AI for law students: use it to learn faster, not to skip the learning

Between 2023 and 2024, researchers at the University of Minnesota did what nobody selling AI for law students had bothered to do: they ran a randomised controlled trial. Sixty students, four realistic tasks, some with GPT-4 and some without, blind-graded. GPT-4 “only slightly and inconsistently improved the quality of participants’ legal analysis but induced large and consistent increases in speed”. The students who gained most had scored lowest without it.

Read that twice. The tool made weak work arrive faster, not better, and it helped most the people least equipped to notice.

LexisNexis’s Mentorship Gap report, published in February 2026, asked nearly 900 UK lawyers where the skills gaps lie: 72% said deep legal reasoning, 69% verification, and only 2% believed AI strengthens learning. Here is how to use ChatGPT and Claude in law school and land on the right side of those numbers.

What the randomised trials actually found

Quality depends on the tool: in the second trial a database-grounded tool and a reasoning model both lifted quality, “a marked contrast with previous research examining older large language models like GPT-4”.

And remember the 11. The reasoning model produced nearly three times the fabricated authorities of students working alone. A model that reasons is not a model that checks; why AI makes up fake cases explains the mechanism.

The equalising effect and its trap

The Minnesota authors put the optimistic case in one sentence: “Because AI tools have an equalizing effect on performance, they may also promote equality in a famously unequal profession.”

The trap follows: the student who gains most from a draft is the one least able to see where it is wrong. Two University of Chicago professors say so plainly. Mark Templeton: “When you use AI, there is a duty to supervise it like you would a junior attorney or paralegal. And to fulfill that duty, you have to be the expert yourself.” Joan Neal bans AI in her contract-drafting course because students “don’t have the base-level knowledge yet to judge the AI output”.

Suffolk’s Dyane O’Leary gave the rule its shape: “The student must be the bread on both sides. What the student puts in, and how the output is assessed, matters more than the tool in the middle.”

Law schools: from bans to required AI certification

Case Western Reserve was the first US law school to require every 1L to earn a legal-AI certification, from January 2025. Suffolk made a generative-AI track compulsory for 1Ls from 2025-26. Chicago requires 1L AI modules from early 2026 but bans AI in the Fall Quarter of its writing programme.

The pattern is sequence, not prohibition: judgement first, tools second, then tools under written rules. If your school has no policy, adopt Chicago’s timetable yourself. What law schools teach about AI compares the curricula, Berkeley’s 100-hour course included.

Study workflows that build skill: Socratic, adversarial and exam mode

O’Leary’s advice for week one is to start “in a low-stakes, casual environment with topics familiar to them — cooking, travel, home repair”. Then bring the tool to the law in three modes, each with you as the bread.

Socratic mode: the model asks, you answer

Anthropic’s open-source Claude for Legal repository on GitHub ships a law-student plugin with a /law-student:socratic-drill command. This prompt does the same in any chatbot.

Drill me on [consideration in English contract law]. Ask one question at a time, from fundamentals upward. After each answer, say what was right and what was missing, then give the correct answer and name the authority I should read, tagged [VERIFY]. Be strict; I would rather be wrong here than in the exam. Stop after ten questions and list my weak spots. Do not cite any case or statute you are not certain exists.

Adversarial mode: you write, the model attacks

This builds the reasoning the 72% are worried about. Write the answer in full without a model, then paste it in.

Here is my answer to a [tort] problem question: <answer>[paste]</answer>. Act as an examiner determined to find every weakness. List: each step of reasoning that does not follow; each issue I missed; each authority I used for a proposition it does not support; and the one question I would least want to be asked. Do not rewrite my answer or add authorities of your own. If it is largely sound, say so rather than inventing objections.

If the critique is soft, reply: “You conceded too easily. Try again as if the mark depended on it.” Prompt engineering for lawyers explains why drafting and critique belong in separate turns.

Exam mode: outline from memory, then get graded

The point is to find the holes before the examiner does, so forbid the model answer.

Here is the marking rubric for [land law]: <rubric>[paste]</rubric>. Here is an outline I wrote from memory in twenty minutes: <outline>[paste]</outline>. For every element in the rubric, mark it PRESENT, PARTIAL or MISSING and quote the line of my outline that earns the mark. Then list the three gaps that would cost most marks. Do not write a model answer and do not add authorities; I will fill the gaps from the casebook.

