a dropout layer with two lines of code instead of writing the underlying matrix operations.
An endpoint of a decision tree where no more questions are asked and a final prediction is given.
Leaves hold the tree's actual answers; how data splits on its way to a leaf is what makes tree models easy to follow.
After splitting on claim amount and case type, a tree reaches a leaf that predicts 'likely to settle'.
The process by which an AI program improves its performance by gaining knowledge.
Learning is what lets AI get better with experience instead of being limited to fixed, hand-coded rules.
A document tool that gets better at flagging relevant files as it processes more examples is demonstrating learning.
A setting that controls how big a step the model takes when adjusting itself during training.
Too big and training overshoots and never settles; too small and it crawls, so the learning rate strongly shapes whether training succeeds.
Lowering the learning rate so a model stops bouncing around and steadily improves each round.
A design principle that gives each user or component only the minimum access and permissions it needs to do its job.
Limiting access shrinks the damage anyone can do if an account is compromised or misused.
A reviewer can open only the case files assigned to them, not the entire document repository, under least privilege.
A method of fitting a line or model to data by making the total of the squared prediction errors as small as possible.
It is the classic, foundational technique for finding the best-fit relationship between numbers in data.
Fitting a straight line through scattered points of square footage versus home price to predict prices.
Grouping the different forms of a word together so they can be treated as a single item.
It lets text analysis recognize that words like run, running, and ran are the same root, improving search and matching.
Searching documents for ran also surfaces run and running once the text has been lemmatized.
A made-up example group used in fairness discussions to illustrate bias without naming a real population.
It lets people reason about fair versus unfair treatment of a group in a neutral, hypothetical way.
A fairness lesson asks whether a loan model approves Lilliputians and Brobdingnagians at equal rates.
A relationship where the output changes in direct, straight-line proportion to the inputs, with no curves or bends.
Linear relationships are the simplest to fit and explain, but many real-world patterns are not linear, which is why deeper models exist.
Predicting total cost as price per page times number of pages is a linear relationship.
A model that makes predictions by giving each input feature a weight and adding the weighted values together.
Linear models are fast, transparent, and a strong baseline; you can see exactly how much each factor pushes the prediction.
A pricing model that adds up weighted points for square footage, location, and age to estimate a home's value.
A model that predicts a number by fitting a straight-line relationship between the inputs and the value you want.
It is the most basic predictive model and a benchmark everything else is measured against; understanding it grounds all of ML.
Estimating how many hours a document review will take based on the number of pages.
An AI model trained on huge amounts of text that can read, summarize, and generate human-like language.
LLMs are the engine behind chat-style tools. They are powerful but can sound confident while being wrong — so grounding them in your documents matters.
An LLM drafts a summary of a 200-page expert report, then cites the exact pages it pulled each point from.
An explanation that covers a single decision or a small subset of decisions rather than the whole model.
Local explanations let you understand why one specific output happened, which is often what a person actually needs to question or trust a result.
A tool shows the exact passages that led it to flag this one document, giving a local explanation for that single decision.
Adapting a product to a specific country or region's language and conventions.
Localization makes a tool usable and natural for people in different regions instead of forcing one-size-fits-all defaults.
A document tool is localized so its interface, date formats, and prompts match the conventions of users in another country.
An error score for probability predictions that punishes confident wrong answers far more harshly than cautious ones.
It rewards a model for being not just right but honestly confident, the standard yardstick for probability-based classifiers.
A model that says 'spam with 99% certainty' on a real email racks up a large log loss for that mistake.
The logarithm of the odds of an event, a number that runs from large negative to large positive instead of staying between 0 and 1.
It is the natural internal scale for many classifiers; a sigmoid step converts log-odds back into an everyday probability.
A model computes log-odds for 'document is relevant', then converts that to a 78% probability for the user.
A model that predicts the probability that something belongs to a category, squeezing its output to a value between 0 and 1.
Despite the name, it is a workhorse for yes/no decisions and is prized for being fast and easy to interpret.
Estimating the probability that a given email is spam, then flagging it if the probability passes a threshold.
The raw, unscaled scores a classification model produces before they are turned into clean probabilities.
Logits are the model's gut numbers; converting them properly is what gives you trustworthy probability outputs.
A model outputs logits for each document category, which a softmax step then turns into percentages adding to 100.
A type of sequence-processing network with a built-in memory that can remember important details over long stretches of text or data.
Before transformers, LSTMs were the main way models handled long sequences without forgetting early context.
An LSTM reading a long witness statement keeps track of who 'he' refers to from many sentences earlier.
A single number that measures how far off a model's predictions are from the correct answers.
Loss is the score training tries to shrink; watching it drop is how you know a model is actually learning.
A model's loss starts high on day one and falls steadily as it learns to classify documents correctly.
A graph showing how a model's error changes over the course of training.
Its shape reveals at a glance whether training is working, stuck, or overfitting, the first chart engineers check.
A loss curve that flattens out signals training has converged and more rounds will not help much.
The specific formula a model uses to turn the gap between its predictions and the truth into a single error number.
Choosing the right loss function tells the model what counts as a mistake, which directly shapes what it learns to do well.
Picking a loss function that penalizes missed spam more than false alarms, so the filter errs toward catching spam.
An imaginary landscape of hills and valleys representing how a model's error changes as its internal settings change.
Training is the search for the lowest valley on this surface; its bumpiness explains why training can get stuck or take many tries.
Picture training as a hiker descending a foggy mountain of error, trying to reach the lowest point.
A value that a quantity is guaranteed not to fall below, used to set a floor on what a result can be.
Lower bounds give analysts a worst-case guarantee, useful when you need to promise a minimum level of performance.
Stating that a search tool will, at minimum, surface 90% of relevant documents sets a lower bound on recall.
A method inside AI where a system learns patterns from data instead of being explicitly programmed with rules.
Most modern AI is machine learning. The quality of its output depends heavily on the data it learned from — garbage in, garbage out.
Show a model thousands of labeled documents and it learns to sort a new document into 'deposition' vs 'expert report' on its own.
A machine's ability to detect and make sense of meaningful events and conditions that affect how the system or its processes run.
It lets systems notice and respond to what's happening around them, which is the basis for monitoring and self-correction.
A processing pipeline observes that a document upload failed and reroutes the job before it stalls the whole workflow.
Using a model to automatically translate text or speech from one language into another.
It lets you work with documents and testimony in languages you do not read, without waiting for a human translator on every page.
Running a Spanish-language contract through a model to get a usable English draft before a certified translation is ordered.
In a dataset where one outcome happens far more often than the others, the majority class is the more common label.
If 99% of cases are 'not fraud,' a lazy model can score 99% accurate by always guessing the majority class while catching zero real fraud.
In 10,000 emails where only 50 are spam, 'not spam' is the majority class.
Hardware, firmware, or software that has been deliberately built or planted in a system to cause harm.
Malware can steal, alter, or destroy the very documents and data an AI tool relies on, so guarding against it protects both the system and the evidence inside it.
An attacker emails a poisoned PDF to a firm; opening it installs hidden software that quietly copies confidential case files.
A math framework for decision-making where the next situation depends only on the current situation and the action taken, not on the full history.
It is the standard way to model problems where an agent makes a sequence of choices and earns rewards, which underlies most reinforcement learning.
A delivery robot choosing its next move based only on where it is right now, not every step it took to get there.
The assumption that the future depends only on the present state, not on how you arrived at that state.
This simplifying assumption makes many sequential decision problems solvable, because you only need to track the current situation.
In a board game, your best next move depends on the current board, not the order of earlier moves that produced it.
A model trained by hiding some words in a sentence and learning to predict the missing words from the surrounding context.
This fill-in-the-blank training is how models like BERT learn rich language understanding used for search and classification.
Given 'The witness signed the [MASK] under oath,' the model learns to predict 'affidavit' or 'statement.'
How important a piece of information is to the people who will rely on it when making a decision.
Knowing what is material helps teams focus attention on the facts that actually change conclusions instead of drowning in minor details.
In a financial review, a small rounding difference is immaterial, but a missing million-dollar liability is material to anyone reading the report.
A popular Python library for drawing charts and graphs of data.
It is the go-to tool for visualizing data and model results so people can actually see patterns, not just numbers.
Plotting how a model's error drops over training rounds to confirm it is learning.
A technique that breaks a big table of user-item interactions into two smaller tables that capture hidden user tastes and item traits.
It is a core method behind recommendation systems, letting a service predict what you'll like even for items you have never seen.
Predicting which movies a user would rate highly by learning compact 'taste profiles' from sparse ratings data.
The mistake of trusting only what is easy to measure with numbers and ignoring important things that are hard to quantify.
Relying solely on convenient metrics can hide real drivers of success or failure, leading to confident but wrong decisions.
Judging a document-review AI only by how fast it processes pages, while ignoring whether it caught the most important passages.
The average size of a model's prediction mistakes, ignoring whether each guess was too high or too low.
It gives one easy-to-read number for how far off, on average, a model's numeric predictions are.
If a model predicts case durations and is off by 5 days on average, the mean absolute error is 5 days.
A score for ranked results that rewards putting the right items near the top of a short list, averaged across many queries.
It measures how good a search or recommendation list is when users only look at the first few results.
Checking whether the most relevant exhibits show up in the top 10 results across hundreds of test searches.
The average of the squared differences between predictions and actual values, which punishes big mistakes much more than small ones.
It is one of the most common ways to measure and minimize error when a model predicts numbers.
A pricing model gets penalized heavily for a single wildly wrong estimate because the error is squared.
Whether a quality of something can actually be assessed against a metric.
If a goal cannot be measured, it is hard to know whether an AI system is meeting it or getting better.
Answer accuracy has high measurability, while a fuzzy goal like 'the AI feels helpful' is harder to measure directly.
The act of assigning a number or category to something in order to describe one of its qualities.
Measurement turns vague impressions into comparable values, which is the foundation for evaluating and improving any AI system.
Scoring how often an AI's answers cite the correct source document gives a measurable quality you can track over time.
A general description of the steps and logic used to carry out a measurement.
Spelling out the method makes a measurement repeatable and lets others check or reproduce the result.
A documented method for grading AI answers might define how a reviewer reads each response and assigns a correctness label.
A statistical model that checks how well observable indicators line up with the underlying concept they are meant to capture.
It tells you whether the things you can measure actually reflect the deeper quality you care about, rather than something unrelated.
A measurement model might test whether several survey questions truly add up to one trait like 'trust in an AI tool.'
Figuring out whether a specific record was part of the data a machine learning model was trained on.
It is a privacy attack: confirming that someone's data was used in training can leak sensitive facts about that person.
An attacker probes a model to determine whether a particular client's case file was in its training set.
A way of splitting a large model and its data across many connected chips so they can train it together.
Huge modern models do not fit on a single chip, so a mesh lets many processors share the work in parallel.
Spreading a giant language model across hundreds of accelerator chips arranged in a grid to train it faster.
A hidden instruction set behind the scenes that tells the model who it is and how to behave, before the user types anything.
It is where you fix the model's role, tone, and rules once, instead of repeating them in every message. Get it right and every later answer stays on-task.
Before any question, the system prompt says: 'You are a paralegal assistant. Cite the exhibit number for every fact and never give legal advice.' Every user query is then answered under those standing rules.
Training a system to get better at learning itself, so it can pick up new tasks quickly from very few examples.
It helps models adapt to brand-new problems fast, instead of needing huge datasets for every new task.
A model that, after seeing just three labeled contract clauses, can reliably spot that clause type in new contracts.
Data that describes other data, such as when a file was created or who authored it.
Metadata helps organize, search, and verify documents, and in legal work it can itself be important evidence.
The hidden creation date and author fields on a Word document are metadata that may reveal when a contract was really written.
A number you track to measure how well a model or system is performing.
Without metrics you cannot tell whether a model is good, getting better, or quietly failing.
Tracking accuracy and error rate to decide if a new model version is actually an improvement.
A set of ready-made tools in the Keras framework for computing performance numbers during model training and testing.
It saves engineers from hand-coding common measurements like accuracy, so they can monitor models reliably.
Adding a one-line accuracy metric to a training script so progress prints automatically each round.
A small, randomly chosen slice of the training data that a model learns from in one step instead of the whole dataset at once.
Learning from mini-batches makes training faster and more memory-efficient than processing every example at the same time.
Updating a model using 32 documents at a time rather than loading all 100,000 into memory.
A training method that improves a model step by step, using a small random batch of examples to estimate which direction to adjust.
It is the workhorse training approach behind most modern models, balancing speed and stability.
Nudging a model's settings after each group of 64 examples until its predictions steadily improve.
A scoring rule used when two parts of a system compete, where one tries to maximize a value while the other tries to minimize it.
It is the tug-of-war that trains generative adversarial networks, pushing one network to make fakes and another to catch them.
A fake-image generator and a fake-detector both improving as each tries to outsmart the other.
Limiting the personal data an AI system collects and uses to only what is genuinely needed for its purpose.
Collecting less personal data reduces privacy risk and helps organizations comply with data protection laws.
A document AI strips out unrelated personal details and keeps only the case facts it needs to answer questions.
In a lopsided dataset, the rarer outcome that shows up far less often than the common one.
The minority class is usually the one you care about most (fraud, disease, defects), yet it is the hardest for a model to learn.
The 50 spam emails inside a pile of 10,000 are the minority class.
A research approach that combines numerical data with descriptive, qualitative data and integrates both to answer a question.
Blending numbers with context gives a fuller picture than either type of data could provide alone.
Evaluating an AI tool by both measuring its accuracy and interviewing lawyers about how it fits their workflow.
A model design that contains many specialized sub-models and routes each input to only the few experts best suited to handle it.
It lets a model be very large and capable while only activating a small slice per request, saving computing cost.
A language model sending a coding question to its code-savvy experts and a legal question to different ones.
The common abbreviation for machine learning, where software learns patterns from data instead of being explicitly programmed with rules.
Most AI tools you use are powered by ML, so the shorthand appears everywhere in product and technical discussions.
A document tool that learns to flag privileged passages from past examples is using ML.
The collection of practices and tools for reliably deploying and running machine learning models in production.
MLOps keeps deployed models working smoothly over time, handling updates, monitoring, and failures so users get dependable results.
An MLOps pipeline automatically retrains and redeploys a document-classification model when its accuracy starts to slip.
A classic practice dataset of tiny handwritten-digit images used to test and teach image-recognition models.
It is the 'hello world' of machine learning, the benchmark beginners use to confirm their model setup works.
Training a first image classifier to tell handwritten 3s from 8s on the MNIST images.
A type of data a model can take in or produce, such as text, images, audio, or video.
Knowing a tool's modalities tells you whether it can read a scanned exhibit image, listen to a recording, or only handle plain text.
A model that accepts both a photo of a document and a typed question handles two modalities.
The trained system that takes an input and produces a prediction or answer, having learned patterns from data.
