Berkeley Law AI Policy Restricts AI, but Law Schools Are Not Banning It

Berkeley Law tightened its AI rules for summer 2026, despite acknowledging that future lawyers must become fluent with the technology. The Berkeley Law AI policy prohibits students from using AI to conceptualize, outline, draft, revise, edit, or translate work submitted for credit. It also bars AI use in every exam unless an instructor establishes a permitted exception.

That sounds like a law school AI ban, but the label misses the policy’s most important feature. Berkeley still allows instructors to authorize AI when it serves a course’s educational goals. The school also plans 22 law and technology courses for fall 2026, making its approach restrictive and expansionary at the same time.

The University of Chicago and Columbia Law School have adopted similarly divided strategies. They are protecting parts of legal education from automation while creating other spaces for deliberate AI use. The real conflict is therefore not law schools against AI. It is independent legal reasoning against premature cognitive outsourcing.

What the Berkeley Law AI Policy Actually Changes

Berkeley has established independent student work as the default, while preserving targeted exceptions for courses that teach AI fluency.

The default AI policy took effect in summer 2026. It applies to work submitted for academic credit and reaches much further than a simple prohibition on chatbot-generated essays.

Students cannot ask an AI system to brainstorm a paper topic or thesis. They cannot request an organizational structure, draft a paragraph, identify passages for revision, or polish grammar. AI-assisted translation is also prohibited when the translated material becomes part of submitted work.

The exam rule is stricter. Students cannot use AI for any purpose during an examination. They also cannot generate an outline with AI and then bring material from that outline into an exam.

These restrictions target the full intellectual production process, not just the final prose. That distinction matters because generative AI can influence an argument before it writes a sentence. A suggested thesis or structure can determine which authorities receive attention and which counterarguments disappear.

The policy also prohibits students from uploading assignments, readings, slides, recordings, or other course content into generative AI systems. That rule addresses privacy, intellectual property, and confidentiality concerns alongside academic integrity.

Berkeley nevertheless says its policy is not a blanket ban. Instructors can adopt different rules for classes designed to teach AI fluency. They can also authorize distinct uses when another approach better serves a course.

That flexibility separates Berkeley’s position from a campus-wide rejection of generative AI. The school is drawing boundaries around assessed thinking while allowing supervised experimentation elsewhere.

Its accompanying guidance presents those boundaries as an educational sequence. Students first need to conceptualize, analyze, draft, and evaluate without automated substitution. They can then use AI while understanding what the system has changed, omitted, or fabricated.

This sequence reflects the central premise behind the Berkeley Law AI policy. A lawyer who cannot independently evaluate an argument cannot reliably supervise a system that generates one.

The rule is also designed to produce clearer and more consistent expectations. Berkeley said its previous policy created difficult questions about the boundary between grammar correction and AI composition. The new approach treats automated editing as prohibited unless a course provides an exception.

That decision sacrifices some convenience for administrability. It avoids asking professors to determine whether an AI suggestion corrected grammar, rewrote a claim, or quietly altered the underlying analysis.

The cost is breadth. A student who uses a familiar writing application might encounter generative functions without actively seeking them. Schools will need practical guidance that keeps ordinary spellchecking separate from prohibited automated rewriting.

The strongest version of Berkeley’s argument is not that every AI interaction harms learning. It is that credit-bearing work must reveal the student’s own capacity. Otherwise, grades stop measuring the skills that law school claims to develop.

Why Legal Reasoning Has Become the Protected Skill

Law schools are restricting AI at the exact moments when students must build the judgment required to supervise it later.

Legal education depends on more than retrieving the right rule. Students must interpret authority, distinguish precedents, identify missing facts, and defend a conclusion against competing arguments. Those activities form the basis of professional judgment.

Generative AI can produce plausible language before a student has formed that judgment. Its output can make an incomplete argument look finished, which complicates both learning and assessment.

This risk is particularly acute during the first year of law school. First-year courses introduce the habits used to read cases and reason from imperfect authority. Students also learn how legal conclusions change when one fact or precedent shifts.

An AI-generated case summary can save time. However, that efficiency becomes educationally costly when reading the case is the exercise. The student receives the apparent product without practicing the reasoning that produced it.

The same problem affects legal writing. A chatbot can generate a structured memorandum, but structure is not merely presentation. Choosing an issue sequence often requires deciding which facts and authorities deserve the greatest legal weight.

