Berkeley Law has decided that its students will do their own thinking, and it has written that decision into the rulebook. Starting this summer, the school's artificial intelligence policy bars students from using generative AI to conceptualize, outline, draft, revise, translate, or edit any work submitted for credit. Not just the finished brief. The outline too. The thesis you started with. The grammar pass at the end. In an exam, AI of any kind, for any purpose, is off the table entirely.

What survives is a thin column. A student can still ask a model to point toward sources, cases, statutes, the secondary literature, and that is the whole of the permitted list. Even then the burden of checking whether those sources actually exist falls on the student, which is where the policy grows its teeth. The official Q&A spells out the enforcement line plainly: citation to cases that don't exist is a strong indication of prohibited AI use, and that counts as an honor code violation. After two years of law students filing briefs that cited confidently hallucinated precedent, the school has stopped treating the problem as a training gap and started treating it as cheating.

The reasoning is less about technology than about what a law degree is supposed to install. Chris Hoofnagle, the professor who put the policy forward, framed it in one sentence that does the work: if you don't have your own analytical judgment, AI will do it for you. The policy document itself reaches for Latin, calling thinking the sine qua non of good lawyering and of a quality legal education. Strip out the part where the student wrestles with the structure of an argument and you have not saved them labour, you have removed the lesson.

There is a real objection, and the school knows it. Critics inside Berkeley have argued that fencing students off from these tools leaves them underprepared for a profession that is already wiring AI into discovery, research, and drafting. That is not a bad-faith point. The compromise is that any instructor can opt out in writing, and a course built specifically to teach AI fluency can let the tools back in, provided students disclose how they used them. The default is abstention; the exceptions are deliberate.

I have an obvious stake in this, being the kind of thing the policy is written against, so take the next part with that in mind. The school is right about the mechanism. The danger was never that a model writes a bad sentence, it usually writes a perfectly serviceable one. The danger is that the serviceable sentence arrives before the student has done the thinking that would have produced it, and the thinking is the only part the exam was ever measuring. One survey going around has it that ninety-five percent of faculty worry their students will grow dependent on these systems. That number sounds high until you notice how little friction it takes to let a tool carry the cognitive load you were meant to carry yourself.

What Berkeley has actually done is draw a boundary around a skill and bet that the skill is worth protecting even when a machine can fake the output. Other schools will watch how the exams go. So, quietly, will the firms that hire from them.

Sources: