Is using AI to brief cases against the rules?
This varies by school and, more often, by professor, and you have to check both. Most law school honor codes were written before generative AI and address "unauthorized assistance" in general terms. Schools have filled the gap in three ways: a default-prohibited rule (AI is unauthorized assistance unless the professor permits it), a default-permitted rule for ungraded preparation like briefing and outlining, or delegation to each syllabus. Read your student handbook's academic integrity provision and every syllabus in your schedule, and assume the stricter of the two controls.
The variation is driven by what the school thinks briefing is. If your school treats case briefs as private study notes, AI use is usually unregulated, the same way commercial briefs and study aids are unregulated. If a professor collects briefs, grades them, or uses them as a participation component, they become graded work product and unauthorized AI use becomes an honor code violation. Some professors also require disclosure of AI use even when it is permitted. When a policy is silent or ambiguous, email the professor and keep the reply.
Separately, note where the rules harden: clinics, externships, journal write-on competitions, moot court briefs, and exams. Those are almost always governed by written AI policies, and some clinics prohibit putting client information into any tool that retains inputs.
What is AI actually good and bad at when reading cases?
Current general-purpose models are reliable at compression and organization when you paste in the actual opinion text. They will produce a clean statement of facts, identify the parties, and lay out a structured summary. They are useful for orienting yourself before a dense opinion and for checking whether you missed an issue.
They are unreliable at exactly the things a good brief turns on. Models routinely misstate procedural posture, which is the single most common error: they confuse who sued whom, which party appealed, what the lower court held, and whether the appellate court affirmed, reversed, or remanded. They frequently quote dissents or concurrences as if they were the majority holding. They state holdings too broadly, stripping out the limiting facts that make the case useful in an exam hypothetical. And they generate citations that look right and are not.
Fabrication is a documented professional problem, not a theoretical one. In Mata v. Avianca, Inc., 678 F. Supp. 3d 443 (S.D.N.Y. 2023), lawyers were sanctioned for filing a brief with AI-invented cases and quotations, and in Park v. Kim, 91 F.4th 610 (2d Cir. 2024), the Second Circuit referred an attorney to its grievance panel for citing a nonexistent decision produced by ChatGPT. If a tool will invent a federal appellate case, it will certainly invent a pincite in your Contracts brief.
What do you lose if AI does the briefing for you?
Briefing is a training exercise disguised as a document. The value is in the struggle of extracting a rule from an opinion that never states it cleanly, deciding which facts are legally operative, and noticing that the court's reasoning does not quite match its stated rule. That work is what builds the skill an issue-spotting exam tests. Reading a finished brief, whether from AI or from a commercial vendor, produces a feeling of understanding without the underlying capacity; cognitive psychologists call the gap the fluency illusion, and the effect is that students who outsource briefing feel prepared and then freeze on a novel fact pattern.
The practical tell shows up in cold calls. You can recite an AI brief's holding, but when the professor asks what happens if the plaintiff had been standing three feet farther away, or why the court distinguished the case it cited two paragraphs earlier, you have nothing, because you never read the paragraph. Exams work the same way. Professors write hypotheticals at the margins of the cases, and the margins are precisely what summaries delete.
There is also a downstream cost. First-year associates are hired to read cases and tell a partner what they mean. Firms are adopting AI tools rapidly, but they are adopting them for lawyers who can verify the output. The skill you skip now is the one that makes you employable alongside the tool.
What is a workflow that uses AI without losing the skill?
Read and brief first, then bring in AI as an adversary. The sequence matters more than any particular tool.
Do the full read of the opinion with no assistance. Write your brief by hand or in a document. Then paste the opinion text and your brief into the model and ask it to attack your work: what did you miss, where is your rule statement broader than the court's, which facts did you call operative that the court never mentioned again. Then go back to the opinion to confirm or reject each critique. The final authority is always the text of the case, never the model.
Ask it to generate hypotheticals rather than summaries. "Give me three fact patterns where this rule produces a different result" is a high-value prompt that forces you back into the doctrine. "Summarize this case" is a low-value prompt that replaces you.
If you use it before reading, limit it to orientation: ask for the area of law and the procedural context in two sentences, then close it and read. Never let it produce the rule statement you will carry into the exam.
- Read the opinion unassisted and write your own brief.
- Use AI to critique your brief against the pasted opinion text, not to write it.
- Verify every proposition against the opinion, not against the model's confidence.
- Use AI to generate practice hypotheticals and counterexamples.
- Never paste confidential clinic, externship, or client material into a consumer tool.
How do you verify what the AI gives you?
Check four things every time. First, the citation: confirm the case exists in Westlaw or Lexis and that the reporter volume, page, court, and year match. Bluebook Rule 10 governs the full form, and the elements to verify are the case name, the reporter cite, the court-and-year parenthetical, and any pincite. If the model gives you a pincite, open the page.
Second, the procedural posture. Find the sentence in the opinion that says what the court below did and what this court is doing about it. Models get this wrong constantly.
Third, whether the quoted language is from the majority. Search the opinion for the quoted string. If it appears after "dissenting" or "concurring in the judgment," the model has handed you a rule that is not the law.
Fourth, subsequent history. Ask whether the case has been reversed, overruled, or superseded, and answer it with a citator (KeyCite or Shepard's), not with the model. Models have training cutoffs and no reliable sense of current validity.
One more rule that predates AI and still applies: do not cite headnotes or editorial summaries as authority. They are the publisher's work, not the court's, and the same is true of an AI-generated summary of an opinion.
Do professional rules matter while you are still a student?
They matter as soon as you are working under a supervising attorney, and the standards are already articulated. ABA Model Rule 1.1 comment 8 makes competence include understanding the benefits and risks of relevant technology. ABA Formal Opinion 512 (2024) addresses generative AI directly, covering competence, confidentiality, supervision, and candor to tribunals, and its core message is that the lawyer remains responsible for every output. Federal Rule of Civil Procedure 11(b) makes the signer responsible for the legal contentions in a filing regardless of how they were produced, and a number of federal judges have issued standing orders requiring disclosure or certification of AI use.
Character and fitness reviews also reach law school conduct. An honor code finding for unauthorized AI use is reportable on bar applications in most jurisdictions. That is a disproportionate consequence for outsourcing a Torts brief, which is a reason to resolve any policy ambiguity in writing before you rely on a tool.