Legal AI for Law Students

Two different things happen when you ask an AI tool for a case citation. One is safe. One is how sanctions happen.

• Grounding: the tool retrieves something it has actually seen: a real case, a real docket, a real holding.
• Generation: the tool predicts what a case would say if it existed. It invents the citation, not maliciously; predicting plausible next words is just how language models work.

The difference is invisible in the output: both look like citations, formatted the same way, and you cannot tell which you have by reading it. You have to check: pull the case, search the quote, confirm the holding.

Leading general tools ground their answers when they can, but “when they can” is not “always,” and even grounded tools misstate holdings. Assume every citation is generated until you prove it is grounded.

Most legal AI tools forget your matter between sessions. That is not a bug: each conversation starts from zero, and legal work is contextual.

What it costs you:

• Case briefing restarts its setup every session
• Memos fragment across days; the tool never sees the full arc
• Outlining needs re-explaining with every follow-up

What to do: paste full matter context once per session and keep that chat open, or carry the history forward into the next one. Tools with longer context windows handle this better, but better is not perfect: budget setup time, and test session continuity when you evaluate any tool.

Courts have repeatedly sanctioned lawyers for filings built on AI-fabricated citations. In the most widely known example, a federal court in New York sanctioned lawyers in 2023 after their brief cited non-existent cases generated by ChatGPT: read the docket yourself in Mata v. Avianca (S.D.N.Y., 1:22-cv-01461). Every one of these sanctions began the same way: a lawyer cited something they had not opened.

The rule is simple: never cite what you have not opened.

Imagine your draft cites six cases an AI handed you. Before that draft goes anywhere, each one gets this:

1. Write down the citation and the holding the AI claimed.
2. Search for the case in Google Scholar or your research tool.
3. Pull the opinion. Confirm the case is real.
4. Ctrl-F the exact quote the AI gave you inside the opinion.
5. Read the context around the quote. Confirm the holding is what the AI said it is.
6. Check the court and year. Binding in your jurisdiction, or merely persuasive?
7. Confirm the parenthetical accurately describes the holding.

Three to five minutes per cite if you are fast. Why so thorough? A case that holds X in the majority might undercut X in a concurrence. An old holding might be overruled. A quote that reads well in isolation might be distinguished three pages later. You cannot know until you read it.

Do this on every cite before you file or send anything. In a rush, especially. And especially when the cite came from AI, because that is where the problem lives.

A good legal prompt has five parts. Leave one out and the output gets worse.

• Role: who the AI is playing. “A legal research assistant,” “a litigation partner reviewing a brief.” Shapes tone and depth.
• Facts: the specific situation, not “a car accident” but the vehicle, the impact, the damages. Specific facts get specific answers.
• Jurisdiction: “North Carolina contract law,” “Second Circuit.” Skip it and you get a generic national answer that is often wrong for your state.
• Output format: a three-sentence summary, IRAC, a checklist. Without a constraint the AI guesses.
• Constraint: what you are NOT asking for. “Case law only, no policy arguments.” “Last five years only.” Often more important than the ask.

A full prompt looks like this:

“You are a first-year law student’s research assistant. I have a contract dispute involving a residential lease in California. The tenant claims the landlord made an implied warranty of habitability. Give me the three essential elements as a numbered list with case citations, then explain each in plain language. Only use California cases, at least one from the last ten years.”

All five parts are there. Drop one and the prompt fails: no jurisdiction gets you a vague federal-state mix; no format gets a long paragraph when you wanted a checklist; no role gets jargon when you wanted a student-level explanation.

Law professors grade in IRAC: Issue, Rule, Analysis, Conclusion. Most AI tools do not produce it unless you name it, so an unformatted answer looks right and grades wrong.

• Issue: the legal question. “Does this breach trigger the implied warranty clause?”
• Rule: the governing law. “Under California Commercial Code section X...”
• Analysis: law applied to these facts. “Here, the contract says Y and the buyer did Z, so...”
• Conclusion: the answer. “Therefore, the buyer has a claim.”

Name it directly in the prompt: “Structure your answer in IRAC format,” then spell out the four parts. Widely used tools recognize IRAC by name; the payoff is output already shaped for your professor’s rubric.

From building and testing citation-checking systems, we keep seeing legal AI fail in four recognizable patterns. Learn the four and you catch the error before it reaches a filing.

1. Invented case. The case does not exist; the tool predicted what a case with that name would say. Catch: search for it. If it is not in Google Scholar or your legal database, it is not real.

2. Real case, wrong holding. The case exists, but the opinion does not say what the tool claims. Catch: pull the opinion and read the actual holding.

3. Real case, invented quote. The quote sounds like the case but is a paraphrase dressed as a direct quote. Catch: Ctrl-F the exact string. If it is not there word for word, it is not a quote.

4. Real quote, wrong attribution. A court really said it, just not that court, that case, or that year. Catch: search the quote inside the cited case; then search it globally to find where it actually lives.

All four look right when the tool returns them. That is the trap, and why verification takes reading, not skimming: you cannot tell which pattern you are holding until the opinion is open.

Honor codes vary, but most now address AI, and they all ask the same question: did you do your own work? Most schools explicitly permit AI use; some require disclosure, some require the tool’s name, some require your prompt transcript. Check your school’s policy.

• Using AI to research cases and summarize holdings, then validating and choosing what matters: your work
• Using AI for a first draft you rewrite in your own words and voice: your work
• Copying an AI paragraph into your memo unchanged: not your work. That is the violation.

The pragmatic test: if you can read your work aloud and explain every sentence and every cite, it is yours. If you cannot, it is not.

Two safeguards: save your prompts, since many schools now ask for the transcript. And do not volunteer disclosure unless the professor asks or the code requires it; know your professor, check the syllabus, and when the rules are ambiguous, ask privately. The safest path: use AI openly, document what you did, be ready to explain it.