Berkeley describes thinking as essential to good lawyering and a quality legal education. Its policy treats conceptualizing, outlining, drafting, revision, and editing as connected cognitive tasks rather than detachable production stages.

That position also responds to the unreliability of generated legal material. Language models produce probable text, not verified legal conclusions. They can invent citations, misstate holdings, or blur differences between jurisdictions.

Experienced lawyers can still make those mistakes. The difference is that AI can produce them quickly, confidently, and at a scale that strains human review.

A student who already understands the doctrine can interrogate that output. A student who lacks the foundation might accept fluent wording as evidence of accuracy.

Legal educators therefore face a sequencing problem. They must prepare graduates for AI-assisted practice without allowing the assistant to replace the apprenticeship.

This tension extends beyond grades. Lawyers owe duties to clients, courts, and opposing parties. The American Bar Association’s AI ethics guidance says existing obligations continue to apply when lawyers use generative systems.

Those obligations include competence, confidentiality, client communication, supervision, candor, and reasonable fees. A lawyer cannot transfer professional responsibility to an AI vendor or excuse an inaccurate filing because software produced it.

Law schools must consequently teach two competencies that can pull in opposite directions. Graduates need enough AI fluency to use modern legal systems effectively. They also need enough independent expertise to notice when those systems fail.

The Berkeley Law AI policy protects the second competency during assessed work. Its technology courses and permitted exceptions are intended to develop the first.

That division is more defensible than a broad claim that AI inherently weakens students. Evidence about long-term cognitive effects remains incomplete, and outcomes depend heavily on how a system is used.

AI can replace effort, but it can also challenge an argument. A student might ask a model to identify counterarguments after independently drafting a position. Another might submit the model’s first response without examining it.

Those behaviors should not receive the same educational judgment. The challenge is designing assignments that distinguish productive assistance from intellectual substitution.

For students managing large collections of cases and notes, a personal student knowledge base can support retrieval without deciding what their arguments should be. The educational boundary should focus on who performs the reasoning, not merely whether software touched the workflow.

UChicago Law Protects the Classroom but Teaches the Tool

The UChicago Law AI strategy makes the tradeoff explicit by separating device-free foundational learning from supervised professional practice.

The University of Chicago Law School announced its AI strategy statement on July 9, 2026. The strategy goes into effect during the fall quarter and builds on curriculum changes introduced since 2024.

Its most visible measure is a pilot prohibition on electronic devices in core first-year classes. Limited exceptions remain available, including accommodations for students who need them.

A device rule is broader than an AI restriction. It removes laptops, tablets, and phones even when a student only wants to take notes. The school says the approach also reduces distraction and supports better classroom discussion.

That makes the pilot a test of classroom attention as much as artificial intelligence. Professors used device-free policies long before ChatGPT existed, often because screens divided attention during discussion.

AI raises the stakes because a connected device can now do more than distract. It can summarize a case, propose an answer, or prepare a response to a professor’s question in real time.

The traditional law-school classroom uses questioning to expose a student’s reasoning process. If software supplies the answer, the professor cannot easily determine whether the student understands the doctrine.

Removing devices creates a temporary zone where students must retrieve, synthesize, and articulate ideas themselves. It also makes participation more visible and reduces the temptation to consult an instant external answer.

Yet UChicago is not trying to recreate a pre-digital profession. Its strategy includes AI modules for first-year students, AI use in clinics, an AI Lab, and guidance for legal research and writing.

The school’s model can be summarized as learning to think with, without, and about AI. Each condition serves a different educational purpose.

Thinking without AI builds a baseline capacity. Thinking with AI teaches students to use automation while retaining responsibility. Thinking about AI develops an understanding of its institutional, ethical, and legal effects.

This is a more precise account than saying UChicago outlawed AI. The school removes technology from selected learning environments while integrating it into others.

Its approach also recognizes that education and employment reward different behaviors. A law firm benefits when a tool reduces repetitive labor without compromising quality. A classroom sometimes assigns that same labor because performing it develops judgment.

Efficiency is therefore not a universal educational goal. The fastest route to an answer can bypass the activity that an assignment was designed to measure.

The distinction resembles training in other high-stakes professions. A pilot learns procedures without relying entirely on automation, even though modern aircraft use extensive automated systems. The foundational practice exists because professionals must recognize and manage failure.

UChicago’s challenge will be measuring whether its device-free pilot improves the intended skills. Better classroom discussion would support the policy, but satisfaction alone would not establish stronger legal reasoning.