Research and writing: no fake citation in your first memo

Dame Victoria Sharp put the rule in one sentence in Ayinde v Haringey: “Freely available generative artificial intelligence tools, trained on a large language model such as ChatGPT are not capable of conducting reliable legal research.”

Then Lnu v. Blanche, Ninth Circuit, June 2026: briefs drafted by unlicensed law graduates, citations unchecked, two attorneys fined $2,500 each and suspended from practice before the court for six months. The standard every graduate inherits: “A competent and diligent attorney must also read and reason.”

So the division of labour is fixed: the model builds structure, the database supplies authority, you read the authority.

Build the analytical skeleton for a memo on whether [client, anonymised] can establish [cause of action] under [jurisdiction] law on these facts: [facts]. Output: (1) the elements, numbered, with the standard of proof for each; (2) for each element, the facts that support it, cut against it, and are unknown; (3) the three most likely defences; (4) the searches I should run in [Westlaw / Lexis / BAILII], phrased as queries. Do not cite any case. Cite a statute only if you are certain it exists, tagged [VERIFY].

Then check what the database returns in six layers: existence; matching parties, court, year and reporter; citator status; the pinpoint; every quotation character for character; jurisdiction. The citation verification guide walks through each.

The skills firms say graduates lack

Sidley Austin London added a mandatory “AI Knowledge Lab” seat to every training contract in September 2026, “a working seat, not a classroom”, as it told Legal Cheek. Harvey puts the new junior job in one line: a first-year who spent 40 hours on a markup “might now spend 5 hours reviewing an agent’s markup”. Review is worth something only if the reviewer could have done the forty.

A senior associate put it bluntly on Reddit, in a comment picked up by the AI and the Future of Law podcast: “For people who already have domain knowledge… AI is an enhancement. And for those who don’t, it’s clearly a replacement.” Will AI replace junior lawyers has the hiring data, less bleak than the headlines; the skills map has the full list.

Confidentiality habits to build before your first client

ChatGPT Free, Plus and Pro and Claude Free, Pro and Max train on your conversations by default; Anthropic flipped its consumer default on 28 August 2025. The opt-outs are two clicks (ChatGPT: Settings > Data Controls > Improve the model for everyone; Claude: Settings > Privacy > Help improve Claude). Business and enterprise tiers do not train by default.

In United States v. Heppner (S.D.N.Y., February 2026) Judge Rakoff held that a defendant’s roughly 31 exchanges with consumer Claude were protected by neither attorney-client privilege nor work product: “Because Claude is not an attorney, that alone disposes of Heppner’s claim of privilege.” The UK Upper Tribunal was blunter about a solicitor who pasted Home Office letters into ChatGPT: that “is to place this information on the internet in the public domain”.

Four habits:

1. Know the tier before you paste. Consumer tiers for casebooks and revision; a no-training tier, or nothing, for anything from a client file.
2. Abstract, do not identify. Brooke Loesby’s ABA Journal example: “I am working on a partnership dispute involving allegations of financial misconduct”, not the client’s name and the $400,000.
3. Never press thumbs-up on anything from a real file. “No training” does not mean no human eyes.
4. Disclose. Chicago’s clinic students must say when they used AI so the class can “review and debate the quality of the AI output”. Do the same, unasked.

Is ChatGPT confidential for lawyers? has the tier-by-tier table.

A first-year AI toolkit (mostly free)

Preparing for interviews: what to say about AI

Interviewers know the two failure modes: the graduate who pastes a client file into consumer ChatGPT, and the graduate who files what a model wrote without reading it. They are listening for a habit, not enthusiasm.

Suppose you are asked how you use AI. A strong answer describes a sequence: “I write the first attempt myself. I use a model to attack it and to build a research plan, never to supply authorities. Everything I cite I have opened in a database and run through the citator. I know which tier I am on, and client facts stay out of consumer tools.” That is ABA Formal Opinion 512’s “reasonable understanding of the capabilities and limitations”. Rehearse it with the Socratic prompt, recast as a sceptical hiring partner.

Where to go next: the learning paths hub collects the training guides, and best AI courses for lawyers rates the paid options. Students and trainees are welcome in AI Lab for Lawyers: four live two-hour sessions, recorded and browser-based, on the confidentiality and verification habits above. A Wolf Theiss associate’s verdict on the September 2026 cohort: “I do feel much more confident now when handling it.”