What actually makes AI use cheating?
Almost every law school honor code prohibits "unauthorized assistance" on graded work and requires that submitted work be your own. AI is not a special category; it is assistance. So the question is never "is AI cheating" in the abstract. It is (1) did the professor authorize it, and (2) did it substitute for the skill the assignment was designed to measure and certify?
That second question is the one students get wrong. If a legal writing memo is graded on your ability to synthesize authority and structure a rule, then having AI produce the rule statement defeats the assessment even if the words are later rewritten. If a Contracts exam tests issue-spotting, an AI-generated issue list is unauthorized assistance regardless of how much you edit it. Conversely, using AI to explain the difference between promissory estoppel and unjust enrichment three weeks before the exam is no different in kind from asking a 3L tutor.
A workable self-test: could you defend every sentence you submitted in a live conversation with the professor, without notes, and would you be comfortable describing exactly how you used AI if asked on the record? If either answer is no, stop.
Which uses are safe in nearly every course?
These uses put AI in the role of a study partner rather than a ghostwriter. They operate on material you already understand or are trying to understand, and they leave no AI-generated text in anything you hand in.
Note the pattern: in each of these, the AI's output dies before it reaches your submission. It changes what you know, not what you turn in.
- Ask it to explain a doctrine you did not follow in class, then verify the explanation against your casebook, a hornbook, or an Examples & Explanations volume before you internalize it.
- Have it generate practice hypotheticals in a subject area, then write your own answer cold and compare.
- Feed it your own completed outline section and ask what is missing or internally inconsistent — then confirm each gap yourself.
- Use it as a cold-call simulator: paste a case you have already read and have it interrogate you on facts, holding, and procedural posture.
- Ask it to reverse-outline a draft you wrote, so you can see whether your structure reads the way you intended.
- Have it translate a professor's terse margin comment into a plain-English description of the problem, so you can fix it yourself.
Which uses will get you in trouble?
Submitting AI-generated prose as your own analysis is the clearest violation, and it is detectable — not usually by detection software, which is unreliable, but by voice mismatch against your other work, by hallucinated citations, and by your inability to explain your own reasoning when asked.
Two less obvious traps. First, using AI to summarize assigned reading you never opened. Many honor codes do not reach this, but it destroys the skill of reading a case for holding versus dictum, which is the entire point of the first year and the thing the bar exam and your first supervisor will demand. Second, using AI to find authority. Generative models produce plausible case names and citations that do not exist. This is how practicing lawyers have been sanctioned — the well-known example is the Southern District of New York's 2023 sanctions order in the Mata v. Avianca litigation, and the Second Circuit has since referred an attorney for discipline over fabricated citations in a brief. If a student does this in a moot court brief or a seminar paper, the school treats it as fabrication of authority, which is a more serious charge than unauthorized assistance.
Also treat any exam as a closed universe unless told otherwise. "Open book" and "open note" policies written before 2023 often do not mention AI, and professors typically read them as excluding it. Do not assume silence is permission during an exam window.
How do you verify AI output before it touches your work?
Adopt an absolute rule: you never cite a case, statute, or article you have not personally opened and read in Westlaw, Lexis, Bloomberg Law, or an official source. Not the AI's summary of it — the document. Then run KeyCite or Shepard's on it. This costs you five minutes per authority and eliminates the single most damaging failure mode.
Check quotations character by character against the original. AI paraphrases drift into quotation marks constantly, and a misquoted holding in a seminar paper is a fabrication problem even when the case is real.
For doctrinal explanations, cross-check against a source with a named human author before you rely on it. AI is confidently wrong about jurisdictional splits, minority rules, and anything that changed recently. It is especially unreliable on state-specific procedure and on any area where the Restatement diverges from majority case law.
On citation format, understand that The Bluebook has no established form for citing generative AI output. If a professor or journal permits AI use and requires disclosure, ask for the exact format they want rather than improvising one. Most journals that have addressed this require a disclosure note or an author's-note statement, not a citation.
What do the professional rules already require of you?
Start building the practicing lawyer's habits now, because they will bind you within three years. Comment 8 to Model Rule 1.1 makes competence include keeping abreast of the benefits and risks of relevant technology — meaning you are expected to understand AI's limits, not to avoid it. Model Rule 1.6 protects client confidentiality, and Model Rule 5.3 makes you responsible for supervising nonlawyer assistance. The ABA's Formal Opinion 512 (2024) addresses generative AI directly and covers confidentiality, competence, supervision, and fees.
The practical consequence for students: never paste confidential material into a general-purpose AI tool. That includes clinic client files, externship work product, unpublished judicial drafts if you are a judicial intern, and anything under a protective order or NDA from a summer employer. Consumer AI tools may retain and train on inputs. A judicial internship confidentiality breach can end a clerkship pipeline permanently, and a clinic breach implicates Rule 1.6 through your supervising attorney's license.
Some courts also require certification about AI use in filings. If you draft anything that will be filed — in a clinic, a pro bono project, or an externship — ask your supervisor about the specific judge's standing order before you touch an AI tool.
How do you find out your school's actual policy?
This genuinely varies, and the variation is driven by three things: the honor code's definition of unauthorized assistance, the individual professor's syllabus, and whether the course is skills-based or doctrinal. Legal writing and moot court programs are typically the most restrictive, because the assignment is the skill. Seminars and upper-level policy papers are often the most permissive, sometimes with a disclosure requirement.
Read three documents in this order: the student handbook's academic integrity section, the syllabus, and any exam instructions. If the syllabus is silent, email the professor with a specific question — not "can I use AI" but "may I use ChatGPT to generate practice hypotheticals and to check my outline for gaps, with no AI text in my submission?" Specificity gets you a usable answer. Save the reply.
If you get permission and the professor asks for disclosure, disclose fully and in writing. Understating your use after receiving permission is worse than never asking.
What happens if you get caught?
Law school honor code proceedings are not like undergraduate ones. Sanctions range from a failing grade to suspension to expulsion, and findings typically go into your permanent file.
The larger consequence is bar admission. Character and fitness applications in every jurisdiction ask whether you have been the subject of academic discipline, and many require the law school dean to certify your record independently. A disclosed honor code violation is survivable; many applicants are admitted after full disclosure and a candid explanation. A concealed one is frequently fatal, because the concealment itself becomes the fitness problem. If you are ever the subject of a proceeding, get a lawyer or a faculty advisor, and never lie to the committee.
The rational calculation is straightforward. The upside of AI-drafted work is a few saved hours on one assignment. The downside is your license.