TL;DR: This glossary defines the key terms behind legal AI in plain English — from artificial intelligence, machine learning, and large language models to RAG, hallucination, TAR, and human-in-the-loop. Each entry is a one- or two-sentence definition you can scan, grouped into core concepts, how legal AI is built, common workflows, and the safeguards that keep it reliable.
Legal AI has its own vocabulary, and the words do real work: the difference between a tool that "uses AI" and one that grounds every answer in a cited source is hidden inside terms like retrieval-augmented generation and citation grounding. This glossary collects the terms a lawyer actually needs, defines each crisply, and links the big ones to fuller explainers. Use it as a reference — skim the group you need, or read it through once to build a working mental model of how these systems behave.
Core AI concepts
Core AI concepts are the foundational ideas that everything else in legal AI builds on — what the technology is, and the broad families it divides into.
| Term | Definition |
|---|---|
| Artificial intelligence (AI) | Artificial intelligence is software that performs tasks normally requiring human intelligence, such as understanding language, recognizing patterns, and drawing inferences. In law it spans research, review, drafting, and more; see what legal AI is. |
| Machine learning (ML) | Machine learning is a branch of AI in which a system learns patterns from data rather than following hand-written rules, improving as it sees more examples. It underpins everything from document classification to predictive coding. |
| Generative AI | Generative AI is a class of model that creates new content — text, summaries, drafts — rather than only classifying or scoring existing data. The contrast with traditional machine learning is explained in machine learning vs generative AI. |
| Large language model (LLM) | A large language model is a generative model trained on vast amounts of text to predict and produce language, and the engine behind most legal AI assistants. Read more in legal LLMs explained. |
| Natural language processing (NLP) | Natural language processing is the field concerned with enabling computers to understand and generate human language, including parsing documents, extracting entities, and answering questions phrased in ordinary prose. |
| Multimodal AI | Multimodal AI is a model that works across more than one type of input — for example text plus images — letting a legal tool read a scanned exhibit, a chart, or a signature block alongside ordinary text. |
| Inference | Inference is the act of running a trained model to produce an output for a new input — the step that happens each time you ask a question and the system answers. Training builds the model; inference uses it, and it is the step you are paying for each time a tool answers a question. |
How legal AI is built
The terms in this group describe how a legal AI system is assembled and how it retrieves the right material — the machinery that turns a question into a grounded answer.
| Term | Definition |
|---|---|
| Training data | Training data is the collection of examples a model learns from. Its quality and coverage shape what the model knows and where it is weak; gaps in training data are a common source of confident errors. |
| Fine-tuning | Fine-tuning is the process of further training a general model on a narrower, task-specific dataset so it performs better on that task — for instance adapting a model to legal language and citation style. |
| Embedding | An embedding is a numerical representation of text that captures its meaning, so that passages with similar meaning sit close together in mathematical space. Embeddings are what make semantic search possible. |
| Vector search | Vector search is retrieval that finds documents by comparing embeddings, returning passages whose meaning is closest to a query rather than those that share exact keywords. It is the retrieval engine inside most RAG systems. |
| Token | A token is the small unit of text — roughly a word fragment — that a language model reads and generates. Model limits and usage costs are usually measured in tokens. |
| Context window | The context window is the maximum amount of text, measured in tokens, that a model can consider at once. A larger window lets a tool reason over longer documents, but very long inputs can still dilute attention. |
| Retrieval-augmented generation (RAG) | RAG is an architecture that retrieves relevant source documents first and has the model answer using only those passages, which sharply reduces fabrication. It is the backbone of trustworthy legal AI; see RAG for legal AI. |
| Semantic search | Semantic search is retrieval by meaning rather than exact words, using embeddings to surface conceptually relevant material even when the wording differs from your query. It is what lets you research in plain language. |
Legal AI workflows
These terms name the everyday tasks and techniques where legal AI meets practice — how you instruct a model, and the established workflows it now accelerates.
| Term | Definition |
|---|---|
| Prompt | A prompt is the instruction or question you give an AI system. Its clarity, scope, and context strongly affect the quality of the answer you get back. |
| Prompt engineering | Prompt engineering is the practice of writing instructions that reliably produce good results — being specific, supplying context, and constraining the task. The lawyer's version is covered in prompt engineering for lawyers. |
| Technology-assisted review (TAR) | TAR is the use of machine learning to prioritize and classify documents in discovery or review, so reviewers focus on the most relevant material first. Its mechanics appear in TAR and predictive coding. |
| Predictive coding | Predictive coding is a specific TAR technique in which the system learns from a reviewer's coding decisions on a sample and applies the same judgment across a much larger document set. |
| Document automation | Document automation is the generation of documents from templates, data, and rules — turning a contract or pleading that once took hours of copy-and-paste into a guided, repeatable assembly. See legal document automation. |
| Optical character recognition (OCR) | OCR is technology that converts images of text — scanned pages, photographs, faxes — into machine-readable, searchable text, so that even a faint photocopy can be reviewed, searched, and translated. |
| Citation grounding | Citation grounding is the practice of tying every AI claim to the exact source passage behind it, ideally with page and quoted text. It is what makes an answer verifiable; the discipline is set out in how to verify AI legal research. |
| AI agent | An AI agent is a system that can plan and carry out multi-step tasks toward a goal — for example breaking a research question into sub-questions and pursuing each — rather than answering a single prompt in isolation. Agents are increasingly used to chain research, review, and drafting steps into one flow. |
Safety and oversight
These terms describe what can go wrong with legal AI and the safeguards that keep it usable — the concepts every responsible practitioner should be able to name.
| Term | Definition |
|---|---|
| Hallucination | A hallucination is a confident but false output — most notoriously, a fabricated case citation that looks real. It is the central risk of generative legal AI and the reason verification is mandatory; see the Mata v. Avianca cautionary tale. |
| Guardrails | Guardrails are the constraints built into an AI system to keep its behavior within safe bounds — refusing unsupported claims, scoping answers to retrieved sources, or flagging low confidence. National bodies such as NIST publish frameworks for managing these risks. |
| Human-in-the-loop | Human-in-the-loop is a design principle in which a person reviews, confirms, or overrides AI output before it is relied on — the model that keeps responsibility with the lawyer. It is explored in human-in-the-loop legal AI. |
How these terms fit together
Read in sequence, the glossary tells a single story. A large language model, a kind of generative AI built with machine learning, is powerful but prone to hallucination on its own. To make it reliable for legal work, you wrap it in retrieval-augmented generation — using embeddings and semantic search to pull real sources into the context window — and you insist on citation grounding so that every claim can be opened and checked. Workflows like TAR, document automation, and AI agents then apply that grounded capability to discovery, drafting, and research, while guardrails and a human-in-the-loop keep the output honest. Independent research centers such as Stanford's Institute for Human-Centered AI study exactly this balance of capability and oversight. Knowing the vocabulary is not academic: it is how a lawyer tells a tool that merely sounds clever from one that can actually show its work.
A citation-first platform puts the safe end of this vocabulary into practice. Judicio's Legal Research grounds every answer in a real source, cites the exact page and passage, and keeps a human in the loop by design — outputs are research aids, not decisions. The same terms describe its other tools: semantic search across the File Library, OCR on scanned uploads, and document automation in Drafting. To see the ideas in action, you can start a 7-day free trial with 500 credits and no credit card, or read the pillar explainer on what legal AI is.
This glossary is educational and its definitions are general; AI outputs are not legal advice, and a qualified lawyer remains responsible for verifying any result.