The model is the actual 'brain' doing the work behind any AI tool, so its quality determines the tool's quality.
Feeding a contract into the model and getting back a list of flagged risky clauses.
A check written over a model's inputs and outputs that flags when something looks wrong.
Assertions act like guardrails that catch likely errors automatically before they reach users.
An assertion fires if an AI answers a legal question without citing any source document, signaling a possible hallucination.
How complex a model is, meaning how rich a set of patterns it is able to learn.
Too little capacity and a model misses real patterns; too much and it memorizes noise instead of generalizing.
A tiny model may fail to learn legal language nuance, while an oversized one may just memorize the training documents.
A short document that accompanies a trained model and reports how it was tested, where it should be used, and how it performs across different groups.
Model cards give users honest, upfront information so they can judge whether a model is suitable and fair for their situation.
A model card for a document AI lists its tested accuracy and notes that it was only validated on English-language contracts.
Sending an input to a small cheap model first and only escalating to a bigger expensive model when the small one is unsure.
It cuts cost and latency by reserving the powerful, pricey model for the genuinely hard cases.
A quick model handles routine document tags, passing only ambiguous ones to a larger, slower model.
The work of diagnosing why a model is making mistakes.
Understanding the cause of failures is the first step to fixing a model and trusting its outputs.
Debugging reveals that an AI misreads scanned documents whenever the page is rotated, pointing to an OCR step to fix.
The gradual decline in a model's performance as time passes.
Models can quietly get worse as the world changes, so spotting decay early prevents bad results from slipping through.
A contract-analysis model grows less accurate as new clause styles appear that it never saw during training.
Making quick, targeted updates to a trained model's behavior for a small area without harming how it handles everything else.
It lets teams correct or update specific facts in a model cheaply, instead of retraining the whole thing.
Editing a model so it stops giving an outdated answer about a changed regulation, while leaving its other answers intact.
An attack where someone repeatedly queries a model to build a close copy of it without permission.
It lets a dishonest user steal a proprietary or sensitive model just by using its interface, threatening intellectual property and security.
A competitor sends thousands of queries to a paid AI service to recreate a near-identical model of their own.
An organization's overall framework of rules and roles for how models get built, validated, and used.
Clear governance ensures models are controlled, accountable, and used responsibly rather than ad hoc.
A firm's model governance policy requires every AI tool to be independently validated before lawyers can rely on it.
A central database that records information about all the models an organization is using.
Knowing exactly which models exist and where they run is essential for managing their risks and keeping them up to date.
A firm's model inventory lists every AI tool in use, its purpose, and when it was last validated.
Moving an AI model from one environment to another — for example, from a test setup to production, or between cloud providers.
Migrations are where things break: settings, versions, and performance can shift, so moving a model safely takes planning and testing.
A team might migrate a tested language model from its research servers to production servers before launching a client-facing tool.
A manual or judgment-based adjustment to a model's inputs or outputs to make up for its known limitations.
Overlays let experts correct for gaps a model cannot handle on its own, but they must be tracked since they override the model.
An analyst manually adjusts a model's risk score upward because it does not account for a recent event the data missed.
Splitting a single large model across multiple chips because it is too big to fit on one.
It is what makes training and running today's giant models physically possible.
Placing the first half of a huge model on one chip and the second half on another so they run as a team.
The governance and controls used to oversee and limit the risks that come from relying on models.
It protects organizations from costly mistakes caused by flawed or misused models through oversight, policies, and checks.
A firm's model risk management program requires senior sign-off before an AI tool is used in client work.
A component that decides which model should handle a given request, based on the task or difficulty.
It automatically sends each query to the most appropriate or cost-effective model, improving speed and savings.
A router sending simple questions to a fast model and complex legal reasoning to a more capable one.
A set of models that operate together as a group.
Many real tools combine several specialized models, so understanding the suite explains how the overall system produces a result.
A document AI might use one model for OCR, another for finding relevant passages, and a third for writing the answer.
The process of teaching a model by showing it data and adjusting it until its predictions get good.
Training is where a model actually learns; the quality and amount of training data shape everything it can do.
Showing a model thousands of labeled clauses so it learns to recognize indemnification language.
The set of processes used to confirm that a model is performing the way it is supposed to.
Validation is how teams build justified confidence that a model can be trusted before relying on its outputs.
Before deploying a legal AI, the team validates it against a set of known questions with verified answers.
A training trick that keeps a model moving in a consistent improvement direction, smoothing out the bumps as it learns.
It speeds up and stabilizes training, helping a model reach a good solution faster and avoid getting stuck.
Like a ball rolling downhill that keeps rolling through small dips instead of stopping at every tiny bump.
Keeping watch over a system's ongoing activities and results to check they stay as expected.
Continuous monitoring catches problems like declining accuracy or outages early, before they affect users.
A team monitors a deployed document AI and gets alerted when its answer accuracy drops below a set threshold.
The capacity to act, reason, and make decisions in a morally responsible way, beyond just causing moral effects.
It raises the question of whether and when an AI system can be held responsible for its actions, or whether people remain accountable.
Debate over moral agency asks whether a self-driving car can be blamed for a crash or whether responsibility stays with its makers.
The moral standing of an entity in terms of how it deserves to be treated.
It frames debates about whether AI systems or other entities are owed any moral consideration in how we treat them.
Discussions of moral patiency ask whether an advanced AI should have any protections, separate from whether it can act morally.
Sorting each input into exactly one of three or more possible categories.
Many real tasks have more than two answers, so this is how models pick the single best label among many.
Tagging each document as 'contract,' 'email,' 'invoice,' or 'memo.'
A mechanism where a model looks at a piece of text from several different angles at once to figure out how words relate.
It is the core engine of transformers, letting models understand long, complex sentences and context.
While reading a clause, the model simultaneously tracks subject-verb links, dates, and party names.
Letting each input receive several labels at the same time, instead of just one.
Real documents often belong to multiple categories at once, so this matches how messy real data actually is.
Tagging one filing as both 'confidential' and 'time-sensitive' and 'medical records.'
A large language model that can understand more than just text, such as images, audio, or video, alongside words.
It can answer questions about a scanned exhibit, a chart, or a photo, not only typed text.
Uploading a photo of a handwritten note and asking the model to transcribe and summarize it.
A model that can work with more than one kind of data, like combining text and images in a single task.
It lets a tool reason across formats at once, such as reading a document image while answering a written question about it.
Showing the model a contract page image and asking which party signed it.
Another name for sorting inputs into one of three or more categories.
It is the same idea as multi-class classification and shows up when a model must choose one label from many.
Deciding whether a review is positive, neutral, or negative.
A statistical method that predicts which of several categories an input belongs to, giving a probability for each.
It is a simple, interpretable way to do multi-class prediction and assign confidence to each possible answer.
Estimating the chance a document is a contract versus an email versus an invoice and picking the highest.
A simple classification method that assumes every input feature contributes to the answer independently of the others.
It is fast, needs little data, and works surprisingly well for tasks like sorting documents or flagging spam — a good baseline before reaching for anything fancier.
Classifying incoming emails as relevant or junk by treating each word as an independent clue about the category.
A classifier built on the Naive Bayes assumption that each feature independently influences which category something belongs to.
It gives a quick, low-cost way to auto-sort large piles of text into categories, which is handy for triaging documents.
Auto-routing thousands of intake documents into case-type folders based on the words they contain.
A training failure where one bad number (a 'NaN', meaning 'not a number') spreads through the model and corrupts everything downstream.
A single NaN can silently ruin a whole training run, so engineers watch for it to avoid wasting hours of compute.
A dividing-by-zero glitch produces a NaN that then poisons every later calculation in the training loop.
The part of AI that produces human-readable text, such as sentences, summaries, or answers.
It is what turns a model's internal results into a written response a person can actually read.
Turning a table of extracted dates into a flowing paragraph that narrates the timeline of events.
The broad field of teaching computers to read, interpret, and produce human language.
Nearly every document-AI feature — search, summarizing, extraction, translation — is built on natural language processing.
Summarizing a 200-page expert report into a one-page brief is a natural language processing task.
The part of AI focused on figuring out what a person actually means when they type or say something, not just the literal words.
It is what lets a tool grasp the intent behind a messy question and respond with something useful rather than a keyword match.
A user asks 'who paid whom and when' and the system understands it as a request to extract payment relationships from the text.
In a yes/no classification, the 'no' outcome — the category you are testing against, such as 'not the thing we're looking for'.
Knowing which class is negative is essential for reading accuracy numbers like false positives and false negatives correctly.
When flagging privileged documents, 'not privileged' is the negative class.
A training shortcut where the model learns from a few wrong examples instead of every possible wrong answer.
It makes training embeddings and recommenders far faster and cheaper without hurting quality much.
When learning word meanings, the model contrasts a real word pair against a handful of randomly picked non-matching words.
An improved version of momentum that 'looks ahead' before taking each step, helping training settle faster.
It can speed up model training and reduce overshooting, getting to a good result in fewer steps.
A hiker who glances ahead before each stride to avoid overstepping, rather than just coasting on prior speed.
A model made of many connected simple units ('neurons') arranged in layers that learn patterns from data.
Neural networks power almost all modern AI, from language tools to image recognition.
A neural network trained on labeled clauses learns to spot indemnification language in new contracts.
A single tiny computing unit inside a neural network that takes inputs, weighs them, and passes on a result.
Neurons are the basic building blocks of every modern AI model; stacking millions of them is what gives models their power.
One neuron in a document model might respond strongly whenever a date-like pattern appears.
A point in a decision tree where a yes/no question is asked or a final answer is given.
Reading the nodes top to bottom shows exactly why the model reached its conclusion — useful when you need an explainable decision.
A node asks 'Was the document signed after 2020?' and routes the case down one of two branches.
Another word for a neuron — one unit within a layer of a neural network.
It is common shorthand in diagrams and discussions of how a network is structured.
A network diagram shows each layer as a column of nodes connected by lines to the next layer.
A single operation or value in a TensorFlow computation graph, the map of steps a model performs.
Understanding nodes helps engineers debug and optimize how a model actually runs.
A node in the graph represents the 'add' step that combines two intermediate results.
Random or irrelevant variation in data that has nothing to do with the real pattern you want to learn.
Too much noise can fool a model into learning junk, so cleaning or accounting for it improves results.
Scanning errors that turn some letters into garbage characters add noise to OCR'd documents.
A question in a decision tree that has more than two possible answers, not just yes/no.
It lets a tree split data into several groups at once, which can model categories like 'state' more naturally.
A node splits documents by jurisdiction into 'federal', 'state', or 'international' in one step.
Treating people and groups in the same fair way rather than favoring or penalizing some over others.
It is a core fairness goal for AI, ensuring systems do not produce unfair outcomes based on attributes like race or gender.
A hiring AI built for nondiscrimination is checked to confirm it does not screen out candidates because of their gender.
A relationship that does not follow a straight line — outputs change in uneven, curved ways as inputs change.
Real-world patterns are usually nonlinear, and capturing them is what makes deep models more capable than simple ones.
Doubling the length of a contract does not double the time to review it — the relationship is nonlinear.
A distortion in data caused when certain people or cases are missing because they did not respond or were not recorded.
If the missing data is not random, the model learns a skewed picture and can make unfair or wrong conclusions.
A satisfaction model trained only on customers who returned a survey misses the views of everyone too unhappy to reply.
When the patterns in data change over time, so what was true yesterday may not hold tomorrow.
A model trained on old data can quietly become wrong as conditions shift, so teams must watch for drift and retrain.
Fraud tactics evolve constantly, so a detector trained last year gradually misses newer schemes.
The way a process is meant to run from its start, through every planned step, to a successful finish.
Defining the normal flow makes it clear what 'working as intended' looks like, so any deviation can be spotted and handled.
A document intake process where a file is uploaded, scanned with OCR, indexed, and marked ready for search follows its normal flow when nothing fails along the way.
Rescaling numeric data so all values sit in a comparable range, like converting everything to a 0-to-1 scale.
It helps models train faster and more reliably by stopping one large-valued feature from dominating the rest.
Putting page counts and dollar amounts on the same scale so neither overwhelms the other during training.
Spotting new data that looks unlike anything the model was trained on.
It flags unusual or never-before-seen cases that may need a human to review rather than trusting the model.
An intake system flags a document type it has never encountered so a person reviews it instead of auto-filing.
Information represented as numbers that you can do math on, like amounts, counts, or dates.
Models handle numbers and text differently, so knowing which fields are numerical guides how you prepare data.
The settlement amount and number of exhibits in a case file are numerical data.
The single number a training process is trying to push up or down, like a score the model aims to improve.
Picking the right objective decides what the model actually learns to optimize — the wrong one teaches the wrong behavior.
A model trained to maximize click-throughs has 'clicks' as its objective, even if that is not what you truly care about.
Factual data that proves something exists or is true, gathered by observing, measuring, or testing rather than by opinion.
Decisions and audits hold up better when they rest on objective evidence instead of assumptions or unsupported claims.
A timestamped log showing exactly which document passages an AI used to answer a question is objective evidence of how the answer was produced.
The math formula that calculates the score a model is trying to maximize or minimize during training.
It is the formal target of learning — every weight adjustment is made to improve this function's value.
During training, the system tweaks settings step by step to lower the objective function that measures total error.
A single piece of information an AI system receives, whether from a user, a sensor, or another source.
Observations are the raw inputs a system reacts to, so their quality and accuracy shape every decision the system makes.
When a user types a question into a document search tool, that query is an observation the system acts on.
Doing model work in advance on stored data, rather than live as new data streams in.
Offline work lets you process big batches calmly without the speed pressure of answering users in real time.
Tagging every document in an archive overnight is offline work; answering a live query the moment it arrives is not.
Generating all a model's predictions ahead of time and saving them, so you just look them up later instead of computing on the spot.
It makes responses instant for users and avoids paying for compute at the moment of the request, at the cost of freshness.
A tool pre-scores every contract in your library overnight, so the risk rating appears instantly when you open one.
Training a model on a complete, fixed dataset all at once rather than as new data arrives.
Offline learning is simpler and lets you train on the whole picture, but the model won't reflect new information until you retrain it.
Building a classifier from a finished archive of past case documents, then deploying it without further updates, is offline learning.
Turning a category into a row of 0s with a single 1 marking which category it is.
Models need numbers, not words, so one-hot encoding is a basic way to feed categories like document type into a model.
For document types contract, email, and memo, an email becomes [0, 1, 0].
Another name for one-hot encoding: marking one category with a single 1 among 0s.
You will see this term used interchangeably with one-hot encoding when converting categories to numbers.
A status field with values open, closed, pending is encoded so 'closed' becomes [0, 1, 0].
Teaching a model to recognize something correctly after showing it just a single example.
It matters when you only have one sample per category and cannot gather thousands of training examples.
A facial-recognition system learns to match a person from one photo on file.
Giving a language model exactly one example of the task inside the prompt before asking it to do the same.
