A plain-language glossary

The AI vocabulary a lawyer actually needs.

No hype, no computer science. Just the terms you'll run into, explained the way I'd explain them to a colleague — and, on the handful where it matters, what each one means for your professional obligations.

The basics

Model
The AI itself — the trained system that does the work. When people name "ChatGPT," "Claude," or "Gemini," they're naming an interface to an underlying model. Different models have different strengths, costs, and limits, which is why choosing the right one for a task is a real decision, not a detail.
Large language model (LLM)
The kind of model most legal work uses. It's trained on an enormous amount of text and, at bottom, predicts language — which turns out to be enough to draft, summarize, analyze, and answer with striking fluency. Fluency is not the same as accuracy, which is the whole reason verification matters.
Prompt
What you ask the model to do. The quality of the prompt shapes the quality of the output more than most people expect — a vague instruction gets a vague result. Learning to prompt well is the closest thing to a core skill in using these tools, though, as agents arrive, it's giving way to something more like supervision.
Token
The unit a model reads and writes in — roughly a word-part. Models are measured, priced, and limited in tokens rather than words. A page of text runs about 500 to 800 tokens. You rarely need to think about them directly, but they're why there's a ceiling on how much a model can consider at once.
Context window
How much the model can hold in working memory at one time — your instructions plus everything it has read in the session. This is the number that matters most in practice: a large context window is what lets a model take in an entire case file and reason across all of it at once, instead of a page at a time. It's the whole reason the thyroid story I tell works.

How the tools behave

Hallucination
When a model states something false with complete confidence — a case that doesn't exist, a statute misquoted, a fact invented out of nothing. It doesn't look like an error; it looks exactly like a correct answer. This is the term every lawyer must internalize.
Why this matters for your practice

Sanctions have already followed lawyers who filed AI-generated briefs citing fabricated cases. The rule is simple and non-negotiable: every citation, every quote, every factual claim an AI produces is verified against the source before it is relied on or filed. The tool drafts; you confirm. That's competence under the rules, not an optional precaution.

Multimodal
A model that can work with more than text — reading images, PDFs, handwriting, audio, or video alongside written input. In practice this is what lets you hand a model a scanned police report, a photograph of an exhibit, or a medical record image and have it work from that directly.
RAG (retrieval-augmented generation)
A method that lets a model answer from a specific set of documents you give it — your files, your case law, your firm's templates — rather than only from its general training. It's how you point a model at your material and get answers grounded in it, which is far more useful, and far safer, than asking a model to recall things from memory.
Fine-tuning
Additional training that adapts a general model to a narrower purpose or style. Most firms never need it — good prompting and RAG cover the vast majority of real work — but it's worth knowing the word so you can recognize when someone is selling you something more complicated than you require.

Agents

AI agent
Software that works toward a goal across several steps — planning, doing, checking its own output, and continuing until the task is complete — rather than answering one question and stopping. The difference from a chatbot is the difference between a search box and a second chair: one responds, the other works through the file.
Agentic
The adjective for tools that behave like agents. When someone describes a system as "agentic," they mean it takes a goal and carries out multi-step work toward it, with some independence between your instruction and the finished result — always, in a properly built legal system, returning to a human for the decisions that matter.
The shift from prompt to supervision
Not a term of art, but the change worth naming. With a chatbot, the skill was writing a good prompt. With an agent, the skill becomes reviewing good work — supervising what the agent produced the way you'd review a junior associate's memo. The lawyer's job doesn't shrink; it moves up a level.

Privacy and control

Open vs. closed models
A closed model runs on the provider's servers; you reach it over the internet. An open model can be downloaded and run on hardware you control. Closed models are usually more capable and easier to use; open models give you physical control over where your data goes.
Why this matters for your practice

For the most sensitive client matters, an open model running locally means the data physically never leaves your machine — no third party, no transmission, nothing to breach. That's not a setting you toggle for peace of mind; it's an architecture you can point to when a client, or a bar authority, asks how their confidences are protected.

Local / on-device
Running a model on your own computer or server rather than sending data to an outside service. Slower and more limited than the big cloud models, but the confidentiality guarantee is absolute in a way a promise never is: the information doesn't travel because there's nowhere for it to travel to.
Enterprise vs. consumer accounts
The same-looking tool can carry very different terms. Consumer (free or personal) accounts often reserve the right to use your inputs to train future models. Enterprise and business accounts typically guarantee they will not train on your data, encrypt it, and contract around confidentiality.
Why this matters for your practice

Putting privileged client information into a consumer account that may train on it is the kind of confidentiality problem that's easy to avoid and hard to undo. The fix is a one-time setup decision: real client work runs on an enterprise-grade account with the right terms. Get that right once and it protects every matter after.

Let's talk

Knowing the words is the start. Using them well is the work.

Book a consultation

Credited in full toward your training or build if you continue.