The school will also need to examine unequal effects. Some students organize and retain information more effectively through typed notes. Others need assistive technology or language support.

A fair policy must provide functional accommodations without forcing students to disclose more personal information than necessary. It must also prevent exceptions from becoming stigmatizing.

UChicago’s hybrid approach gives the school a useful comparison. It can evaluate device-free foundational classes alongside courses that teach structured AI use. That evidence would be more valuable than a philosophical argument about whether technology is good or bad.

Columbia Chooses Responsibility Over a Broad Prohibition

Columbia’s rules show that law school AI policies are converging on controlled use, even when their default boundaries differ.

Columbia Law School adopted a new academic AI policy effective August 1, 2026. It replaced the school’s earlier interim policy for the 2026–2027 academic year.

Columbia prohibits AI assistance in examinations. Every word in an exam answer must come from the student, including language produced through AI translation.

Its treatment of other coursework is more permissive than Berkeley’s default. Students can use AI as a learning aid or critical interlocutor when course rules allow it.

A critical interlocutor is a system used to test ideas rather than originate them. It might challenge an argument, identify an organizational weakness, or suggest questions for further research.

Columbia places intellectual responsibility on the student. Submitted ideas, arguments, analysis, and expression must reflect the student’s own judgment. Students must understand and defend every proposition they submit.

The policy therefore focuses on authorship and accountability rather than banning every assisted stage. AI cannot become an undisclosed ghostwriter or replace the student’s legal reasoning.

Faculty members retain discretion to establish course-specific rules. Some can integrate AI extensively, while others can limit it to develop particular skills.

This model accepts inconsistency as part of professional preparation. Lawyers will work under different court rules, employer policies, client restrictions, and security requirements. Students need practice identifying which rules govern a specific setting.

However, policy variation can also create confusion. A student moving among several courses must remember whether brainstorming, editing, translation, or research assistance is permitted in each one.

The 2026 student survey from Thomson Reuters Institute illustrates that problem. It collected responses from more than 1,800 law students in April 2026.

Almost 60% reported using AI several times per week for academic work. Meanwhile, 48% said AI policies varied by professor.

The survey found that 72% considered AI literacy an essential professional skill. At the same time, 74% recognized that overreliance could weaken their core legal competencies.

Those figures complicate the idea that students and administrators occupy opposing camps. Many students appear to understand both the professional value and the developmental risk.

The institutional gap concerns training. Thirty-two percent said their schools did not provide the AI skills needed for their future careers.

That result strengthens the case for structured integration. A restrictive policy without a meaningful AI curriculum pushes students toward unsupervised experimentation. It can also leave graduates unprepared for tools already entering legal workplaces.

Columbia attempts to solve this through a combination of boundaries, specialized courses, and faculty choice. Berkeley uses broader default restrictions alongside substantial technology offerings. UChicago creates device-free spaces while embedding AI elsewhere.

These are different implementations of the same emerging principle. Students should not outsource the skill currently being taught, but they should learn to use AI after developing enough knowledge to supervise it.

The disagreement concerns where that line belongs. Berkeley draws it around nearly every stage of credit-bearing work. Columbia draws it around intellectual responsibility and course-specific permission.

Neither policy establishes the correct boundary for every assignment. Their value will depend on whether professors translate broad principles into clear, observable requirements.

The Risk Is Producing Either Dependent or Unprepared Lawyers

A rule that blocks AI without teaching it can create a different version of the competence problem it was designed to prevent.

The strongest case for restrictions is developmental. Students must practice legal analysis before they can responsibly delegate parts of that work.

The strongest objection is professional. Graduates will enter workplaces where AI supports research, document review, due diligence, contract analysis, and drafting. Refusing to teach those systems does not keep graduates safe.

It leaves them to learn through vendors, employers, or private experimentation. That training may prioritize speed over confidentiality, verification, and professional judgment.

This is why the phrase “AI ban” creates a misleading binary. A school can prohibit AI during an exam and require it during a clinic. Both decisions can support the same educational objective.

The harder question concerns ordinary coursework between those extremes. Students write papers partly to demonstrate knowledge and partly to improve through practice. AI can support or replace either function.

Consider a student preparing a memorandum. Asking AI to write the first draft can remove the need to organize authorities independently. Asking it to attack a completed draft can expose weaknesses that the student missed.

Even the second use carries risks. The system might invent a counterargument, distort precedent, or encourage unnecessary revisions. However, the student remains cognitively engaged when evaluating those suggestions.