A single example often steers the model toward the right format and style without any retraining.
You show the model one sample 'question to summary' pair, then ask it to summarize a new deposition the same way.
Handling a multi-category problem by training a separate yes/no classifier for each category against everything else.
It is a simple way to reuse two-choice classifiers for problems with many possible labels.
To sort documents into contract, brief, or invoice, you train three detectors: 'contract or not', 'brief or not', 'invoice or not'.
Handling many categories by training a separate classifier for every possible pair of categories.
It is an alternative to one-vs-all that can be more accurate but needs many more sub-classifiers.
For three document types you train classifiers for contract-vs-brief, contract-vs-invoice, and brief-vs-invoice, then vote.
Doing model work live, responding to each new piece of data as it arrives in real time.
Online work keeps answers current the instant something changes, unlike pre-computed offline work.
Scoring a fraud risk the moment a transaction happens is online work.
Computing a model's prediction on demand, right when a request comes in, instead of using a saved answer.
It gives the freshest possible answer for the exact input asked about, but you pay compute and a slight delay each time.
You upload a brand-new contract and the tool analyzes it on the spot rather than from a pre-computed score.
Updating a model bit by bit as new data streams in, instead of training on a fixed dataset all at once.
Online learning keeps a model current with fresh data, which is useful when conditions change over time.
A spam filter that adjusts its rules each time a user marks a new message as junk is doing online learning.
A structured map of the concepts in a subject area, their properties, and how they relate to each other.
An ontology gives software and people a shared vocabulary so different systems can understand and reason about the same information consistently.
A legal ontology might define how 'witness', 'deposition', and 'testimony' relate, letting a tool connect related documents automatically.
The quality of an AI system being hard to understand, where even its makers and users can't easily see how it reaches its conclusions.
Opacity makes it difficult to trust, debug, or challenge an AI's decisions, which is a serious concern in high-stakes settings.
A deep model that flags a contract clause as risky without showing why is opaque, leaving reviewers unsure whether to rely on it.
An AI model whose trained internal numbers, called weights, are released publicly so anyone can download and run it.
Open weights let teams run a model on their own systems, inspect it, and customize it without depending on a single provider.
A firm wanting full control over confidential data might run an open-weight model on its own servers instead of a hosted API.
A single computational step in a model's graph, such as adding, multiplying, or applying a function to numbers.
Models are built from many small operations chained together; understanding them helps when debugging or tuning performance.
A multiplication of two number tables inside a neural network is one operation in its computation graph.
Turning an AI system or concept into something actually in use so its effects can be measured.
An idea or model only delivers value once it is operationalized into real workflows where its performance can be tracked.
Moving a document-summarization model from a research notebook into the live review pipeline, with metrics on its accuracy, is operationalization.
The person who controls or oversees a system during human-in-the-loop, remote, or semi-autonomous operation.
A human operator provides judgment and a safety check, stepping in when an automated system reaches the limits of what it should decide alone.
A reviewer who approves or rejects each AI-suggested document classification acts as the operator in the loop.
When a person takes a deliberate action to agree to share their information, rather than it being shared by default.
Opt-in puts people in control of their own data, which builds trust and is often required by privacy laws.
A user must tick a box before a tool may share their uploaded documents with a third-party service.
When a person takes a deliberate action to refuse having their information shared, which would otherwise happen by default.
Opt-out lets people stop unwanted data sharing, though it puts the burden on them to act rather than defaulting to privacy.
A user clicks a link to tell a service not to share their account data with outside partners.
The algorithm that decides how to adjust a model's internal settings to reduce its errors during training.
The optimizer controls how fast and how well a model learns; a good choice can mean the difference between training succeeding or failing.
An optimizer nudges the model's weights a little after each batch so its predictions get steadily more accurate.
The tendency to see members of groups you don't belong to as more alike than they really are.
When this human bias leaks into training data or labels, models can treat unfamiliar groups as interchangeable and judge them unfairly.
Labelers from one region tag profiles from another region with less nuance, flattening real differences between those people.
The result or consequence that follows from an action or process.
Judging an AI system by its outcomes, not just its internal steps, shows whether it actually achieves what people need.
Faster, more accurate document review is the outcome a legal team cares about when adopting an AI search tool.
A data point that sits far away from most of the others, looking unusual or extreme.
Outliers can be valuable signals (real anomalies) or harmful noise that throws off a model, so spotting them matters.
In a set of contracts averaging 10 pages, a single 4,000-page filing is an outlier.
The task of automatically finding the data points that don't fit the normal pattern.
It catches errors, fraud, or rare important cases hiding in large piles of data that a person would never review by hand.
Scanning thousands of expense entries to flag the few that look nothing like the rest.
Data that a system sends out to an external destination.
Output is what users actually see and act on, so its correctness and format determine how useful the system is.
A source-linked answer returned by a document search tool is its output.
The final layer of a neural network that produces the model's actual answer or prediction.
It shapes the form of the result — a single number, a probability, or a set of category scores — that you ultimately read.
In a document classifier, the output layer gives the probability that a file is a contract versus a memo.
When a model memorizes its training examples so closely that it fails on new, unseen data.
An overfit model looks great in testing but performs poorly in the real world, which makes its results untrustworthy.
A model that perfectly labels the 500 contracts it trained on but misreads a fresh contract has overfit.
Repeating or duplicating examples from a rare category so the model sees enough of them to learn it.
Without it, a model can ignore an important but uncommon class entirely because it is drowned out by common ones.
If only 2% of filings are fraud cases, you copy those fraud examples extra times so the model takes them seriously.
A formal theory describing when a model can learn a rule from examples well enough, most of the time, given enough data.
It gives a mathematical guarantee that learning is actually possible, helping researchers know how much data a problem needs.
PAC learning explains why a spam filter trained on enough labeled emails will usually classify new emails correctly.
A folder bundling all the code and information needed to train and run a machine learning model.
Packaging everything together makes a model portable and reproducible, so it can be deployed reliably across environments.
Shipping a document-classification model as a package lets another team load and serve it without rebuilding it from scratch.
A value inside a model that the system adjusts on its own during training to make better predictions.
Parameters are what a model actually 'learns' — a model with billions of them has billions of tuned dials.
A large language model that drafts contract summaries may have hundreds of billions of parameters set during training.
A dedicated machine (or set of machines) that stores and updates a model's parameters while many other machines train in parallel.
It is a core piece of infrastructure that lets very large models train across many computers at once.
When training a recommendation model across 100 servers, a parameter server keeps the shared weights in sync.
A single adjustment to a model's internal values during training, nudging it toward better predictions.
Training is just millions of these small updates; understanding them explains how models gradually improve.
After each batch of training documents, the model performs a parameter update to reduce its errors.
A way to adapt a large pre-trained model to a new task by adjusting only a small fraction of its values instead of all of them.
It makes customizing huge models far cheaper and faster, since you don't retrain everything from scratch.
Adapting a general language model to summarize medical depositions by tuning just a few extra layers.
A model that boils data down to a fixed number of parameters, no matter how many training examples it sees.
Parametric models are compact and fast to use, though their fixed size can limit how much detail they capture from large datasets.
Linear regression is parametric because it always summarizes the data with the same small set of coefficients.
A process that contains one or more smaller sub-processes, activities, or tasks within it.
Seeing the parent process helps people understand how individual steps fit into a larger workflow and where they belong.
A full case-preparation workflow is the parent process for sub-steps like document intake, indexing, and review.
An extra bit added to a block of data so a system can check whether the data was changed or corrupted.
Parity is a simple safeguard for detecting data errors during storage or transmission, helping keep records intact.
A parity bit lets a storage system notice when a stored document chunk has been accidentally altered.
A measure of how much a result changes when you tweak just one input while keeping the others fixed.
Partial derivatives are the building blocks of how models figure out which direction to adjust each parameter.
Calculating how much the prediction error shifts if only one weight changes, holding the rest steady.
Any contributor to a process, whether a computer system, data, business rule, or a person.
Listing every participant in a workflow clarifies who and what is responsible for each part, which helps with design and accountability.
In a document review flow, the OCR engine, the indexing rules, and the human reviewer are all participants.
Involving many different stakeholders in discussion and decision-making to reach agreement.
Bringing affected groups into the process helps AI systems reflect real needs and reduces blind spots that one team might miss.
Inviting attorneys, paralegals, and clients to shape how a document tool works is participation in its design.
A skew that appears when the people who choose to take part in data collection differ from those who don't.
It quietly distorts results, so conclusions about a whole population may only reflect the willing few.
A survey on courtroom experience that only volunteers answer may overrepresent people with strong opinions.
The rule that decides how a large model or its data is split across multiple machines for parallel training.
A good split keeps all the machines busy and balanced, which is essential for training huge models efficiently.
Choosing whether to split a model by layers or by data when training across a fleet of GPUs.
An agent that follows a fixed set of behaviors and only learns from what happens, without choosing its own actions.
Passive agents are useful for studying how well a fixed strategy performs, in contrast to active agents that explore and pick their own moves.
An agent that always follows a preset route and just records the rewards it earns is a passive learning agent.
A programming framework built for training extremely large neural networks across many accelerator chips.
It is the kind of tooling that makes training frontier-scale models practical for engineering teams.
A research team uses Pax to train a multi-billion-parameter language model spread over thousands of TPU chips.
The simplest kind of artificial neuron that takes several inputs, weighs them, and outputs a single yes-or-no style result.
It's the historical seed of modern neural networks — understanding it demystifies how deep learning began.
A perceptron could take a few features of an email and decide a single thing: spam or not spam.
How well a model does its job, measured either by prediction quality or by computational speed depending on context.
It's the catch-all word for whether a model is 'good,' so it's important to clarify which kind of performance is meant.
A document classifier with high performance accurately tags exhibits and does so fast enough for real-time review.
A method that ranks how much each input feature matters by scrambling it and seeing how much the model's accuracy drops.
It reveals which inputs a model actually relies on, helping you trust and explain its decisions.
Shuffling the 'document age' column to see if a case-outcome model really depends on it.
A score of how surprised a language model is by a piece of text — lower means the model predicts the words more confidently.
It's a standard way to measure how well a language model has learned, with lower perplexity signaling a better model.
A legal-domain language model shows lower perplexity on contract text than a general model does.
Any information that relates to a person who can be identified, directly or indirectly, such as a name, ID number, or location.
Personal data is protected by privacy laws, so handling it carelessly can cause real harm to people and legal trouble for organizations.
Names, addresses, and account numbers inside uploaded case documents are personal data that must be safeguarded.
The full sequence of automated steps that takes raw data through preparation, training, and deployment of a model.
A well-built pipeline makes machine learning repeatable and reliable instead of a one-off manual experiment.
A pipeline that ingests new deposition PDFs, cleans the text, retrains the classifier, and publishes it weekly.
Splitting a model across machines so different stages process different batches at the same time, like an assembly line.
It speeds up training of very large models by keeping every machine working instead of waiting its turn.
While one GPU handles the early layers of batch two, another finishes the later layers of batch one.
A JAX function that automatically spreads a computation across many accelerator chips for you.
It hides the hard plumbing of parallel computing, letting researchers scale models without rewriting low-level code.
Using pjit to run a large model's math across 64 TPU cores with a single function call.
A JAX function that runs the same computation on multiple chips at once, each handling a different slice of data.
It's a key tool for training on many accelerators in parallel, multiplying speed without extra effort.
Using pmap to train one model copy per GPU, each on a different batch of documents.
In a learning agent, the strategy that maps each situation to the action the agent should take.
The policy is essentially the agent's 'brain' — improving it is the whole goal of reinforcement learning.
A game-playing agent's policy tells it which move to make given the current board.
A family of methods that improve a learning agent by directly nudging its action strategy toward higher rewards.
It's a foundational technique behind agents that learn complex behaviors, including parts of how chatbots are tuned.
Adjusting a robot's control strategy step by step so it earns more reward for walking without falling.
A step in image models that shrinks a region of data down to a single summary value, like its maximum or average.
It makes vision models smaller and faster while keeping the important features, so they recognize objects efficiently.
Condensing a patch of a scanned exhibit image into one value to detect a signature regardless of exact position.
Extra information added to each word so a model knows the order words appear in, since it reads them all at once.
Without it, a Transformer would treat 'dog bites man' the same as 'man bites dog' — order is meaning.
Positional encoding lets a model understand that a date appears before, not after, a clause in a contract.
The viewpoints, assumptions, and circumstances a researcher brings to their work, which shape how the work is done and what comes out of it.
Recognizing positionality helps people see how their own perspective may bias data, design, or results, leading to fairer and more honest AI work.
A team building a hiring model reflects on how their own backgrounds might influence which signals they treat as important.
A way of making a hard-to-read model understandable after the fact, using methods like example-based, visual, or simplified explanations.
Post-hoc explanations let people make sense of complex models that weren't built to be transparent, supporting trust and review.
Highlighting which words in a document most influenced an AI's risk rating is a post-hoc explanation of an otherwise opaque model.
Adjusting a model's outputs after it makes predictions, often to make results fairer across different groups.
It's a practical lever for correcting bias without retraining the whole model.
Tuning the score threshold separately for two groups so a screening tool treats both equitably.
A method that reduces bias by adjusting a model's predicted labels after the model has made them.
Post-processing offers a way to make outcomes fairer without retraining the model or touching its training data.
After a model scores applicants, a post-processing step adjusts the decision thresholds so different groups are treated more equitably.
A single number summarizing how well a model balances precision and recall across all its decision thresholds.
It's especially useful when the thing you're looking for is rare, where plain accuracy can be misleading.
Measuring PR AUC for a tool flagging the few privileged documents hidden among thousands.
A way of judging whether a measured difference is large enough to actually matter in the real world, beyond just being statistically detectable.
A result can be statistically significant yet too small to mean anything in practice, so practical significance keeps decisions grounded in real impact.
In a discrimination case, analysts ask not only whether a hiring gap is statistically real but whether it is large enough to be practically meaningful.
A ready-to-use, prebuilt model component you can train without writing it yourself.
It speeds up building standard models by giving you tested components instead of starting from a blank page.
Using a prebuilt classifier estimator to sort documents into categories with minimal custom code.
A method that reduces bias by adjusting the training data before a model learns from it.
Fixing bias in the data itself can stop a model from learning unfair patterns in the first place.
Rebalancing a training set so under-represented groups appear more often is a pre-processing step to reduce bias.
A model already trained on a huge general dataset, ready to use or adapt rather than build from scratch.
Starting from a pre-trained model saves enormous time and data — it's the foundation of most modern AI tools.
Taking a pre-trained language model and adapting it to understand the vocabulary of expert reports.
The first, large-scale training stage where a model learns general patterns from a massive dataset.
It builds the broad knowledge a model later specializes; it's the costly step done once before fine-tuning.
A language model is pre-trained on the open internet before anyone adapts it for legal document review.
Out of everything a model flagged as positive, the share that was actually correct.
