As Lawyers Face Sanctions for A.I. Errors, Law Schools Rethink Training

Earlier this month, we wrote about the University of Chicago Law School restricting devices in its core 1L classrooms to protect how students learn to think.

We then examined Berkeley Law’s new AI policy, which makes no-AI the default for most work submitted for credit unless a professor expressly authorizes it.

Now, UC Law San Francisco is taking the opposite approach, requiring every JD student to complete an AI-enabled lawyering lab during their second year.

The requirement arrives after several years of courts sanctioning lawyers for filing fabricated authorities and other unverified AI-generated material.

In 2023, two New York lawyers were sanctioned after submitting six fictitious cases generated by ChatGPT. The problem was not discovered through some sophisticated technical audit. Opposing counsel and the court simply could not find the cases.

The problem has since reached appellate courts. In May 2026, the Supreme Court of Georgia vacated a trial court order after an assistant district attorney submitted AI-generated authorities that included nonexistent cases, inaccurate quotations, and misrepresented holdings. The prosecutor was suspended from practicing before the court for six months and ordered to complete additional training.

The following month, the Ninth Circuit suspended two lawyers after briefs contained nonexistent cases and quotations attributed to real opinions that did not contain them.

The recurring failure was not simply that lawyers used AI.

It was that they treated plausible output as verified law.

Against that backdrop, one school is restricting technology. Another is restricting AI. A third is requiring it.

They are not actually disagreeing.

They are beginning to define when AI belongs in legal education, and when it does not.

The timing matters

UC Law SF’s six-session lab covers legal research, drafting, workflow design, confidentiality, accuracy, privilege, bias, and professional responsibility. Students will also build an AI tool as a capstone project, often without prior coding experience.

But the most important detail is when the requirement occurs.

Students will complete it during their 2L year, after spending a year developing the legal reasoning, research, and writing skills necessary to evaluate what an AI system produces.

That is remarkably consistent with Chicago’s approach.

Chicago is protecting the core 1L classroom from technological interference while incorporating AI elsewhere in the curriculum. UC Law SF is building the foundation first and then requiring students to learn how technology fits into professional practice.

The sequence is the policy.

A student cannot reliably evaluate an AI-generated legal analysis without knowing what a sound legal analysis looks like. They cannot catch a missing issue, false citation, weak analogy, or flattened counterargument if they have never worked through those problems themselves.

AI should come downstream of effort, not upstream of understanding.

But once that understanding exists, refusing to teach AI leaves the education incomplete.

AI competence is not prompt competence

UC Law SF is also making the lab mandatory rather than leaving it as an elective.

That matters because AI will not remain confined to technology-focused practices. It is entering legal research, drafting, discovery, due diligence, document review, and ordinary firm workflows.

Every graduate does not need to become a developer. Every lawyer using AI does need to understand how the system can fail.

That means knowing whether a task should be delegated, what information can safely be entered, how the output will be verified, whether the cited authority supports the proposition, and who remains responsible when the answer is wrong.

The answer to the last question is still the lawyer.

The sanctioned lawyers did not need better prompt templates. They needed to open the cases, read the authorities, compare the quotations, and verify that the law supported what they were asking a court to do.

Building an AI workflow forces students to confront those decisions. They must define the task, the inputs, the expected output, the likely failure points, and where human review must remain.

Learning to govern an AI system's boundary conditions is vastly more valuable than memorizing prompt shortcuts.

Restrictions and requirements belong together

Berkeley’s default prohibition and UC Law SF’s requirement make sense once the purpose of the assignment is considered.

In a first-year writing course, the student’s ability to organize and express a legal argument may be the thing being assessed. Having AI perform that work defeats the assignment even when the final answer appears correct.

In an AI-enabled lawyering lab, the student’s ability to direct, test, constrain, and evaluate an AI system is the thing being assessed. Prohibiting AI there would defeat the purpose just as completely.

The useful question is not simply whether AI was used.

It is: What intellectual work was the student supposed to perform?

Sometimes protecting that work requires prohibiting AI. Sometimes it requires students to use AI and demonstrate that they can recognize its limitations.

What this means for JurisNote

These policies reinforce why JurisNote cannot be built around a single answer to the question, “Is AI allowed?”

The answer will vary by institution, professor, course, assignment, and stage of the student’s education.

Law schools should be able to use JurisNote’s briefing, outlining, flashcard, scheduling, and study tools without enabling generative AI. Institutions can also disable AI broadly or restrict particular categories of use.

When AI is enabled, it should remain visible and reviewable. JurisNote grounds AI answers in the student’s own materials, connects generated brief claims to supporting passages, labels generated content, and requires students to review generated flashcards before accepting them.

Those controls do not make AI inherently educational.

They preserve the student’s responsibility for deciding whether the output is correct.

Chicago validates the need for spaces where students work without AI. Berkeley validates the need for enforceable institutional boundaries. UC Law SF validates the need for structured environments where AI is deliberately taught as a professional tool.

An edtech platform built for law school cannot force a single philosophy onto a spectrum of pedagogy. It must support all three.

The future of legal education is not AI everywhere.

It is not AI nowhere.

It is AI used deliberately, at the right stage, for the right task, and by students who know enough to recognize when it is wrong.