Inside Berkeley Law’s New AI Policy, and What It Means for Legal Education

What Berkeley Law’s AI policy prohibits, and what it allows

Berkeley didn’t ban AI. As a matter of policy, it made the use of AI prohibited as a default rule for almost every part of a student’s substantive work. Here’s the rule itself:

“The use of AI is prohibited for aid in conceptualizing, outlining, drafting, revising, translating, or editing any work submitted for credit. AI use is prohibited for any use for any purpose in any exam situation. Students may not upload course materials—including assignments, readings, slides, class recordings, or other class content—into generative AI systems. AI can be used for research on papers ONLY for the limited purpose of identifying sources, such as cases, statutes, or secondary sources. Students are responsible for the accuracy of their research and all other aspects of their submitted work. Citations to sources that do not exist will raise a presumption of prohibited AI use.”

The categories of prohibited use listed in the policy include brainstorming, organizational structure, summarizing, identifying repetitive passages, revising or polishing a paper, generating an exam outline, and asking AI to translate a paper originally written in another language.

This is not a blanket ban, and law professors may include AI in the curriculum if they choose. And even where they don’t, students retain a narrow allowance for research.

Instructors can deviate from the default rule in writing, with appropriate notice, and can require students to disclose any authorized AI use. So professors who want to teach AI as part of their courses can opt in.

The student research exception is narrower. AI can be used for research on papers “only for the limited purpose of identifying sources, such as cases, statutes, or secondary sources.” Students remain responsible for the accuracy of the research and all other aspects of their submitted work. And citations to sources that do not exist will raise a presumption of prohibited AI use.

What the policy gets right

A few things are worth defending in Berkeley’s policy, even if you don’t agree with where it draws the line.

It’s a clear rule. Law schools across the country have been wrestling with what to tell students about AI use, and most of them haven’t said much. This leaves students wondering whether or under what circumstances AI use is permitted, and whether other students are secretly juicing their work with artificial intelligence. Berkeley has made the default explicit. That clarity is itself a service to students, even if some students would prefer a different default.

It puts responsibility for citation accuracy on the student. The policy says students are responsible for everything they cite, and that citations to nonexistent sources will be presumed to be AI-generated. That’s a good norm to build into law students. The AI hallucination problem in court filings is real enough that it’s worth teaching at the law school level.

It requires disclosure. Where AI use is authorized, students have to disclose it. That’s a good, instructive policy, and the direction law schools should be heading.

It’s a floor, not a ceiling. Professors who want to teach AI can opt in. Berkeley isn’t pretending that AI doesn’t exist; it’s setting a strict default and pushing the affirmative decision to use AI down to the course level. That’s a structural choice that some professors will appreciate, especially those who already restrict laptops, Wi-Fi, or exam software to preserve specific pedagogical experiences.

Where the policy goes too far

The policy is also too broad in places, in ways that work against its own goals.

Enforcement is the most obvious issue. The rule against jaywalking in New York City was a fine rule that got broken constantly with no real consequences. Berkeley’s policy faces the same problem. Any student who decides to be a “secret cyborg,” to use AI quietly and never disclose it, is going to be hard to catch. The detection tools are unreliable, and the most sophisticated AI use leaves no trace. A blanket prohibition that can’t be enforced creates a two-tier system: students who follow the rules and students who don’t, and the rule-followers may end up at a disadvantage.

The chilling effect on curiosity is the bigger issue. Berkeley’s list of prohibited uses includes brainstorming, summarizing, and identifying repetitive passages. These are exactly the kinds of low-stakes uses that help students explore unfamiliar material, find their way into a topic, and build intuition. A student who would otherwise ask Claude or ChatGPT to summarize a hard case before reading it, or to brainstorm angles on a paper topic, will now think twice. Some of those uses might have helped them learn faster. Some of them might have surfaced ideas they wouldn’t have found on their own.