High precision means few false alarms — crucial when acting on a wrong flag is costly.
If a tool tags 100 documents as privileged and 95 truly are, its precision is 95 percent.
A graph showing the trade-off between catching more positives and keeping those flags accurate as you change the cutoff.
It helps you pick the right threshold for your situation — chasing completeness or chasing accuracy.
Plotting precision against recall to choose how aggressively a tool should flag potentially privileged files.
The output a model produces when you give it new input — its best guess based on what it learned.
Predictions are the entire point of a deployed model; everything else exists to make them accurate.
Feeding an unseen contract to a model and getting its prediction of which risk category it falls into.
A gap between the average of a model's predictions and the average of the real answers it should produce.
A large gap signals something is off, prompting you to investigate the data or the model before trusting it.
If a model predicts 30 percent of documents are relevant but the true rate is 50 percent, it has prediction bias.
Organizing analyses of structured and unstructured data to find patterns and correlations that can predict new situations.
Predictive analysis turns past data into useful foresight, helping people anticipate likely results before they happen.
Studying patterns across past case outcomes to estimate how a new matter might unfold is predictive analysis.
Reporting and insights that answer the question of what is likely to happen next, with a high degree of confidence.
Predictive analytics helps organizations plan ahead by forecasting future events instead of only reviewing the past.
A tool that estimates how long a document review will take based on similar past projects is using predictive analytics.
A fairness check that's satisfied when a model is equally accurate in its positive predictions for every group.
It's one formal definition of treating groups equally, useful when justifying that a model is fair.
A screening tool meets predictive parity if a 'high risk' label is correct at the same rate for each demographic.
Another name for the fairness condition where a model's positive predictions are equally accurate across groups.
Knowing the synonyms keeps you from being confused when different sources use different fairness vocabulary.
Confirming a risk tool's flagged cases come true at equal rates for two populations.
Cleaning and reshaping raw data into a tidy form a model can actually use before training or prediction.
Models are only as good as their inputs; careful preprocessing often matters more than the model choice.
Stripping headers, fixing OCR errors, and standardizing dates in scanned exhibits before feeding them to a model.
Analysis that goes beyond predicting what will happen and recommends what specific action you should take in response.
It turns data into a concrete next step, so people are not left guessing what to do once a problem or opportunity is spotted.
After flagging that a case backlog is growing, a tool suggests which filings to prioritize first to clear it fastest.
An assumption about what's likely true that you hold before seeing any data.
Stated priors make hidden assumptions explicit, which matters when data is scarce and assumptions sway results.
Before seeing returns, assuming most documents in a batch are not privileged is a prior belief.
A person's freedom from unwanted intrusion into their personal life or affairs.
Respecting privacy is both an ethical duty and a legal requirement when systems handle people's personal information.
A document AI tool redacts witnesses' home addresses before sharing a case file with a wider team.
Building privacy protections directly into a system from the start rather than bolting them on later.
Designing for privacy upfront avoids costly fixes and reduces the chance that personal data is exposed by accident.
A new intake tool is built to collect only the client details it truly needs, storing nothing extra by default.
Tools and methods that protect privacy by cutting down or hiding personal data while keeping a system useful.
These technologies let organizations get value from data without unnecessarily exposing the people behind it.
A tool blurs faces in case photos so investigators can review scenes without identifying bystanders.
A value of a sensitive characteristic that marks a group which has historically had a systematic advantage.
Spotting which groups have been favored helps reveal whether an AI system is repeating that unfair advantage.
When auditing a loan model, analysts compare outcomes for a historically advantaged group against others.
A model that predicts a number along with how uncertain it is, instead of giving a single bare figure.
Knowing the confidence around a prediction lets you act differently on a sure estimate than a shaky one.
Estimating a case's likely document count as '5,000, give or take 800' rather than just '5,000'.
A curve that shows how likely a continuous value is to fall near any given point.
It's the basic math language for describing uncertainty in numbers a model predicts.
A density function describing how likely a document's length is to be near various page counts.
A written, ordered set of steps for carrying out a particular process, activity, or task.
Clear procedures make work repeatable and consistent, so different people get the same reliable result.
A firm's procedure for handling new evidence lists each step from logging it to running it through the analysis tool.
A sequence of connected activities, events, and decisions that an organization runs to get work done.
Understanding work as a process lets teams measure, improve, and automate the steps that matter.
The case-review process moves a file through intake, analysis, and a final summary before sign-off.
A defined map of how a process is meant to run from start to finish, including its exceptions.
Seeing the full flow makes it easier to spot bottlenecks and decide where automation can help.
A diagram of the deposition prep flow shows each handoff and what happens when a document is missing.
Any operation performed on personal data, such as collecting, storing, using, sharing, or deleting it.
Privacy laws regulate processing, so knowing what counts helps organizations stay compliant when handling data.
Uploading client records, indexing them for search, and later deleting them are all forms of processing.
The combination of software and hardware that an application actually runs on.
The environment affects how reliably and securely a tool performs, so it must be accounted for when deploying.
A document AI tool tested on a laptop is then moved to its cloud processing environment for real cases.
A person or organization that handles personal data on behalf of, and under instructions from, a controller.
Knowing who is the processor versus the controller clarifies who is responsible for protecting the data.
A cloud vendor that stores a law firm's case files acts as the processor for that firm.
A professional who plans, builds, and launches AI-powered products and guides their ongoing success.
A product manager keeps an AI project tied to real user needs and business goals, not just technical novelty.
The product manager for a document tool decides which new features matter most to the legal teams using it.
The person who sets direction and priorities for an AI team and keeps everyone aligned to one vision.
A clear product owner prevents a diverse team from pulling in different directions and stalling progress.
The product owner decides the team should fix search accuracy before adding any new tool features.
Turning a top-performing model into a real, dependable product that people can use in live services.
A model only delivers value once it is hardened into something stable that handles real-world use.
A promising case-summary model is wrapped into a polished tool that lawyers can run on their own files.
Automatically analyzing personal data to evaluate or predict things about a person, like behavior, health, or preferences.
Profiling can affect important decisions about people, so it raises fairness and privacy concerns that need oversight.
A system that predicts which clients might miss a deadline based on past behavior is performing profiling.
The instruction or question you give an AI model to get a response.
The wording of a prompt strongly shapes the answer. Clear, specific prompts produce more reliable, on-point results.
"List every date mentioned in this report with the page number" is a prompt that returns a sourced timeline.
The craft of writing the instructions you give a model so it produces the output you want.
The same model can give great or useless answers depending on the prompt, so design is a core skill.
Rewording a request from 'tell me about this contract' to a structured prompt that returns risks and obligations.
Systematically refining the wording, examples, and structure of prompts to reliably get better model outputs.
It's how teams squeeze accurate, consistent results out of a model without changing the model itself.
Iterating on a prompt template until a model consistently extracts dates and parties from every contract.
Training a small set of learnable prompt values that steer a frozen model, instead of editing the model's own weights.
It customizes a big model for a task very cheaply while leaving the original model untouched.
Learning a compact prompt prefix that makes a general model reliably summarize medical records.
Getting a model to do a task purely by phrasing the request well, without retraining its internals.
It lets non-engineers steer powerful models just by writing instructions, dramatically lowering the bar to use AI.
Asking a model to 'summarize this deposition in five bullet points' and getting a usable result with no fine-tuning.
A characteristic, such as race or gender, that divides people into groups whose outcomes should be treated equally.
Tracking these attributes lets teams check whether an AI system treats different groups fairly.
Auditors compare a hiring model's results across gender to make sure no group is unfairly screened out.
A characteristic, like race or age, that law or policy says must not be used as the basis for decisions.
Using a protected class to make decisions can be illegal and unfair, so systems must avoid relying on it.
A screening tool is reviewed to confirm it never bases recommendations on an applicant's religion or age.
An early working model of a new product built to show its key features and how it performs.
A prototype lets a team test an idea and gather feedback before investing in a full build.
The team builds a rough prototype of a timeline tool to see if reviewers find the auto-generated dates useful.
Granting a person or machine the access rights and permissions they need within a system.
Careful provisioning ensures people can do their jobs while keeping sensitive systems and data secure.
A new paralegal is provisioned read-only access to the case archive but not permission to delete files.
An innocent-looking feature that secretly stands in for a sensitive trait like race or gender.
A model can discriminate even without the sensitive attribute, because a proxy leaks it in — a key fairness trap.
A ZIP code acting as a proxy for race, causing a model to treat groups unequally without ever seeing race.
Stand-in answers used to train a model when the true answer you really want isn't directly available.
They make training possible when ideal labels are missing, but a poor proxy can quietly teach the wrong thing.
Using 'document was later cited' as a proxy label for 'document was important' when importance isn't recorded.
Trimming away parts of a decision tree (or model) that add little value, to keep it simpler and avoid overfitting.
A pruned model is faster and generalizes better to new data instead of memorizing quirks of the training set.
Cutting back branches of a case-routing decision tree that only matched a handful of odd past files.
Processing personal data so it can no longer be tied to a specific person without separately held extra information.
It reduces privacy risk while still letting data be re-linked when genuinely needed, with the key kept safe and apart.
Client names in a dataset are replaced with codes, and the code-to-name key is stored separately under lock.
A scoring function that predicts how much total reward an AI agent will earn if it takes a particular action in a particular situation.
It is the core 'is this move worth it?' calculation that lets a learning agent choose good actions instead of random ones.
A document-triage agent uses a Q-function to estimate whether opening an exhibit now leads to a faster correct answer than reading the index first.
A reinforcement learning method where an agent learns the value of each action by trying things and updating its estimates from the rewards it gets.
It lets a system discover a good strategy through trial and error without anyone hand-coding the rules.
An agent learns the best order to search through case files by repeatedly trying sequences and keeping the ones that reach the answer fastest.
How well the features of a product or service meet the needs it is supposed to satisfy.
Quality determines whether users can actually trust and rely on a tool's output for real work.
The quality of a summary tool is judged by how accurately and completely it captures each document.
A cut point that divides a sorted set of values into equal-sized groups, like the median splitting data into two halves.
Quantiles describe how data is spread out, which helps you spot typical values versus extreme outliers.
Sorting 1,000 contract review times and finding the 90th-percentile quantile tells you the duration that only the slowest 10% exceed.
Grouping numeric values into buckets that each hold roughly the same number of items, instead of buckets of equal width.
It keeps every bucket meaningfully populated even when the data is lopsided, so a model treats common and rare ranges fairly.
Sorting documents by page count into ten equal-sized groups so each bucket has the same number of files, not the same page range.
Shrinking a model by storing its numbers at lower precision so it uses less memory and runs faster.
It makes large models cheaper and quick enough to run on smaller hardware, usually with only a tiny drop in accuracy.
A quantized language model runs on a laptop to summarize depositions instead of needing an expensive server GPU.
A waiting line data structure that holds items and releases them in the order they arrived, first in, first out.
Queues let a system buffer incoming work so requests are processed in order without overwhelming the model.
Hundreds of uploaded contracts wait in a queue and get analyzed one batch at a time as capacity frees up.
The social process by which a group is assigned a racial identity, whether or not its members see themselves that way.
Recognizing how groups become racialized helps reveal hidden bias in how data and AI categorize people.
Researchers examine whether a model's categories reflect socially imposed labels rather than how people self-identify.
A method where the AI retrieves relevant documents first, then answers using only that retrieved text.
RAG is how a tool answers from YOUR files instead of guessing from memory — the main defense against hallucination.
Ask a question across 500 deposition PDFs and get a source-linked answer drawn only from the matching passages.
A grid of data where the rows can have different lengths instead of all being the same size.
Real text data is uneven — sentences have different word counts — so ragged tensors let models handle varying-length inputs without forcing everything into one fixed shape.
Storing 100 deposition answers where one is 5 words and another is 500 words, without padding the short ones to match.
A prediction method that builds many decision trees on random slices of the data and averages their votes.
Combining many trees is far more accurate and stable than one tree, and it resists overfitting — a workhorse for everyday business prediction.
Predicting whether a contract clause is high-risk by polling hundreds of trees, each trained on a random subset of past clauses.
A strategy where an AI agent picks its next action completely at random instead of based on what it has learned.
It is the baseline starting point in reinforcement learning — useful for exploring options early and for measuring whether a trained policy is actually better than dumb luck.
A game-playing agent making purely random moves on its first attempt before it has learned anything.
An item's position in a sorted list, like 1st, 2nd, or 3rd, where the order matters but the gaps between positions don't.
Many AI results are about ordering things — most relevant document first — and rank captures that ordering without pretending the spacing between items is meaningful.
Listing search hits as the 1st, 2nd, and 3rd most relevant exhibits to a query.
The number of dimensions a block of numbers has — a single number is rank 0, a list is rank 1, a grid is rank 2, and so on.
Knowing a tensor's rank tells you how the data is shaped, which matters when wiring data into a model so the shapes line up.
A grayscale image is a rank-2 tensor (height by width); a color image is rank 3 (height, width, color channels).
A table of numbers where some rows or columns are just repeats or combinations of others, so it carries less independent information than its size suggests.
Redundant data can make a model's math unstable or unsolvable, so spotting rank deficiency helps explain why some models fail to train cleanly.
A spreadsheet column that is always exactly twice another column adds no new information.
A type of machine learning that orders items by how relevant they are to a query or user.
Good ranking surfaces the most useful results first, which is the heart of search and recommendations.
A document search tool ranks the passages most relevant to a question at the top of the results list.
A person who reviews and labels data or judges how good an AI's answers are.
Human raters create the gold-standard labels that models learn from and are measured against — the quality of an AI often traces back to the quality of its raters.
A reviewer marking whether each AI-generated case summary is accurate, partial, or wrong.
Software that gives a human operator preset, automated steps to run tasks across several programs during their own workflow.
RDA speeds up repetitive desktop work by letting a person trigger automated sequences instead of doing each click by hand.
A clerk clicking one button that copies case details from an email into three separate systems.
A second pass that reorders an initial list of search results to put the most relevant ones on top.
Fast first-pass search casts a wide net; re-ranking carefully sorts the catch so the best matches surface first, sharpening RAG answers.
After pulling 100 candidate passages, a re-ranker decides which 5 most directly answer the question.
Models built to think through a problem in steps before answering, instead of replying with the first thing that comes out.
They do far better on multi-step tasks — math, logic, untangling a complex argument — because they spend extra effort working it out. The trade-off is they are slower and cost more per answer.
Asked whether two depositions contradict each other, a reasoning model lines up each witness's statements step by step, then concludes — rather than guessing yes or no instantly.
Out of all the things that truly should have been found, the fraction the model actually caught.
High recall means few important items slip through — critical when missing one matters, like overlooking a key document in discovery.
If 100 documents are privileged and the model flags 92 of them, recall is 92%.
Out of all the truly relevant items, what fraction appear in the top k results the system returns.
