What is AI actually good at in law school?
General-purpose models like ChatGPT, Claude, and Gemini are strong at three things: explaining a concept at whatever level you ask for, generating unlimited practice hypotheticals, and giving fast structured feedback on writing you produce. Those map neatly onto the parts of 1L that are hardest to do alone — you cannot generate your own issue spotters that surprise you, and you cannot get your professor to read five practice essays.
They are weak at exactly the thing students most want to use them for: telling you what the law is. A model producing text from general training data does not know that your Contracts professor teaches the UCC § 2-207 fight her way, that your Torts course follows the Third Restatement rather than the Second, or that your jurisdiction rejected the majority rule. Your exam grades your professor's course, not the internet's average view of the law.
The practical upshot: use AI where the correct answer is something you supply or something you can immediately verify. Use your casebook, supplement, and class notes where the correct answer has to come from an authority.
How do you use AI without letting it do the learning for you?
Learning happens when you generate and retrieve, not when you read fluent explanations. Every workflow below is designed so you produce first and the AI reacts second. If you reverse the order, you will feel productive and remember nothing.
Build the loop into a weekly routine: after each topic, write the rule from memory, then check it; after each unit, write a timed practice answer, then get feedback; before each class, brief the case yourself, then stress-test the brief.
- Cold-call prep: brief the case yourself, paste your brief plus the opinion, and ask the model to play the professor and ask five follow-up questions you cannot answer from the brief alone.
- Issue spotters: ask for a 45-minute fact pattern covering three named doctrines from your syllabus, write the answer under time, then paste your answer with your professor's model answer or a released bar exam sample answer as the grading rubric.
- Rule articulation: write out the elements of a doctrine from memory, then ask the model to identify what you omitted — and go confirm each flagged gap in your casebook.
- Reverse explanation: explain the holding of a case to the model in your own words and ask it to identify where your explanation is imprecise. Vague spots are the ones you do not actually understand.
- Outline compression: paste your own 60-page outline and ask for a one-page attack sheet of decision trees. The compression is useful; the underlying synthesis must be yours.
How should you prompt so the output is usable?
Ground the model in text. The single biggest quality difference comes from pasting the actual source — the opinion, the statute, your notes, the professor's slides — into the prompt and instructing the model to answer only from that text and say so when the text does not resolve the question. This is why Google's NotebookLM is often better for study than a bare chatbot: it answers from documents you upload and cites back to them.
Assign a role and a standard. "You are a 1L Civil Procedure professor grading on a curve; identify every issue I missed and every place I stated a conclusion without applying a rule to a fact" produces far more useful output than "how did I do?" Ask for criticism explicitly, because these models default to agreeable.
Never ask an ungrounded model for citations. If you need authority, go to Westlaw or Lexis and find it. If you want the model to help you understand an authority, bring the authority to the model.
Which tools should you use?
You already pay for the best legal ones through tuition. Westlaw and Lexis both give law students full academic passwords, and both now bundle generative features — Lexis+ AI and Westlaw's Precision/CoCounsel tools — that run retrieval against their actual case databases rather than inventing text. They are far safer for anything touching authority, and learning them now is directly transferable to summer work.
For study support, general models are fine and often better at explanation. Commercial study aids — Quimbee, CALI lessons, Examples & Explanations, Emanuel and Understanding treatises — remain more reliable than any chatbot for black-letter rules, because a human expert wrote and edited them. Treat AI as the thing that quizzes you on the treatise, not the thing that replaces it.
Note that these tools are irrelevant during most exams. If your school uses Examplify or a similar lockdown browser, you have no internet and no AI. Build a paper attack outline you can use without one.
What are the integrity and ethics rules you have to check?
Policies vary enormously and the variation is driven by school honor codes, individual professor syllabi, and the format of the assessment. Some schools have adopted blanket AI policies; more have delegated the question to faculty, so the same student may face a total ban in Legal Writing and permitted use in a seminar. Take-home exams and graded memos are where students get caught, and an honor code violation is reportable on your bar character and fitness application. Read the syllabus, and if it is silent, email the professor and keep the reply.
The professional rules are already real for you. ABA Model Rule 1.1 comment 8 makes competence include understanding the benefits and risks of relevant technology, Rule 5.3 covers supervising nonlawyer assistance, and ABA Formal Opinion 512 (2024) addresses generative AI specifically, including confidentiality and client consent. If you are in a clinic or externship, do not paste client facts into a consumer chatbot — that is a Rule 1.6 problem, not a study habit problem.
The cautionary cases are real and recent. In Mata v. Avianca, a federal judge in the Southern District of New York sanctioned lawyers in 2023 for filing a brief with cases ChatGPT invented, and in Park v. Kim the Second Circuit referred an attorney to its grievance panel in 2024 for citing a nonexistent decision. Courts publish these opinions with the lawyers' names in them.
What should you never delegate?
Do not have AI read cases for you. Reading judicial opinions slowly and badly, then better, is the actual skill 1L teaches, and the discomfort is the point. A summarized case gives you the holding without the reasoning, and exams test reasoning.
Do not have AI write your outline. Outlining works because you make the choices about hierarchy and relationships; a generated outline is someone else's synthesis that you will not be able to navigate under time pressure.
Do not trust AI on Bluebook. Citation formatting is rule-driven and models routinely blend Bluebook and ALWD conventions, invent reporter abbreviations, and misapply the practitioner Bluepages rules versus the academic whitepages rules. Use the Bluebook itself, Table T1 for jurisdictions, and the "copy with reference" function in Westlaw and Lexis, which produces a correctly formatted cite from the actual document.
How does this connect to the bar and to practice?
The NextGen bar exam, which states began adopting starting in July 2026, tests lawyering skills — issue spotting, legal analysis, client counseling — in integrated performance-style items. That format rewards exactly the practice-and-feedback loop AI is good at supporting and punishes students who have only ever read summaries. Confirm which exam your jurisdiction administers and when.
In practice, firms increasingly expect associates to use AI tools competently and to verify their output completely. Building the habit now — draft first, generate second, verify always, cite from the database — is the habit that keeps you out of a sanctions opinion.