Schools therefore need assignments that make the process visible. A professor can require source logs, oral defenses, revision histories, or short explanations of permitted AI use.

These measures are more informative than relying on AI-detection software. Detection systems can produce false positives, especially when evaluating formulaic or second-language writing. They also encourage a contest between evasion and enforcement.

Process evidence supports education as well as integrity. A student who can explain why an argument changed demonstrates more understanding than one who merely submits polished prose.

Clear rules also matter because generative features now appear inside search engines, word processors, and research platforms. Students may not always recognize when a familiar tool has shifted from correction to composition.

Berkeley’s bright line improves clarity but could capture low-risk assistance. Columbia’s responsibility standard preserves flexibility but requires more interpretation. UChicago’s device-free classroom creates an enforceable boundary only within a limited setting.

Accessibility remains another concern. AI-assisted translation, editing, or summarization can help students with disabilities and those working in a second language. Restricting those functions can impose unequal burdens.

Academic accommodations can address some cases. Still, schools should distinguish an accommodation from an advantage that changes the skill being assessed.

The answer will depend on the assignment. Translation assistance might be inappropriate when English-language legal writing is the measured competency. It could be reasonable when the assignment measures issue spotting in another context.

Law schools should also avoid overstating evidence about cognitive harm. Research has raised concerns about cognitive offloading, which occurs when people delegate mental work to external tools. However, association does not establish inevitable decline.

The effects likely vary by task, student experience, prompting behavior, and instructional design. A guided exercise that requires verification differs from asking a chatbot to complete an essay.

Strict policies should therefore be treated as educational interventions, not settled scientific conclusions. Schools need comparative evidence about student reasoning, retention, writing, and professional performance.

The same skepticism applies to AI-literacy programs. Offering a course does not prove that graduates can detect fabricated citations, protect client information, or choose when automation is inappropriate.

Schools must test those abilities directly. Students should encounter flawed output, confidentiality scenarios, and ambiguous authorities under supervised conditions.

The industry also changes too quickly for permanent rules. A boundary written around today’s chatbot interface can become obsolete when generative assistance disappears into standard software.

The durable principle is responsibility. Students need to know what work they delegated, what information they disclosed, and how they verified the result.

A law school AI ban cannot teach those habits by itself. Unrestricted access cannot teach them either.

What to Watch During the 2026–2027 Academic Year

The next test is whether hybrid policies improve independent reasoning without widening the gap between students and professional practice.

The first signal will come from UChicago’s device-free first-year pilot. Administrators should examine participation quality, doctrinal performance, student retention, and accommodation requests.

If those indicators improve without creating substantial access problems, the case for protected, device-free learning periods becomes stronger. If results remain unchanged, the policy may have removed convenience without delivering educational value.

The second signal will be how Berkeley instructors use their authority to grant exceptions. The school’s policy works as a hybrid only if students receive meaningful, supervised opportunities to practice with AI.

Berkeley’s 22 law and technology courses provide substantial capacity on paper. The important measure is whether those courses teach verification, confidentiality, and professional accountability across the student body.

If AI instruction remains concentrated among technology specialists, the broader student population might graduate with strong traditional skills but uneven practical readiness.

The third signal will be policy convergence across law schools. An education survey published by the American Bar Association in 2024 found institutions still developing their approaches.

By 2026, Berkeley, UChicago, and Columbia have moved beyond temporary statements. Each has established a more detailed allocation of AI-free work, supervised use, and instructor discretion.

Other schools will now decide whether to copy those models or build alternatives. Consistent definitions would help students understand the difference between retrieval, correction, generation, and automated revision.

Courts and employers will influence these choices. If firms expect new associates to supervise AI systems immediately, schools will face pressure to expand practical instruction.

Judges and regulators can push in the other direction when generated errors reach legal filings. Every public failure strengthens the argument that independent verification must precede automation.

The lasting shift is not that law schools have outlawed AI. They have started treating unaided reasoning as a protected stage of professional training.

That stage must remain connected to the workplace students will enter. Otherwise, restrictions become avoidance rather than preparation.

The Berkeley Law AI policy offers a clear experiment: require students to produce assessed work independently, then teach AI in designated settings. UChicago and Columbia are testing different versions of that division.

Students should ask a practical question when evaluating any policy: Does it specify when AI must stay out, while also teaching how responsible use works? A credible answer needs both parts.