Users only look at the first few results, so this measures whether the right items make it into that short visible list.
If 10 documents are relevant and 7 of them land in the top 20 search results, recall at 20 is 70%.
Automatically finding patterns in data with algorithms and using them to act, such as sorting items into categories.
Recognition is the basis for many AI tasks, from reading scanned text to spotting objects in images.
Optical character recognition turns a scanned contract into searchable, editable text.
A system that suggests items a person is likely to want based on their past behavior and what similar people liked.
It is the engine behind 'you might also like' suggestions and can surface relevant documents or precedents you didn't know to search for.
Suggesting related expert reports based on the ones a paralegal has already opened.
A person's right to have inaccurate personal data about them corrected or completed by the organization holding it.
Letting people fix wrong data protects them from decisions made on faulty information.
A client notices their recorded address is wrong and asks the firm to correct it in the case system.
A simple rule inside a neural network that passes positive numbers through unchanged and turns negative numbers into zero.
This tiny on/off rule is what lets neural networks learn complex patterns efficiently; it is the most common building block in modern models.
A network layer that keeps a value of 3 but replaces a value of -2 with 0 before passing it forward.
A neural network that processes a sequence one step at a time while remembering what came before.
It was the early standard for understanding text and speech where order matters; knowing it helps explain why Transformers later replaced it.
Reading a witness statement word by word, keeping context from earlier words to interpret later ones.
A group authorized to act like an attacker and probe a system's defenses to find weaknesses.
Red teams expose real vulnerabilities before genuine adversaries can, making systems safer and more robust.
A red team tries to trick a document AI into leaking confidential case data so the gap can be fixed first.
A class that describes object identifiers, whose instances stand in for objects by representing them indirectly.
Reference classes let software point to and manage objects without copying the objects themselves around.
In a system tracking exhibits, a reference points to an exhibit record instead of duplicating all its details.
Critical self-examination that asks why and how research is done, questioning its purpose, ethics, and value.
Reflexivity helps researchers and AI teams catch their own assumptions and biases before they shape results.
A team building a fairness tool pauses to ask whose interests their chosen metrics actually serve.
A type of supervised learning that predicts a number on a continuous scale rather than a yes/no category.
Regression is how models answer 'how much' or 'how many' questions, which underpins forecasting and any prediction that lands on a sliding scale.
Estimating the likely repair cost of a building from its age, size, and damage measurements.
A model that predicts a number, like a price or a duration, rather than a category.
Whenever you need to forecast a quantity instead of a yes/no label, regression is the tool — it underlies most everyday business predictions.
Estimating how many days a case will take to resolve based on its features.
A technique that gently penalizes a model for being too complex so it doesn't just memorize the training data.
It is the main defense against overfitting, helping a model stay accurate on new data it has never seen.
Discouraging a model from relying too heavily on any single quirky feature in the training documents.
A dial that controls how strongly a model is penalized for being too complex.
Tuning this number balances fitting the training data well against staying simple enough to work on new data.
Turning the dial up forces a fraud-detection model to use fewer, more reliable signals.
A way of training an AI by letting it try actions and rewarding the good outcomes so it learns by trial and error.
It is how AI learns tasks with no fixed answer key — from game-playing to aligning chatbots with human preferences.
An agent learning the best sequence of negotiation moves by getting rewarded for favorable settlements.
Reinforcement Learning from Human Feedback
Training a language model to behave better by having people rank its answers and rewarding it for the ones people prefer.
RLHF is the key step that makes raw language models helpful, honest, and polite instead of just plausible-sounding — it is how modern chatbots got usable.
People rate two draft answers from a legal assistant, and the model learns to favor the clearer, more accurate style.
How well a retrieved item actually matches what the user was asking about.
A search or RAG system is only as good as the relevance of what it pulls back; irrelevant results lead to wrong or useless answers.
Judging whether a passage a search tool returned genuinely addresses the question about a witness's whereabouts.
How closely repeated estimates agree with the first estimate, showing whether a result holds steady over time.
Reliable systems give consistent answers, so users can trust that a repeated measurement or prediction will not drift unexpectedly.
Running the same document analysis twice and getting essentially the same extracted dates each time.
The detailed process of cleaning, organizing, updating, and safely storing data, including deleting what is no longer needed.
Good data remediation keeps a system's records accurate and secure, which directly improves how well AI tools can use that data.
Cleaning a messy case file folder by fixing mislabeled PDFs, deleting duplicates, and moving the rest to a secure store.
A copy of a model or part of a training job running on separate hardware so work can happen in parallel.
Replicas let large models train and serve faster by splitting the load across many machines.
Running 8 copies of a model across 8 chips to train on a huge document set in a fraction of the time.
The tendency for written records to mention unusual events far more often than ordinary ones, skewing what a model learns is common.
Models trained on text can wrongly assume rare dramatic events are normal because people write about them more, distorting predictions.
Text mentions 'murder' far more than 'breathing,' so a naive model may overestimate how common violent crime is.
The way raw data is converted into numbers that a model can actually work with.
The quality of how you represent data often matters more than the model itself — good representations make patterns easy to find.
Turning a contract's words into numeric features so a model can compare it to other contracts.
How closely results agree when the same thing is measured again under changed conditions.
Reproducible results show that an outcome is real and not a fluke of one particular setup, which is essential for trust and review.
Another analyst re-running an experiment on a different machine and reaching the same conclusion.
Something that must be present or satisfied for a goal, system, or outcome to be achieved.
Clear requirements define what a tool must do before it is built, preventing wasted effort and mismatched expectations.
Stating that a document tool must accept PDFs over 32 MB before development begins.
The gap between what a model predicted and what the data actually showed, capturing the part the model could not explain.
Examining residuals reveals where a model is going wrong and how much of the real pattern it is missing.
If a model predicts a 5-day delay but the real delay was 8 days, the residual is 3 days.
A system's ability to prepare for, withstand, and quickly recover from disruptions, attacks, or changing conditions.
Resilient systems keep working through accidents and attacks, so a single problem does not cause a complete failure.
A document service that keeps serving files even when one of its backend servers crashes.
Building and using AI in a way that aligns with shared goals, values, and democratic principles.
Responsible AI guards against harm and misuse, helping ensure tools serve people fairly and ethically.
Choosing not to deploy an AI feature until its fairness and privacy effects have been reviewed.
The outcome you get once a process has been carried out to completion.
Defining the expected result keeps work focused on a concrete deliverable rather than effort for its own sake.
The summarized timeline a tool produces after finishing its analysis of a set of documents.
How much data is kept long-term, measured by storage size and by how many months or years it is stored.
Setting retention limits balances having records available against privacy and storage costs, and is often a legal requirement.
A policy that automatically deletes uploaded case files 90 days after a matter closes.
The step of finding and pulling the most relevant pieces of information from a large collection.
Retrieval is the search engine inside RAG systems — get it wrong and the AI answers from the wrong material.
Searching across 10,000 case files to fetch the passages that mention a specific clause.
An approach where the AI first looks up relevant documents and then writes its answer using that retrieved text.
It grounds answers in real source material instead of the model's memory, which is the main defense against made-up facts.
Pulling the three most relevant deposition excerpts before drafting a source-backed summary.
The total reward an agent expects to collect over the rest of an episode, often counting near-term rewards more than distant ones.
Agents aim to maximize long-term return, not just the next reward — this is what makes them plan ahead instead of acting greedily.
A negotiation agent accepting a smaller immediate gain because it leads to a much larger total payoff later.
A score an AI agent receives after an action, telling it whether that action was good or bad.
Rewards are the signal that drives all reinforcement learning — the agent's whole goal is to collect as much reward as possible.
Giving a negotiation agent points for reaching a favorable settlement and penalties for impasse.
A model trained to predict how much humans would like a given AI response, used to guide further training.
It lets a system improve without a human grading every single answer — the reward model stands in for human judgment at scale.
A model that scores draft chatbot replies the way human reviewers would, so the chatbot can be trained on those scores.
A way to keep a model simple by penalizing large values across all its internal weights, shrinking them toward zero.
It curbs overfitting while keeping every feature in play, making predictions steadier on new data; it is another name for L2 regularization.
Stopping a price-prediction model from leaning too hard on any one feature by gently shrinking all its weights.
A combined measure of how likely an event is and how big its consequences would be, which can be positive or negative.
Understanding risk lets teams weigh potential harms and opportunities before acting, so decisions are made with eyes open.
Weighing the chance an AI tool misreads a document against the impact such an error could have on a case.
Measures taken during design, build, and evaluation to manage security, economic, and performance risks in AI systems.
Risk controls are the concrete safeguards that turn responsible-AI intentions into protections against bias, errors, and breaches.
Adding human review of AI outputs to catch errors before they reach a final report.
How much risk an organization or team is willing to accept in order to reach its goals.
Risk tolerance sets the line for what is acceptable, guiding which AI uses are approved and which are held back.
A firm deciding it will not let AI draft client-facing documents without attorney sign-off.
A single number, in the same units as what you are measuring, that summarizes how far an estimate typically is from the true value.
RMSD gives an at-a-glance sense of a model's typical error size, making it easy to compare how accurate different models are.
Reporting that a delay-prediction model is off by about 2 days on average.
Short for recurrent neural network — a model that reads a sequence step by step while remembering earlier steps.
RNNs were the original approach for text and speech tasks and are useful background for understanding why Transformers took over.
Processing a sentence one word at a time, using earlier words to make sense of later ones.
An AI system that keeps working well in messy real-world conditions and resists attempts to fool it.
Robust AI stays reliable when inputs change unexpectedly or when someone tries to trick it, which matters for safety-critical use.
A document reader that still extracts text correctly from a faded, slightly skewed scan.
A system's ability to keep performing well across a wide variety of conditions.
Robustness shows whether a tool only works in ideal cases or holds up when faced with the variety of real inputs.
An OCR model that reads many different fonts, layouts, and scan qualities without large drops in accuracy.
A graph showing the trade-off between catching true positives and raising false alarms as you change a model's decision threshold.
It shows at a glance how well a yes/no model separates the two classes across all possible cutoffs, helping pick the right threshold.
Plotting how a privilege-detection model's hits versus false flags shift as you make it stricter or looser.
The top-level folder where a project stores its model checkpoints and related files.
Knowing where saved model versions live matters for reloading the right one and avoiding overwriting good results.
The main folder holding every saved snapshot of a model trained on contract data.
A single number summarizing how far off a model's numeric predictions are on average, in the same units as what it predicts.
It is the go-to score for number-predicting models, and because it punishes big misses heavily, it flags models that occasionally go badly wrong.
An RMSE of 4 days when predicting case duration means typical predictions are about 4 days off.
A model's ability to recognize an object no matter how it is rotated in the image.
Documents and objects appear at different angles, so this property lets a vision model identify them whether they are upright, sideways, or tilted.
An image classifier still recognizing a signature stamp whether the scanned page was rotated 90 degrees or upside down.
A single entity or observation in a data table, with columns describing its properties.
Each row is one example the model can learn from, so more rows generally mean more cases to learn the problem.
In a spreadsheet of documents, one row holds the filename, date, and page count for a single PDF.
Preconfigured software that uses business rules to run sequences of tasks across multiple systems on its own, with people stepping in only for exceptions.
RPA automates routine, rule-based work end to end, freeing people from repetitive multi-system data entry.
A bot that automatically pulls new filings from a portal and files them into a case management system overnight.
A system property meaning it will not, under defined conditions, put human life, health, property, or the environment in danger.
Safety focuses on reducing both expected harms and unexpected ones, which is central to deploying AI responsibly.
Designing an AI tool so it flags uncertainty instead of confidently stating a wrong conclusion that could mislead a decision.
A flaw where the data you collected does not represent the real population because of how it was gathered.
If your sample skews one way, the model learns a distorted picture and makes unfair or wrong predictions for the missing groups.
Training a settlement-prediction model only on cases that went to trial misses the majority that settled, skewing every estimate.
The ability to add or remove computing resources to match how much work needs to be done.
Scalable systems handle bursts of demand without slowing down or wasting money on idle capacity when demand is low.
A document tool spinning up extra servers when many users upload large files at once.
Adjusting numeric features so they share a common range, like 0 to 1, so no single feature dominates by sheer size.
Without scaling, a feature measured in millions can drown out one measured in fractions, slowing or breaking learning.
Before modeling, contract values in dollars and contract counts get rescaled so both contribute evenly.
The observed pattern that model performance improves predictably as you add more data, parameters, and compute.
Scaling laws explain why bigger models trained on more text keep getting better, guiding how labs spend training budgets.
A lab uses scaling laws to forecast that doubling training data will cut a model's error by a known amount before spending the money.
A popular free Python library with ready-made tools for common machine learning tasks.
It lets people build and test models without writing the math from scratch, making ML accessible to beginners.
A data analyst trains a quick classifier on case-outcome data in a few lines of scikit-learn before scaling up.
A single number an evaluator assigns to rate the quality of a model's output.
Scores turn fuzzy quality into a comparable metric so teams can tell which model or prompt is better.
An autorater gives each generated case summary a score from 1 to 5 for accuracy and completeness.
The step where a system assigns a number to each candidate item to rank how good or relevant it is.
Scoring decides what rises to the top of a recommendation or search list, directly shaping what users see.
After filtering to relevant exhibits, the system scores each one so the most pertinent appears first.
When a disability lowers someone's score on a selection test and they lose an opportunity as a result.
Screen-out is a form of discrimination automated screening can cause, so fairness reviews must check for it.
An automated resume filter rejecting a qualified candidate because a screen reader formatted their resume differently.
A system's resistance to deliberate, unauthorized actions meant to harm or damage it.
Security protects sensitive data and system integrity against intentional attacks, which is vital when handling confidential documents.
Encrypting uploaded case files so an intruder who breaches the server cannot read them.
The process of finding groups of similar records within a data table.
Segmentation reveals natural groupings in data, helping teams tailor analysis or handling to each type of record.
Sorting a pile of documents into clusters like contracts, emails, and reports based on their content.
A distortion that arises when the items chosen for your dataset are not picked at random from the whole population.
It quietly skews conclusions, so a model can look accurate on your data yet fail on the cases you never sampled.
Reviewing only flagged documents teaches a model what flagged documents look like, not what all documents look like.
A mechanism that lets each word in a sentence look at every other word to understand context.
Self-attention is the core trick behind modern language models, letting them grasp meaning that depends on distant words.
Reading a clause, the model uses self-attention to tie 'it' back to the right party named earlier in the sentence.
A computing system that understands enough about its own inputs, processing, and outputs to judge and improve itself toward its purpose.
Self-awareness lets a system catch its own mistakes and adapt, reducing the need for constant human oversight.
A pipeline that notices its accuracy is dropping and flags itself for retraining.
A system's ability to read its own sensor data, check that the data is valid, and report on its processes and results.
Self-diagnosis lets a system spot and communicate its own faults early, before they cause bigger failures.