There’s a serendipity argument here too. Law librarians talk about the value of perusing the stacks and finding books you didn’t know existed. AI is a different kind of serendipity engine. Ask it about a doctrine and it might point you to a concept or a case you’d never have encountered. Cutting off that mode of discovery for first-year students assumes that we know exactly how learning happens, and we don’t.

Research on AI and legal reasoning cuts the other way

Recent research makes that point empirically rather than philosophically. Professor Dan Schwarcz at the University of Minnesota recently ran a study with collaborators at the University of Michigan testing the hypothesis that AI hurts legal reasoning. Two groups of law students were given cases, statutes, and regulations. Group one had less AI access; group two had more. Both groups went through four stages of work.

The hypothesis going in was that the AI-assisted group would do the early work faster but worse at the harder analytical work, and that once the AI was taken away, the no-AI group would outperform the AI group because they’d built the mental models themselves.

That’s not what happened.

• Stage one (synthesizing cases, statutes, and regulations): The AI group did the work faster and better, as expected.
• Stage two (multiple-choice questions about the law, with AI removed from the AI group): The two groups did equally well. The AI group didn’t lose any ground from having used the tool.
• Stage three (applying client facts to the law, again with AI removed from the AI group): The AI group actually did better than the no-AI group. The researchers’ hypothesis was wrong in the opposite direction.
• Stage four (both groups given AI): The AI had mixed results, helping weak writers but hurting strong writers.

Worth noting: The AI tended to raise the floor for weaker writers more than it raised the ceiling for stronger writers. Some already-strong writers actually got worse results when they used AI, because they appeared to accept lower-quality output than they would have produced on their own. But the overall finding is striking: AI use during early stages of learning the law appears to strengthen legal reasoning, even after the AI is removed.

The hypothesis on why is interesting. Getting the “CliffsNotes” version right away, without going down rabbit holes of misunderstanding, may help students build a better mental model of the doctrine in the first place. Rather than replacing thinking, AI might accelerate the foundation that makes deeper thinking possible.

This study is an empirical study, and the researchers’ starting hypothesis turned out to be wrong. That’s harder to dismiss than a thought experiment. It also cuts directly against Berkeley Law’s policy.

The legal research carve-out is backwards

One specific piece of the policy deserves its own scrutiny. That’s the policy’s carve-out for using AI in legal research.

Berkeley’s policy says AI can be used “for the limited purpose of identifying sources, such as cases, statutes, or secondary sources.” This is the one thing that generative AI foundation models are worst at. ChatGPT, Claude, and Gemini hallucinate cases more than any other category of legal output. The 1,400+ documented cases worldwide of AI-generated errors making it into court filings are overwhelmingly bad case citations, not bad brainstorming.

The carve-out doesn’t distinguish between general-purpose chatbots and purpose-built legal AI tools with hyperlinked citations to real authority. A student could comply with the policy by using ChatGPT to “identify sources” for a paper, and walk straight into a citation to a case that doesn’t exist. A different student, using a legal AI tool grounded in real case law, would get verifiable citations every time.

If UC Berkeley had drawn a line between general-purpose chatbots and verified legal AI, the carve-out would make sense. Instead, the policy permits exactly the kind of AI use most likely to produce errors and prohibits the kinds of use (brainstorming, summarizing) where errors are lowest-stakes.

Beyond Berkeley Law’s AI policy: How should law schools teach AI?

Policies like Berkeley’s are answers to a question most law schools haven’t been asked to articulate out loud yet. How should law schools train lawyers to use AI well?

The calculator analogy is useful here. Math students learn arithmetic before they’re given calculators, so that when they later use calculators, they can recognize when an output is wrong. The discipline of mathematics expanded in the calculator era. The same is true of accounting and spreadsheets. The power tools didn’t replace the thinking. Instead, they raised the ceiling on what was possible.

