Within the span of a few weeks, three law schools made decisions about artificial intelligence through three different instruments. Berkeley Law (my alma mater) adopted a default rule, effective Summer 2026, that forbids students from using AI to conceptualize, outline, draft, revise, translate, or edit any work submitted for credit, on the ground that “thinking remains the sine qua non of good lawyering.” The dean of the University of Texas School of Law, Bobby Chesney, sent his faculty a letter rejecting the idea that AI training and traditional rigor sit in “a zero-sum relationship,” and moved to protect assessment by supervision rather than prohibition. And Boston College Law appointed its first Faculty Director for AI Initiatives, charged with teaching students to use the tools “without losing critical human judgment and decision-making skills.”
In tone, the three read as disagreement. All three, though, serve the same unexpressed goal: to graduate students who can analyze, research, write, and exercise professional and ethical judgment, and to assess whether they have learned to do so (ABA Standards 302 and 314). AI strains that mandate from both sides, threatening the skills the standards require while adding a new one, the ability to supervise a machine that now produces legal work. What divides Berkeley, Texas, and Boston College is the instrument and the sequence: when and how to build legal judgment and AI fluency into the same lawyer.
The exceptions in Berkeley’s rule
Berkeley’s policy is a careful statement of the first competency. It forbids AI for conceptualizing, outlining, drafting, revising, and editing any credited work, bars it entirely from exams, and prohibits students from uploading readings, slides, or recordings into a model, all to preserve the skills a lawyer builds by doing the work unaided. The policy then carves out one permission: AI may be used “for research on papers ONLY for the limited purpose of identifying sources, such as cases, statutes, or secondary sources.” Read against current practice, that exception all but concedes that legal research now runs through AI. The AI-assisted research built into Westlaw, Bloomberg, and Lexis+ AI is rapidly becoming central to how lawyers and students find authority; using those platforms almost requires using generative AI. My read is that Berkeley could not forbid AI-assisted research without forbidding research altogether as it is now conducted, hence the carveout.
The default prohibition includes one other exception, as well. The policy lets instructors set their own rules for courses “designed intentionally to teach AI fluency,” and Berkeley is launching one this fall: AI and the Practice of Law, a three-credit course for 2Ls and 3Ls with more than 100 hours of hands-on training, taught by Wayne Stacy, executive director of the Berkeley Center for Law and Technology. Stacy endorses the default ban and excludes 1Ls on the same reasoning, treating the course as what comes after the foundation rather than a departure from it. What distinguishes one lawyer from another once everyone holds the same tools, he says, is judgment, “and we need to build that first.”
In effect, Berkeley is seeking to defer the use of AI until students have built a proper foundation. Once the foundation is in place, an upper-level elective stands ready to teach it. However, by holding AI off by default and concentrating its instruction in a single upper-level elective, Berkeley also signals to the AI-forward firms its graduates are entering that the degree does not by itself certify readiness to practice with these tools; that readiness, the supervisory judgment I have described before as evaluating generated work rather than producing it, arrives only if students have taken Stacy’s course. Berkeley may have sound reasons for drawing that line, but the policy does not explain them.
While the policy’s pedagogical goal is defensible, its reach is not: it forbids far more than it can detect. Whether a student used AI to brainstorm a topic, outline an argument, tighten a paragraph, or upload a reading into a model leaves essentially no trace. Apart from the proctored exam and the presumption that a fabricated citation signals prohibited use, the rule runs on honesty it cannot verify. More importantly, an unenforceable prohibition does not bind evenly; it constrains the students who comply and quietly rewards those who break it and go uncaught, which inverts what an integrity rule should do and may even exacerbate inequity. A prohibition this broad would do more as a statement of the faculty’s values than as a disciplinary rule.
Texas separates the challenges
In a June 16 letter to his faculty, University of Texas School of Law Dean Bobby Chesney separates the three challenges posed by AI: a skills challenge (what AI competence to teach), an assessments challenge (how to keep graded work honest when AI is everywhere), and an educational-rigor challenge (the de-skilling risk). His letter rejects the idea that traditional pedagogy and AI training sit in “a zero-sum relationship,” and Texas pursues the skills side in earnest, giving students access to Harvey, Legora, Westlaw, and Lexis and writing learning objectives that name “the signal importance of output verification” and the capacity to produce “verified work product.” Those objectives describe the supervisory competence directly, the one Berkeley’s default forecloses.