A scanning tool that detects a clogged sensor and warns operators that recent scans may be low quality.
A system that can sense when it is not running correctly and fix itself without anyone stepping in.
Self-healing systems recover from problems automatically, reducing downtime and the need for manual intervention.
A service that detects a stuck process and restarts it on its own to restore normal operation.
A way to train models on raw unlabeled data by having them predict hidden parts of the input.
It unlocks learning from huge amounts of text or images that nobody had time to label by hand.
A model learns language by repeatedly guessing the next word across millions of documents, no human labels needed.
A technique where a model labels new data itself and then trains further on its own confident guesses.
It stretches a small set of labeled examples by leaning on plentiful unlabeled data, saving labeling effort.
A document classifier labels thousands of unmarked files, keeps only the high-confidence ones, and retrains on them.
A framework of metadata labels applied to data so machine learning algorithms can find relationships between pieces of information.
A semantic map gives raw data structure and meaning, guiding AI toward the connections that matter when new data is added.
Tagging every document field as a party, date, or claim so a tool can link related facts across files.
Search that finds results by meaning rather than by matching exact keywords.
It surfaces the right passage even when it uses different words than your query, the foundation of modern document AI.
Searching 'who paid for repairs' finds a clause about 'reimbursement of restoration costs' despite no shared words.
Training that mixes a small amount of labeled data with a large amount of unlabeled data.
Labeling is expensive, so combining a few labeled examples with many unlabeled ones can boost accuracy cheaply.
A few hundred hand-tagged contracts plus thousands of untagged ones together train a stronger clause detector.
A personal characteristic, like race, gender, or age, that deserves special legal or ethical care in a model.
Models that key off sensitive attributes can discriminate, so these features need careful handling to stay fair and lawful.
An auditor checks that a screening model does not give different outcomes based on an applicant's age or gender.
A what-if test that checks how much the final results change when you tweak one of the inputs or assumptions.
It tells you which settings really drive an outcome, so you know what to double-check and what is safe to leave alone.
An AI document classifier's accuracy is tested while nudging its confidence cutoff up and down to see which threshold setting it depends on most.
Turning human-style sensory input like sight or speech into a digital form a computer can read and process.
It is the bridge that lets machines work with things people see or hear, which is the whole basis of tools like OCR and speech-to-text.
Scanning a stack of handwritten deposition exhibits and converting them to searchable text through OCR.
Automatically judging whether a piece of text expresses a positive, negative, or neutral feeling.
It turns mountains of free-text opinions into a quick read on attitude, useful for triage and trend-spotting.
Scanning witness statements to flag which ones carry hostile versus cooperative tone.
Any task that turns one sequence, like a sentence, into another sequence, like its translation.
It frames translation, summarization, and answering as input-text-to-output-text, the shape most language tools use.
Feeding a long deposition transcript in and getting a short summary out is a sequence-to-sequence task.
A way of running AI in the cloud where you just call an API and the provider handles all the servers and scaling for you behind the scenes.
It lets a team use a model without buying, configuring, or maintaining hardware — you pay for what you use and the infrastructure is invisible.
A small firm could add document summarization to its app by calling a serverless AI endpoint, with no servers of its own to manage.
A set of coordinated steps that takes some input, adds value by transforming it, and produces an output a customer needs.
Framing software as a service keeps the focus on the useful result a user gets rather than the internal mechanics.
A document-summary service accepts a long PDF and returns a short brief.
Running a trained model in production so it can answer real requests from users or other systems.
A model only delivers value once it is serving live queries; serving is the bridge from lab to product.
After training, the document classifier is deployed for serving so it can tag new uploads instantly.
Machine learning approaches where you first hand-pick the useful features and only then let the model learn from them, as separate steps.
It contrasts with deep learning, which figures out its own features, so knowing the difference explains why some models need more manual setup.
Manually extracting word counts from documents before feeding them to a simple classifier, rather than letting a neural network learn the patterns itself.
The dimensions of a data array, such as how many rows and columns it has.
Mismatched shapes are a top cause of errors when feeding data into a model, so knowing shape prevents crashes.
A batch of 32 images sized 28 by 28 pixels has the shape 32 by 28 by 28.
A setting in boosted trees that scales down each new tree's contribution so the model learns slowly and steadily.
Smaller steps reduce overfitting and usually produce a more accurate final model, at the cost of needing more trees.
Lowering the shrinkage value forces a gradient-boosted model to take gentler corrections each round.
A math curve that squashes any number into a value between 0 and 1, often read as a probability.
It converts a model's raw score into a clean 0-to-1 probability, the basis of yes/no predictions.
A model's raw output of 2.5 passes through a sigmoid to become roughly a 92 percent chance of 'relevant'.
A way of analyzing accuracy that separates genuine correct catches from false alarms when deciding whether something is present.
It helps measure how well a system distinguishes a real signal from background noise, which matters whenever there is a yes/no judgment under uncertainty.
Evaluating a fraud-detection model by tracking both the real fraud it catches and the honest transactions it wrongly flags.
A score that measures how well each item fits its assigned cluster versus other clusters.
It helps judge whether a clustering actually found meaningful groups or just split things arbitrarily.
A high silhouette score confirms that documents grouped by topic genuinely cluster together rather than overlapping.
A formula that scores how alike two items are, often two meaning-vectors.
Search and recommendation rely on similarity measures to find the items closest in meaning to your query.
Cosine similarity compares your question's vector to each passage's vector to surface the closest matches.
Noticing what is happening in a system or its surroundings, understanding what it means, and anticipating what might come next.
A system or operator with good situational awareness can react to changes and risks in time instead of being caught off guard.
A monitoring dashboard that shows an AI tool's current load and flags when error rates start climbing so an operator can step in.
A model's ability to recognize an object whether it appears large or small in the image.
Real photos show objects at many sizes, so a useful vision model must not be fooled by scale.
A signature detector spots the same mark whether it is a full-page flourish or a tiny initial.
A shortcut that compresses huge data into a smaller summary to speed up similarity calculations.
It makes finding near-matches across massive datasets fast enough to be practical.
Sketching lets a system narrow billions of document vectors to a handful of likely matches in milliseconds.
The view that technology and the people using it form one connected system shaped by the wider society around it.
It reminds us that an AI tool's real-world effects depend on human and social context, not just the code, so design has to account for both.
Rolling out a legal AI assistant means considering not only the model's accuracy but also how attorneys trust it, verify it, and fit it into their workflow.
A function that turns a list of raw scores into probabilities that add up to 100 percent.
It is how a model expresses confidence across several possible answers when only one is correct.
For a document that must be one category, softmax outputs 70 percent contract, 20 percent memo, 10 percent email.
Running a system under set conditions, recording what happens, and judging whether it behaves as it should.
Testing is how teams catch problems before users do and gain confidence that software works as intended.
Running a document-extraction tool against sample contracts with known answers to confirm it pulls the right dates and parties.
A feature whose values are almost always zero, with only a few nonzero entries.
Sparse features are common with text and categories, and handling them efficiently saves huge amounts of memory.
A column marking which of 50,000 possible legal terms appear in a document is mostly zeros for any single file.
A memory-saving way to store data that records only the nonzero values and their positions.
It lets systems handle enormous mostly-empty datasets without wasting space on all the zeros.
Instead of storing 50,000 zeros, the system notes only the three terms a document actually contains.
A list of numbers that is mostly zeros, with only a few meaningful nonzero entries.
Sparse vectors power keyword search and are paired with dense vectors in hybrid search for better recall.
A keyword vector lights up only the few terms a query mentions, leaving thousands of slots at zero.
The fraction of a dataset or model that is zero rather than carrying a value.
High sparsity is both a storage opportunity and a modeling challenge, since most signal hides in few entries.
A term-by-document grid for a case file is highly sparse because each document uses only a tiny slice of the vocabulary.
A step in image models that shrinks a region of the picture down to a single summary value.
Pooling cuts computation and makes recognition more robust to small shifts in where an object sits.
After detecting edges, pooling condenses each patch so the network focuses on the strongest signals.
A document that spells out, completely and precisely, what a system must do and how to check that it does it.
A clear specification gives everyone a shared, testable definition of done, reducing guesswork and disputes about whether the system works.
A spec stating that a redaction tool must mask every Social Security number in a filing, with a checklist to verify each case.
A variant of hinge loss that squares the penalty, punishing confident wrong answers more steeply.
It trains classifiers that not only get the right side but stay safely clear of the decision boundary.
A support vector classifier uses squared hinge loss to push borderline document labels further from the dividing line.
A way to measure error by squaring the gap between a prediction and the true value.
Squaring punishes big misses harder than small ones, steering models to avoid large errors.
Predicting a settlement of 100k when the truth is 130k yields a squared loss of 30k times 30k.
Training a model in phases, often starting simple and adding complexity as it learns.
Building up gradually can make training faster and more stable than tackling the full problem at once.
A network first trains on a few layers, then more layers are added in later stages.
Any person, group, or organization that has an interest in a system or is affected by what it does.
Identifying stakeholders early ensures a project meets real needs and accounts for everyone it touches, not just the people building it.
Stakeholders for a case-management AI tool include the attorneys using it, the clients whose data it holds, and the firm's IT and compliance staff.
A number that says how spread out a set of values is around their average.
It tells you whether your data clusters tightly or varies wildly, which shapes how much to trust an average.
Two case datasets can share the same average award yet differ sharply in standard deviation, meaning one is far more variable.
The activity, task, or input that marks the beginning of a process.
Knowing exactly what kicks off a process makes workflows easier to map, automate, and troubleshoot.
Uploading a new case file is the start event that triggers an automated document-indexing workflow.
Making predictions in advance in bulk and storing them for instant lookup later.
It delivers answers with near-zero wait at request time, ideal when the same predictions are reused often.
Overnight, a system scores every document for relevance so morning users get results with no delay.
A model that is trained once and then used as-is, without ongoing updates.
It is simple to manage but can grow stale as the world it learned about changes.
A clause classifier trained last year keeps running unchanged even as new contract language appears.
The property that a dataset's statistical patterns stay the same over time.
Models assume the future resembles the past, so a loss of stationarity quietly erodes accuracy.
If filing patterns shift after a new law, the data is no longer stationary and the old model drifts.
A result is statistically significant when it is unlikely to have happened just by random chance.
It helps you tell a real effect apart from a fluke, so you do not act on patterns that are only noise.
An A/B test shows a new model version answers more accurately, and statistical significance confirms the gain is real and not just luck in the sample.
Numerical data about groups, and the science of collecting, analyzing, and interpreting that data.
Statistics is the foundation for measuring how systems perform and for drawing trustworthy conclusions from data.
Computing the average and spread of a model's accuracy across thousands of test documents.
How easily a user can guide or control an AI model's output so it matches a desired style, format, or behavior.
A steerable model does what you ask without constant fighting — staying on tone, length, or scope — which makes it dependable for real work.
Telling an assistant to answer only in short, plain-English bullet points and having it consistently obey is good steerability.
One update to a model's settings during training, usually after processing a batch of examples.
Training progresses step by step, and counting steps is how practitioners track and schedule learning.
After each step, the model nudges its weights slightly to reduce error on the latest batch.
How big a change a model makes to its settings on each training update.
Too big and training overshoots; too small and it crawls, so step size strongly affects success.
Cutting the step size in half makes the model adjust more cautiously and avoid wild swings.
A fixed set of generalized beliefs about the members of a group that oversimplifies them and often resists being corrected.
AI systems can absorb stereotypes from their training data and repeat unfair generalizations, so spotting them is key to fair outcomes.
A resume-screening model that quietly favors one group because the historical data it learned from carried a stereotype.
Describes something that includes randomness, so its behavior or outcome varies rather than being fixed every time.
Many AI methods rely on randomness, which is why the same model can give slightly different answers on different runs.
A language model can produce two different wordings for the same prompt because its output generation is stochastic.
A training method that updates the model using one small random batch of examples at a time.
By not waiting for the whole dataset each step, it trains huge models far faster and is the workhorse of deep learning.
Rather than reviewing all 500 documents before adjusting, the model tweaks itself after each small random handful.
Completing a whole transaction or process automatically from start to finish through connected software, with no manual steps in between.
It removes hand-offs and waiting, making processes faster and less error-prone.
A filing is uploaded, automatically classified, indexed, and routed to the right matter without anyone touching it along the way.
A flawed argument that attacks a distorted, easier version of someone's position and then claims to have beaten their real one.
Recognizing this trick helps you evaluate reasoning honestly, including arguments an AI might generate or summarize.
Asking an AI to critique a brief and getting back an objection that misstates the brief's actual claim is a strawperson response.
A test that pushes a system beyond its normal limits, like very heavy load or scarce resources, to see how it holds up.
It reveals where a system breaks under pressure so you can fix weak points before they fail in real use.
Sending thousands of documents at once to a search tool to confirm it stays responsive during a big case dump.
How far an image filter slides across the picture between each step.
A bigger stride scans faster and shrinks the output, trading detail for speed.
A stride of two moves the scanning window two pixels at a time, halving the work and the output size.
A training goal that balances fitting the data well against keeping the model simple.
It guards against overfitting by penalizing needless complexity, so the model generalizes to new cases.
The training process accepts slightly higher error on known documents in exchange for a simpler, more reliable model.
Data that follows a predefined model and is organized in a consistent, predictable way, like rows and columns in a table.
Its tidy organization makes it easy for computers to store, search, and analyze without extra interpretation.
A case database table with one row per exhibit and columns for date, party, and document type.
A smaller process nested inside a larger one, which can appear or repeat within several parent processes.
Breaking a big workflow into reusable sub-processes makes it easier to design, maintain, and reuse common steps.
An OCR step is a sub-process reused inside both the intake workflow and the search-indexing workflow.
Using a smaller random portion of the data instead of all of it.
It speeds up experiments and can reduce the influence of overly common examples.
To test a pipeline quickly, an engineer trains on a subsample of 5,000 documents instead of the full million.
In reinforcement learning, the new decision strategy a system adopts after an update.
Each successor policy is the agent's improved playbook, and the cycle of replacing policies is how it learns to act better.
After a round of training, the game-playing agent switches to its successor policy that wins more often.
A logged value, like accuracy or loss, recorded during training so you can watch progress.
Summaries are how teams monitor whether a model is learning and catch problems early.
Charts of loss summaries reveal that training stalled after the tenth epoch.
Training a model on examples that come with the correct answers already attached.
It is the most common form of ML, learning to predict labels by studying many labeled examples.
Showing a model thousands of documents pre-tagged 'privileged' or 'not' so it learns to tag new ones.
A classic algorithm that separates two groups by drawing the widest possible gap between them.
It is a reliable, well-understood classifier that works well even with limited data.
A support vector machine cleanly divides relevant from irrelevant documents by maximizing the margin between the two piles.
Artificial data generated by a model to imitate real data, used for training when real examples are scarce, private, or expensive to collect.