The same logic probably applies to AI in law. The best AI users in legal practice tend to be lawyers with strong doctrinal foundations. They recognize when an output is missing an exception, when the analysis lacks depth, when the tool cites only one case but misses the seminal case. They treat AI output as a first draft. That critical judgment is what law schools are trying to teach, and judgment is hard to build if students are either (a) never exposed to AI or (b) using AI as a substitute for the underlying analysis.

The open question is how long the doctrinal training needs to last before students are turned loose with AI. Some professors will argue to ban AI for all three years. Some will argue a single semester. The most favorable policy is probably in the middle, with significant variation by subject area and by students’ planned post-graduation work. 

There’s also a deeper change happening in the work itself. Many of the tasks that junior associates used to do—research, first drafts, due diligence, document review—are increasingly being done by machines. Those tasks were how associates built the tacit doctrinal expertise that makes great lawyers. If the tasks go away, the training opportunities go with them. Law schools are choosing more than whether to allow AI. They’re choosing how to develop tacit expertise when the traditional pipeline is changing underneath them.

Why this is really an assessment problem

Strip away the AI debate, and Berkeley’s policy is really about something else entirely: how to grade students fairly.

Law schools across the country, like all of higher education, are wrestling with how to fairly assess students in a world where some are using AI and some aren’t. How do you grade a paper written manually with grammatical errors against a paper polished by Gen AI with no errors? How do you cold-call a class when half the students might be typing the question into ChatGPT and reading back the answer in real time? How do you give a take-home exam that some students are completing themselves and others are completing with significant AI assistance?

Many law schools are getting rid of papers. Others are moving back to handwritten exams. Some are abandoning take-home assignments entirely. The assessment infrastructure that supported a half-century of legal education is being rebuilt in a two-year period.

It’s a little like the recent enhanced games in athletics, where some athletes openly used performance-enhancing drugs and competed against others who didn’t. Some of the unenhanced athletes still won. But the comparison highlights how hard fair assessment becomes when participants have radically different toolkits.

Berkeley’s policy is, in part, an attempt to solve that assessment problem by removing the variable. If AI use is uniformly prohibited, the assessment can be uniform too. That’s an understandable goal, even if the means are debatable.

What UC Berkeley Law’s AI policy means for law schools

Berkeley Law’s AI policy isn’t the final word on any of this. It’s an early move from a renowned law school, and it deserves to be evaluated for what it is rather than dismissed as a refusal to engage with AI. A few things to take away from it.

• Clarity is a virtue. A clear default rule, paired with an opt-in mechanism for instructors who want to teach AI, is more useful than the ambiguous silence at many other law schools.
• Citation responsibility is the right norm. Holding students responsible for the accuracy of every cited source, with a presumption that fake citations are AI-generated, is going to be standard practice in the profession. Teaching it now is the right call.
• The empirical evidence complicates the case. The Schwarcz study suggests that AI use during early-stage learning may strengthen legal reasoning, not weaken it. Future policies should account for that.
• The legal research carve-out is backwards. General-purpose chatbots are the worst at the exact task Berkeley permits. A better-drawn policy would distinguish between foundation models and purpose-built legal AI tools with hyperlinked, verifiable sources.
• The real question is assessment. Behind every law school AI policy is a deeper question about how to fairly evaluate students in a world where some are AI-augmented and some aren’t. Berkeley’s policy is one answer. There will be many others.

The competitive picture in legal AI keeps shifting, and law school policies are going to keep shifting with it. Berkeley made an early, strong move. It may or may not look right in five years. What’s certain is that the assessment problem and the tacit-expertise problem aren’t going away, and the schools that figure out how to teach AI literacy and rigorous doctrinal thinking at the same time are going to produce the lawyers who do best in the profession that’s emerging.

Used well, AI raises the ceiling on what law students can learn. Used carelessly, AI short-circuits the very expertise that makes lawyers most able to supervise AI use effectively in practice. The work ahead for legal education is to strike the right balance, advancing traditional doctrinal coursework, assessing students fairly, and preparing them to practice in a changed profession.