On the assessment challenge, Chesney is blunt that “some traditionally common assessment methods are conspicuously vulnerable,” the take-home exam most of all. Texas nearly eliminated take-home exams, moved to in-class exams under lockdown software, and added oral and participation components, relocating graded work to settings where unaided performance can be observed. In instructional design, this is known as constructive alignment: an assessment should measure the outcome it is meant to measure. A paper an AI chatbot can write at home measures the chatbot’s drafting ability, and, generously, the student’s ability to prompt; a supervised exam measures the student’s reasoning. For the harder case of the writing seminar, Chesney points to colleagues who require students to document their AI use at each step and grade “the quality of that AI use,” which he calls “aligning evaluation with important emergent skills”; this approximates the logic of backward design: beginning from the capability graduates should have and working back to an assessment that reveals it. He admits the seminar problem is not yet solved.
On rigor, the letter locates the problem with care, tracing de-skilling to Plato’s Phaedrus while noting that not every skill a new technology displaces counts as a loss. Because AI cannot be reliably excluded from any unsupervised setting, “supervised settings have become even more valuable than they used to be,” the “sole context in which a professor can be certain that a student is engaged in the hard but essential work of analysis and communication unmediated by AI.” Rather than ban the tool, Chesney would teach against the temptation and make the minutes of class “an ever-more valuable asset” through sustained Socratic dialogue. I have written that the functions only a human instructor performs, challenging a student’s reasoning and reading the question behind the question, are what scarce class time should protect; the letter reaches the same conclusion from a dean’s chair.
Boston College’s appointment
Boston College’s response to the AI crisis in legal education was to appoint a faculty director. Maureen Van Neste, named Faculty Director for AI Initiatives, will shape classroom policy and curriculum around what Dean Odette Lienau called using the tools “responsibly, without losing critical human judgment.” Van Neste described AI fluency as more than a technical skill: “We want students asking the right questions, about the tool and about what it produces. Should we use this technology for this task at all? Whom does it serve? Is the work product accurate and complete?” Those questions are a description of the supervisory competence, and treating it as something to be taught and assessed, through a curriculum, a student AI badge, and a faculty director who sets classroom policy, supplies the affirmative version of what Berkeley’s default leaves out. The school’s four governing principles, set in 2023, describe the capability in the same terms: graduates should be able to use the tools, “assess and benchmark a tool’s work product,” fit them within a client strategy, and weigh their ethical implications.
The Jesuit language of discernment gives Boston College a vocabulary for the judgment question that the other two policies gesture at without stating outright, and the appointment sits alongside the university’s new Krantz Institute for Artificial Intelligence, Ethics, and Humanity. Whether a faculty director and a set of principles produce the competence they describe is a harder question than whether they articulate it well, and the curriculum that would deliver it has not yet been built. The aspiration is stated clearly; the assessments that would make it more than aspiration are still to come.
What the three are really deciding
The three schools’ responses to this moment reflect a question all law schools are—or should be—asking: once an AI tool can produce competent legal work, what should a law school assess, and where, and how? That question is partly about accreditation. Standard 314 requires every school to assess whether its students reach the stated learning outcomes, and AI has dissolved the old assumption that a take-home assignment measures unaided ability. It is also about ethics: the duty of technological competence under Model Rule 1.1, comment 8, a binding obligation to understand the benefits and risks of relevant technology, points toward teaching supervised use rather than away from it. A graduate allowed to touch AI only at the margins of the curriculum has not been trained in the competence the rule requires of them.
I am not completely opposed to Berkeley’s position as reflected in its policy. The strongest argument for its default is that skill is sequential: judgment about a model’s output presupposes the underlying ability the output imitates, and a student cannot supervise a draft she could not have written herself. During the first year, when those foundations are laid, a prohibition is defensible triage rather than technophobia. But whether a default prohibition, even paired with opt-in courses like Stacy’s, builds the supervisory competence widely enough is unclear to me. Berkeley already grants through its research carve-out that finding and evaluating authority is now done with AI, yet outside that carve-out and the courses instructors design for it, supervising a model stays an elective rather than a baseline of the degree. The past year’s citation sanctions have fallen on lawyers who could not tell a real authority from a fabricated one, a judgment the profession now needs broadly and early, not only from the students who opt into the class that teaches it.
This post draws on Berkeley Law’s Artificial Intelligence Policy, effective Summer 2026, and the school’s announcement of Wayne Stacy’s AI and the Practice of Law course; Dean Bobby Chesney’s June 16, 2026 letter to the Texas Law faculty, “AI and Legal Education” (as reported by Bloomberg Law); and Boston College Law’s announcement of Maureen Van Neste’s appointment as Faculty Director for AI Initiatives, together with the university’s Krantz Institute for Artificial Intelligence, Ethics, and Humanity. It builds on earlier posts on the delegation framework, answer quality versus learning, and revised Standard 314.