It lets teams train and test systems without exposing sensitive real records, and fills gaps where real examples are rare.
Instead of training on real client files, a team might generate fake-but-realistic case documents that carry no confidential information.
A new input created by combining or transforming existing features rather than measured directly.
Well-chosen synthetic features hand the model useful patterns it would struggle to discover on its own.
Dividing total pages by number of exhibits creates a synthetic 'density' feature that helps flag complex filings.
A set of interacting parts arranged together to accomplish one or more stated goals.
Thinking in systems helps you see how components combine to produce results that no single part delivers alone.
A legal AI tool is a system combining a document store, a search engine, and a language model to answer questions from files.
A behind-the-scenes message that sets an AI assistant's role, rules, and tone before the user types anything.
It is the main lever for shaping how a model behaves, keeping answers on-task and within guardrails.
A system instruction tells the assistant to answer only from provided case files and cite every source.
A larger setup where several independent systems work together to deliver a capability none of them could provide alone.
Many real products are built this way, so understanding the term clarifies how separate tools combine into something bigger.
An AI legal toolbox where the query, timeline, and critique tools each run on their own but together support a full case review.
Bias built into the routine rules and practices of institutions that quietly advantages some groups and disadvantages others, even without anyone intending it.
Because it hides in normal procedures and data, systemic bias can creep into AI models and produce unfair outcomes at scale.
A hiring model trained on past decisions reproduces an institution's long-standing patterns that favored certain applicants.
A language model from Google that treats every task as turning one piece of text into another piece of text.
By framing everything as text-in, text-out, one model can translate, summarize, classify, and answer questions without a different design for each job.
Feed T5 'summarize: <long expert report>' and it returns a short summary; feed 'translate English to French: ...' and it translates.
An open-source software framework for training very large text-to-text models like T5 at scale.
It is the engineering plumbing that lets researchers train giant language models across many chips efficiently.
A research team uses T5X to train a custom legal-document model across hundreds of accelerators.
A reinforcement learning method that stores the value of every state-action pair in a simple lookup table.
It is the clearest way to understand how an agent learns which actions pay off, before scaling up to neural-network versions.
A board-game bot keeps a table listing how good each move is from each position and updates it as it plays.
The correct answer a model is trying to predict, also called the label.
Without knowing the right answers during training, a model has nothing to learn from.
When training a model to flag privileged documents, the 'target' is the human-marked yes/no for each file.
A slowly updated copy of a learning model used to give stable goals during reinforcement learning.
Chasing a constantly moving target makes learning unstable, so a frozen copy keeps the goal steady enough to converge.
A game-playing agent compares its predictions against a target network that only updates every few thousand steps.
The specific job a machine learning model is built to do, such as classifying or predicting something.
Defining the task clearly is the first step; everything else (data, metrics, model choice) flows from it.
The task might be 'decide whether each email is responsive to a discovery request.'
A way of classifying things into types and subtypes based on their natural relationships.
A good taxonomy gives consistent labels and structure, making large collections easier to organize, search, and reason about.
Tagging case documents under a taxonomy of motions, exhibits, and correspondence so they can be filtered quickly.
A security safeguard that the system itself enforces through its hardware, software, or firmware rather than through human rules.
Built-in technical controls protect data automatically and consistently, without relying on people to remember a procedure.
Automatic encryption of uploaded case files at rest is a technical control on a document tool.
The belief that technology is always the right answer to a problem.
This mindset can lead teams to deploy AI where a simpler or human approach would work better, causing waste or harm.
Insisting an AI model handle a nuanced legal judgment that an experienced attorney should clearly make is technochauvinism.
A setting that controls how random or creative a language model's answers are.
Low temperature gives consistent, predictable answers (good for facts); high temperature gives varied, creative ones (good for brainstorming).
Set temperature near zero so a contract-summary tool gives the same dependable wording every time.
Data that is tied to specific points in time and whose order matters.
Treating time-ordered data carelessly (like shuffling it) destroys the patterns you actually want to learn.
A timeline of when each email was sent during a case is temporal data.
A container for numbers arranged in one or more dimensions, the basic data structure machine learning runs on.
Almost everything fed into or produced by a model, text, images, predictions, is represented as tensors.
A page of scanned text becomes a tensor of pixel values before a vision model reads it.
A specialized Google chip designed to run machine learning math extremely fast.
TPUs let teams train and run large models much faster and cheaper than ordinary processors.
Training a large language model on TPUs finishes in days instead of weeks.
The number of dimensions a tensor has, like whether it is a single number, a list, or a grid.
Knowing a tensor's rank is essential for making model inputs and outputs line up correctly.
A single price is rank 0, a row of prices is rank 1, and a spreadsheet of prices is rank 2.
The number of elements a tensor holds along each of its dimensions.
Mismatched shapes are one of the most common errors when feeding data into a model.
A batch of 32 documents each turned into 768 numbers has shape [32, 768].
The total number of individual values a tensor contains.
It tells you how much memory a piece of data will use, which matters for big models.
A tensor of shape [32, 768] has a size of 24,576 values.
A visual dashboard for watching how a model trains, showing charts of accuracy, loss, and more.
Seeing training progress in graphs makes it far easier to spot problems than reading raw numbers.
An engineer watches the loss curve flatten in TensorBoard to know when a model has finished learning.
A widely used open-source platform from Google for building and running machine learning models.
It is one of the main toolkits engineers use to create and deploy AI systems in production.
A team builds a document-classification model in TensorFlow and serves it inside their case-review tool.
An interactive website where you can train a small neural network in your browser and watch it learn.
It lets beginners see, with no coding, how knobs like learning rate and layers affect a model.
A new analyst drags sliders in TensorFlow Playground to see how adding layers changes the model's decision boundary.
A system for putting a trained model into production so applications can send it data and get predictions.
Building a model is only half the job; serving it reliably to real users is what makes it useful.
A trained privilege-detection model is deployed with TensorFlow Serving so the review app can query it live.
How many times a particular word appears in a single document.
It is the starting point for scoring relevance; words that appear often in a document often signal its topic.
If 'negligence' appears 12 times in a brief, its term frequency in that brief is 12.
The rule that decides when an episode of training or interaction should stop.
Without a clear stopping rule, an agent could run forever or stop too early to learn anything useful.
A simulation ends (the termination condition) when the game is won, lost, or 200 moves have passed.
Checking how well a trained model performs on data it has never seen before.
A model that aces its study material but fails the real exam is useless; testing reveals true performance.
After training on past cases, a model is tested on a fresh set of documents to see if it still classifies them correctly.
A score measuring how wrong a model's predictions are on data it was never trained on.
It is the honest measure of model quality; if it is much worse than training loss, the model has memorized rather than learned.
Low training loss but high test loss warns that a document classifier won't generalize to new cases.
A portion of data held back and used only to judge a finished model's performance.
Because the model never saw this data during training, it gives a fair estimate of real-world accuracy.
Out of 10,000 labeled documents, 1,500 are reserved as a test set the model only sees after training.
A framework of methods and metrics for checking that a system meets its design requirements and is fit for its intended use.
It gives a disciplined way to prove an AI system actually works as promised before and during deployment.
Before launching a contract-review model, a TEVV process measures its accuracy against known cases and confirms it suits the firm's workflow.
A continuous stretch of words or characters within a larger piece of text, marked by where it starts and ends.
Many AI tasks work by pointing to the exact span of text that answers a question or contains an entity.
A question-answering tool highlights the text span 'March 14, 2021' in a deposition as the answer to 'when did it occur?'
An approach where every AI task is framed as taking text in and producing text out.
It unifies many different jobs under one simple format, so a single model can do translation, summarizing, and answering.
A text-to-text model answers 'classify sentiment: This report is excellent' with the output 'positive.'
A formula that scores how important a word is to a document by weighing how often it appears against how common it is everywhere.
It powers classic keyword search by surfacing distinctive words and ignoring filler words like 'the.'
Across thousands of contracts, TF-IDF flags 'indemnification' as meaningful in a clause but downweights 'and.'
An outside entity that takes part in an interaction that is mainly between two other parties.
Third parties such as vendors or APIs introduce dependencies and risks that need to be tracked, especially for data security and trust.
A document tool that sends text to an outside model provider relies on that third party to handle the data securely.
A way of organizing risk control into three separate groups: the team doing the work, an independent oversight team that challenges them, and a third team that audits the whole thing.
Separating who does the work from who checks it keeps any one group from marking its own homework, which is the core of trustworthy governance.
When deploying a document-AI system, the build team runs it, a risk team reviews its decisions, and an internal audit confirms both did their jobs.
A cutoff value that decides how confident a model must be before it says 'yes' to a prediction.
Moving the threshold trades off catching more true cases against raising more false alarms.
Set the threshold so a document is flagged privileged only if the model is at least 80% sure.
Studying data points collected over time to find patterns and predict what comes next.
Lots of real-world data, sales, sensor readings, case filings, arrives over time, and predicting trends has clear value.
Analyzing monthly filing volumes over five years to forecast next quarter's caseload.
One step in a sequence, such as a single word in a sentence or one moment in a time series.
Sequence models process data one timestep at a time, carrying memory forward as they go.
A model reading a sentence handles 'the,' then 'expert,' then 'testified' as three timesteps.
The small chunk of text (roughly a word-piece) that a language model reads and counts.
Models are priced and limited by tokens, not pages. Knowing this explains why very long documents cost more and must be split up.
The word "deposition" might be one token; a 10-page motion might be ~6,000 tokens.
The process of splitting text into the small units a language model reads.
How text is split affects how many tokens you pay for and how well the model handles rare or technical terms.
Tokenization turns a paragraph from a contract into a sequence of word-pieces before the model analyzes it.
The tool that chops text into the small pieces (tokens) a language model can actually read.
Models don't read raw text; the tokenizer decides how words are split, which affects cost and how well the model handles unusual words.
A tokenizer might break 'indemnification' into 'indem,' 'ni,' and 'fication' before the model processes it.
How often the correct answer appears among the model's top few guesses rather than just its single best guess.
For tasks with many possible answers, being in the top five suggestions is often good enough and a fairer measure.
A document-routing model gets top-3 accuracy if the right category is among its three highest-ranked suggestions.
A way of generating text where the model picks its next word only from its top k most likely choices.
It keeps a language model's output varied but on-topic by ruling out wildly unlikely words.
With top-k set to 40, a summary tool chooses each next word from only the 40 most probable options.
A text-generation method where the model picks the next word from the smallest set of choices whose probabilities add up to a chosen percentage.
Also called nucleus sampling, it adapts how many choices to consider on the fly, balancing creativity and coherence.
With top-p at 0.9, the model considers only the most likely words that together account for 90% of the probability.
One branch of a neural network that processes a particular kind of input, often paired with another branch.
Two-tower designs let a recommender separately understand users and items, then compare them quickly at scale.
In a recommendation model, one tower encodes the user and another encodes the document, and their outputs are matched.
Google's custom chip built to run machine learning calculations very quickly.
TPUs make training and serving large AI models faster and more cost-effective than general-purpose chips.
A team rents TPUs in the cloud to fine-tune a model on millions of legal documents over a weekend.
A single physical Tensor Processing Unit, the basic building block of Google's AI hardware.
Understanding the chip helps make sense of how much computing power a training job actually uses.
One TPU chip handles part of a large model's math while many chips work together on the rest.
A board holding multiple TPU chips that work together as one unit.
It is a common unit of compute teams rent or allocate when training models.
A single TPU device packs several chips so a model can train faster than on one chip alone.
The coordinating process that hands out work to and collects results from many TPU workers.
Coordinating dozens of chips requires a conductor; the master keeps the whole training run in sync.
The TPU master splits a training batch across worker chips and gathers their results each step.
A cloud resource that gives you access to TPUs without managing the underlying hardware.
It lets teams use powerful AI chips on demand instead of buying and maintaining their own.
An engineer spins up a TPU node in the cloud to train a model, then shuts it down to save money.
A large cluster of many TPU chips wired together to train very big models.
Training today's largest language models requires the combined power of a whole pod of chips.
Training a frontier-scale model uses a TPU pod with thousands of chips working in parallel.
Any allocatable unit of TPU computing power, from a single chip to a full pod.
Thinking of compute as a resource helps teams plan and budget for training jobs.
A project requests enough TPU resources to fine-tune its model within a fixed deadline.
A portion of a TPU pod assigned to a single training job.
Slicing lets several teams share one big pod, each getting just the chips they need.
A team grabs a small TPU slice for a quick experiment instead of reserving the whole pod.
The specific generation and configuration of TPU hardware you choose to use.
Newer or larger TPU types run faster but cost more, so the choice affects speed and budget.
Picking a higher-performance TPU type cuts a training run from days to hours.
A process that does a share of the actual computation on TPU chips under a coordinator's direction.
Workers are where the real number-crunching happens during large distributed training.
Each TPU worker computes results for its slice of the data and reports back to the master.
The ability to follow the recorded history of something back through every step so you know exactly where it came from and how it was handled.
Without a clear record of an item's path, you cannot prove it is genuine or untampered, which matters for both evidence and audits.
A scanned exhibit logs every handoff and edit, so a reviewer can trace it back to the original source document.
The process of teaching a model by showing it many examples and adjusting it to make better predictions.
Training is how a model actually learns; the quality and quantity of examples largely decide how good it becomes.
Training a model on 50,000 labeled emails teaches it to spot which ones are responsive to a request.
The collection of examples a model studies in order to learn how to do its job.
A model can only be as good and as fair as the data it learned from, so the choice of training data shapes everything it later does.
A document-classification model learns to sort contracts after being trained on thousands of labeled sample contracts.
A running score of how wrong the model's predictions are on the examples it is learning from.
Watching training loss drop tells you the model is learning; if it stops dropping, learning has stalled.
As training proceeds, the training loss falls from 2.4 to 0.3, showing the model is fitting the examples.
The collection of labeled examples a model learns from.
It is the model's textbook; a biased or small training set produces a weak or skewed model.
Ten thousand documents each marked privileged or not form the training set for a privilege-detection model.
A harmful gap between how a model is treated during training and how it is treated when actually used.
A model can look great in training yet fail in production if the live data is prepared differently than the training data.
A classifier trained on clean text underperforms in production because live OCR text has typos training never saw.
The full sequence of states, actions, and rewards an agent experiences during one run.
Learning from whole trajectories lets an agent figure out which early moves led to later success or failure.
A game agent reviews the trajectory of an entire match to learn which opening moves paid off.
A coordinated set of steps carried out across systems or people that produces one intentional, consistent, and checkable result.
Treating a multi-step action as a single transaction means it either fully completes or fully fails, preventing half-finished states that corrupt data.
Uploading a case file, indexing it, and confirming it is searchable all succeed together as one transaction, or none of them takes effect.
Reusing a model already trained on one task as a head start for a new, related task.
It saves enormous time and data, you adapt a powerful general model instead of training one from scratch.
