There are hundreds of AI terms. Most of them you will never need. These are the ones that come up in this work, either because they explain why the tools fail or because they name something a rule already requires of you. Gold-edged entries are the ones worth knowing cold.
How these tools work
Generative AI
Software that produces new content, such as text, images, or code, in response to what you type.
Large language model (LLM)
The engine behind tools like Claude, ChatGPT, and Gemini. It was trained on an enormous amount of text to predict what words should come next.
Stochastic parrot
A blunt description of what the model is doing: assembling word sequences that are statistically likely, without understanding them.
Why it matters: this is the whole reason a fake case can look exactly like a real one. The model is not looking anything up. It is producing text that resembles a citation.Training data
The text a model learned from. It has a cutoff date, and it does not include your files unless someone put them there.
Why it matters: recent amendments and new opinions may simply be absent.Context window
How much text the model can hold in view at once, counting your prompt, your attachments, and its own answer.
Why it matters: in a long document review, material can fall out of view, and the model will not tell you that it did.Where it goes wrong
Hallucination
When the model states something false with complete confidence.
Why it matters: confidence is not a signal of accuracy. The tone is identical whether the answer is right or invented.Fabricated citation
The specific hallucination that gets lawyers sanctioned: a case, statute, or quotation that does not exist, in correct citation format.
Why it matters: every reported sanction case turns on the same missing step. Nobody opened the source.Bias
Systematic skew in output that traces back to the data the model learned from.
Why it matters: it can show up quietly in how a client, a neighborhood, or a charge gets characterized.Prompt injection
Instructions hidden inside a document or web page that the model reads and follows as if you had typed them.
Why it matters: if you feed the tool material you did not write, treat its output as unverified.How you reduce the risk
Prompt engineering
Writing the request carefully so the output is focused and checkable. Less mystique than the name suggests.
Structured prompt
A prompt carrying four parts: role (who the model should act as), context (the facts and jurisdiction), task (exactly what to produce), and format and limits (the shape of the answer and what it may not do).
Why it matters: the less you specify, the more the model invents to fill the gap.Zero-shot and few-shot prompting
Zero-shot is asking with no examples. Few-shot is including one or two samples of the output you want.
Why it matters: examples steer format and tone harder than instructions do.Persona, or role
Telling the model who to be, such as a Texas criminal defense attorney preparing a suppression hearing.
Chain of thought
Asking the model to work through its reasoning step by step instead of jumping to an answer.
Why it matters: it improves multi-step reasoning and, more usefully, shows you where the logic breaks.Grounding
Requiring the model to answer from documents you provide rather than from memory. In this session, the D in TCDLAi.
Why it matters: it is the difference between a guess and a citation you can check line by line.Retrieval-augmented generation (RAG)
The technical name for grounding. The tool retrieves passages from a set of documents and answers from those passages.
Why it matters: it cuts hallucination sharply. It does not eliminate it, and it will still bridge a gap the sources do not cover.Verification, or the inspect step
Opening the actual source and confirming that the quote, holding, and pinpoint are really there. The lowercase i in TCDLAi.
Why it matters: asking the model to check itself is a fast first pass, not a substitute for reading the opinion.Guardrail
A limit you impose on the tool or on yourself, such as cite only sources I provide, or flag anything you are unsure of.
Human in the loop
A person reviews and approves the output before it is used. In practice, that person is you, and the review is not optional.
Your data and your duties
Consumer tool versus enterprise tool
A free or personal account generally offers weaker contractual protection than an enterprise agreement with terms on retention, training, and confidentiality.
Why it matters: this distinction, not the model itself, usually decides whether client facts may go in.Training on your inputs
Whether the provider uses what you type to improve its models. Some tools do this by default and allow you to turn it off.
Why it matters: read this setting before the first client-adjacent prompt, not after.Retention and memory
Retention is how long the provider keeps what you typed. Memory is a feature that carries details between conversations.
Why it matters: a convenience feature can quietly become a store of client information.De-identification
Stripping the details that identify a matter, such as names, dates, locations, and cause numbers, while keeping the legal shape of the problem.
Why it matters: it is what makes a consumer tool usable at all for real work.Custom instructions
Standing directions that apply to every conversation, such as always cite only what I paste, and flag anything unsupported.
TCDLAi
The framework used in this session: Target the issue, Compile the facts, Define the law from your sources, List strategies, Analyze them, and inspect the result. The lowercase i is deliberate. That step stays with you.