Take a model that already understands general English and adapt it to read legal contracts with far fewer examples.
A neural network design that reads all parts of a text at once and weighs how each word relates to the others.
It is the architecture behind nearly every modern language model, including the ones powering today's AI assistants.
A transformer reading a clause connects 'it' back to 'the agreement' mentioned three sentences earlier.
A model's ability to recognize something no matter where it appears in an image.
It means an image model spots an object whether it is in the corner or the center, making vision robust.
A signature-detection model finds a signature whether it sits at the top or bottom of a scanned page.
Presenting information openly and clearly so people can understand what a system does and hold it accountable.
When users can see how and why an AI reached a result, they can trust it, question it, and catch mistakes.
A document-AI tool shows which source passages it used for an answer instead of just stating a conclusion.
A sequence of three consecutive words or items treated as a single unit.
Looking at three-word chunks captures short phrases and context that single words miss.
In 'breach of contract,' the whole three-word trigram carries meaning a single word cannot.
A case the model correctly identified as NOT belonging to the category.
Counting these is part of measuring whether a model correctly leaves alone the things it should leave alone.
A non-privileged document that the model correctly marks 'not privileged' is a true negative.
A case the model correctly identified as belonging to the category.
True positives are the wins, the things you wanted the model to catch, and it did.
A privileged document the model correctly flags as privileged is a true positive.
The share of all actual positive cases that the model successfully caught.
Also known as recall, it tells you how much of the important stuff you are not missing.
If 90 of 100 truly privileged documents are flagged, the true positive rate is 90%.
A person's confidence, based on their experience, that a system will help them reach their goals even when there is risk or uncertainty.
People only adopt and rely on AI tools they trust, and misplaced trust can be as harmful as no trust at all.
A reviewer learns to trust a search tool after it repeatedly returns accurate, source-linked passages.
How well a system can be counted on to keep information confidential, accurate, and available across all the threats it might face.
Trustworthiness is what earns user trust; it describes the system's actual dependability rather than just how it feels to use.
A document store is trustworthy if it keeps case files private, unaltered, and accessible whenever the team needs them.
AI that is valid, safe, secure, accountable, explainable, privacy-respecting, and fair with harmful bias kept in check.
These traits together are what make an AI system responsible to deploy in high-stakes settings like legal work.
A trustworthy document-AI tool gives accurate, explainable answers, protects client data, and is regularly checked for bias.
A test where a machine passes if a human, while reading its typed answers, cannot reliably tell it apart from a real person.
It became the classic benchmark for asking whether a machine can imitate human conversation, and it still shapes how people judge whether AI 'sounds human.'
If you chatted with a support assistant and genuinely could not tell whether a person or a program was typing back, that program would have passed an informal Turing Test.
A design with two separate network branches, one for users and one for items, whose outputs are compared to make matches.
By encoding users and items separately, it can search huge catalogs for good matches very fast.
A document-recommender uses one tower for the reviewer's interests and another for each document, then matches them.
A false alarm, where a test concludes something is true when it actually is not.
Knowing your false-alarm rate helps you judge how often a system flags problems that are not really there.
A document classifier raises a Type I error when it flags a routine memo as privileged even though it is not.
A missed detection, where a test concludes nothing is wrong when something actually is.
Type II errors are dangerous because the system stays silent about a real problem that slips through.
A document classifier makes a Type II error when it treats a genuinely privileged document as ordinary and lets it through.
Leaving out a sensitive trait like race or gender from the data, in the hope that the model can't then discriminate on it.
It sounds fair but often isn't — other data (like ZIP code) can quietly stand in for the removed trait, so bias sneaks back in.
Removing gender from a hiring screen, yet the model still favors one group because it leans on a correlated signal like the name of a college club.
The state of not having enough accurate knowledge to be sure about a situation or outcome.
Recognizing uncertainty keeps you from over-trusting a prediction and prompts you to gather more information before acting.
A document-AI tool that flags low-confidence answers is signaling its uncertainty so a human can double-check.
When a model is too simple to capture the real patterns in the data, so it does poorly even on the examples it was trained on.
An underfit model gives weak predictions everywhere — it hasn't learned enough, which usually means it needs more capacity, better features, or longer training.
A contract-risk classifier that flags everything as low risk because it never learned the warning signs is underfitting.
When a group appears too rarely in a dataset to be fairly captured by a model.
If a group is underrepresented in training data, the model tends to work worse for that group, creating unfair results.
A document model trained mostly on contracts from one region may handle rarer regional formats poorly because they were underrepresented.
Throwing away some examples from the most common category so a lopsided dataset becomes more balanced for training.
When one outcome is rare, models tend to ignore it; trimming the majority class helps the model actually learn the rare case.
In a fraud dataset that is 99% legitimate, keeping only a fraction of the legitimate transactions so the model can focus on the rare fraud cases.
When it is impossible to give an explanation of a system's decision that is both fully accurate and easy to understand at the same time.
Some accurate AI decisions cannot be made simple without distorting them, which creates a real tension for accountability.
A complex model may pick the right document every time yet be unexplainable in plain terms a court would accept.
A model that reads text in one direction only, using just the words before the current word to make sense of it.
It mirrors how text is generated left to right, which is why many chatbots predict the next word using only what came before.
A model deciding the next word in 'the witness testified that the...' looks only at the words to the left, not anything after.
A language model that predicts the meaning or next word of a text using only the words that come before it, never after.
This one-way design is the basis of most text-generating AI, since writing naturally flows from past words to the next one.
A model that completes 'pursuant to the terms of this...' by reading only the preceding words.
A piece of data that has its input features but no known answer attached to it yet.
Most real-world data is unlabeled; making use of it (without paying humans to label everything) is a major focus of modern AI.
A folder of 10,000 scanned exhibits with no human tags saying which are invoices, emails, or memos.
Information that has no fixed format or predefined organization, like free-form text or images.
Most real-world documents are unstructured, so AI tools must make sense of messy text before they can search or analyze it.
Scanned letters, depositions, and email bodies are unstructured data that a tool must read before it can answer questions about them.
Training a model on data that has no answer key, so it finds structure or groupings on its own.
It lets you discover patterns in piles of unlabeled data — like natural clusters of similar documents — without anyone tagging them first.
Letting an algorithm group thousands of case files into themes on its own, without being told the categories in advance.
A technique that predicts how much a specific action will change an outcome for a particular person, not just what the outcome will be.
It tells you who is actually persuadable by an action, so you spend effort only where it makes a real difference.
Estimating which clients would respond to a follow-up call by signing, versus those who would have signed anyway.
Telling the model to treat certain examples as more important during training, often to make up for there being fewer of them.
It's a way to keep rare-but-important cases from being drowned out without deleting data, by counting each one more heavily.
After undersampling common transactions, giving the kept ones extra weight so the overall picture stays accurate.
How effectively, efficiently, and satisfyingly specified users can reach their goals with a product in its intended setting.
A powerful tool that is hard to use goes unused, so usability often decides whether AI actually helps people.
A document-search tool with clear results and a simple query box has high usability for reviewers under deadline.
Evaluating a product by watching real users try to complete typical tasks with it and noting where they struggle.
Watching actual users reveals problems and satisfaction levels that designers cannot guess on their own.
Before launch, a team watches reviewers search case files in a new tool to find where the workflow confuses them.
A person or group that interacts with a system or benefits from it while it is being used.
Knowing exactly who the users are shapes how a tool should be designed, explained, and safeguarded.
The users of a document-AI tool include the reviewers who run searches and the team that relies on the results.
A table of numbers that captures each user's tastes, learned so the system can match people to items they'll like.
It's half of how recommendation engines work — pairing this user profile with an item profile to predict what someone will enjoy.
A streaming service stores a compact taste profile per viewer to predict which new shows they'll watch.
A design approach that actively involves users, designs and tests in repeated cycles, and draws on many disciplines to fit real needs.
Building around real user needs from the start produces tools people can actually use instead of ones they fight against.
Designers of a document tool interview reviewers, prototype, and refine repeatedly so the search fits how reviewers really work.
Checking how well a model is doing during development by testing it on data it didn't learn from, before the final test.
Validation catches problems early and helps you tune a model without cheating by peeking at the final test set.
After teaching a contract-clause classifier on most of your documents, you check its accuracy on a held-back slice to see if it's actually learning.
A number measuring how wrong a model's predictions are on the validation data during training.
Watching validation loss tells you when a model stops improving and starts memorizing instead of learning the right patterns.
Training loss keeps dropping but validation loss starts rising, a sign the model is overfitting and you should stop.
A portion of your data set aside to tune and check a model during development, separate from both the training data and the final test data.
Without a separate validation set you'd tune your model on the same data you measure it with, fooling yourself into thinking it's better than it is.
You split 1,000 labeled exhibits into training, a validation set for tuning, and a test set saved for the final score.
A principled way of designing technology that deliberately accounts for human values throughout the whole design process.
Considering values like privacy and fairness from the beginning prevents harmful blind spots that are costly to fix later.
A team building a document tool weighs client confidentiality and fairness at every design step, not just at the end.
A training difficulty where the signal used to update a deep network shrinks to almost nothing, so the early layers barely learn.
It explains why very deep or long-sequence models were hard to train, and why newer designs were invented to fix it.
An old-style network reading a long document forgets the opening sentences because the learning signal fades before reaching the early layers.
A measurable characteristic of something being observed that can take on different values or categories.
Variables are the basic units of data analysis; choosing which ones to record determines what patterns you can later find.
When analyzing a set of documents, the document type and page count are variables you can record for each one.
A ranking of how much each input feature contributes to a model's predictions.
It tells you which factors actually drive a decision, making a model easier to explain and trust.
A model predicting case outcomes shows that filing date matters far more than document length.
A measure of how spread out a set of values is around their average.
High variance means values are scattered widely, which affects how reliably you can predict or summarize them.
If document lengths range from one page to hundreds, the variance in page count is high.
A type of neural network that learns to compress data into a smooth set of numbers and then generate new, similar data from it.
VAEs are an early building block of generative AI, letting a model create new examples rather than just classify existing ones.
A VAE trained on thousands of signature images can produce new, plausible-looking handwriting samples.
An ordered list of numbers that a computer uses to represent something like a word, image, or document.
Vectors are the basic language of machine learning, and turning text into vectors is what makes meaning-based search and AI possible.
A paragraph from a deposition is turned into a vector of numbers so the system can compare it to other passages.
A database that stores embeddings and quickly finds the passages most similar in meaning to a query.
It is the index that makes searching thousands of documents by meaning fast — the retrieval half of RAG.
Ask a question and the vector database returns the 10 most relevant paragraphs from across 500 PDFs in milliseconds.
Able to be checked for correctness by a person or a tool.
If a claim or result is verifiable, others can confirm it independently rather than just taking it on faith.
A source-linked AI answer is verifiable because a reviewer can open the cited document and confirm it.
A number inside a model that says how much one input should influence the answer, and that the model adjusts as it learns.
Weights are literally what a model learns; the millions or billions of weights are the model's stored knowledge.
A model deciding if an email is spam gives a high weight to the word 'free' because it strongly signals junk mail.
Multiplying each input by its weight and adding all the results together, the basic math step inside a neuron.
This single calculation, repeated billions of times, is the core arithmetic that lets neural networks turn inputs into predictions.
Estimating a house price by multiplying square footage, bedrooms, and location each by their importance, then adding them up.
A popular tool that records every detail of a model's training runs so you can track, compare, and visualize how training went.
When you train or fine-tune a model, you need a logbook of what settings produced what results; this kind of tracking dashboard is that logbook.
After training a document-classification model, you open a dashboard to see which learning rate gave the lowest error.
A model that learns from a huge number of input features directly, rather than building deep layers of abstraction.
Wide models are good at memorizing specific combinations of features, which helps with tasks like recommendations where exact pairings matter.
A recommendation model that directly learns the pairing 'users who bought item A also bought item B.'
The number of neurons in a single layer of a neural network.
Width (how many neurons per layer) along with depth (how many layers) controls how much a network can learn; wider layers can capture more patterns at once.
Adding more neurons to a layer so the model can track more distinct features in a contract at the same time.
The idea that averaging many independent guesses often beats any single expert's guess.
This principle is why combining many models (an ensemble) usually predicts more accurately than one model alone.
Many people guessing the number of jellybeans in a jar; the average of all guesses lands surprisingly close to the truth.
Representing each word as a list of numbers so that words with similar meanings end up with similar numbers.
Word embeddings let computers grasp that 'attorney' and 'lawyer' are related, powering search and language understanding beyond exact word matching.
The words 'plaintiff' and 'claimant' get close embeddings, so a search for one can surface documents using the other.
A classic technique that turns words into number lists by learning which words tend to appear near each other in text.
It was a breakthrough that taught computers word meanings from raw text, laying the groundwork for today's language AI.
Training on legal text, the model learns 'deposition' and 'testimony' appear in similar contexts and gives them similar number lists.
A compiler that rewrites and speeds up the heavy math behind machine learning models so they run faster on chips like GPUs and TPUs.
It lets the same AI model train and answer faster without changing the model itself, which lowers cost and wait time.
A document-classification model is compiled with XLA so it sorts thousands of contract pages per minute instead of crawling through them.
A logic rule that is true only when exactly one of two inputs is true, but not both and not neither.
XOR is the classic example of a pattern a simple straight-line model cannot learn, which is why deeper, nonlinear neural networks exist.
A light turns on only if one of two switches is flipped but goes off if both are flipped or both are off.
The vertical scale on a training chart, usually showing the model's error (loss) as it learns over time.
Watching the loss on the y-axis drop tells you the model is actually improving during training, and where it stops improving.
Plotting training error on the y-axis against the number of training steps to see when learning levels off.
A fast object-detection method that looks at an entire image just once to find and label all the objects in it at the same time.
Because it processes an image in a single pass, YOLO is fast enough for real-time use, like spotting items in a video feed as it plays.
Scanning a scanned exhibit photo and instantly boxing every visible object such as a vehicle, sign, or person.
A way of rescaling numbers so they are measured in 'how many standard deviations away from the average' a value sits, centering everything around zero.
When features are on wildly different scales (dollars vs. page counts), models can be thrown off; z-score normalization puts every feature on a common footing so each gets fair weight.
Before training a model on case data, you convert raw document counts and damages amounts to z-scores so a million-dollar figure doesn't drown out a 12-page count.
When a model correctly handles a task or category it was never explicitly trained on, using only general knowledge it already has.
It means a tool can tackle new kinds of requests without anyone collecting and labeling examples first, saving setup time and cost.
An AI that was never shown labeled 'force majeure' clauses still flags them in a new contract because it understands the concept from its broad training.
Asking an AI to do something with just an instruction and no worked examples in the prompt.
It is the simplest, fastest way to use an AI; if a plain instruction works, you avoid the effort of crafting example-filled prompts.
You type 'Summarize this deposition transcript in five bullet points' with no sample summary attached, and the model produces one.