The Worst AI Policy in Higher Ed
How Berkeley Law Boalt-ed From Expertise in Favor of Abstinence
Last week, one of the top law schools in the United States, the University of California, Berkeley School of Law, released its final policy on artificial intelligence, effective summer 2026. In the span of a breezy 1.5 pages, the school outlined the challenge AI poses to legal education and how it plans to address this problem. Despite these intentions, this AI policy is, in my estimation, the worst AI policy in higher education I have seen.
As the AI Faculty Fellow emeritus at my medium-sized public university, I have been privy to conversations around AI policy in higher education since my institution started an ad hoc Generative AI working group in early 2023. It has been a marathon, not a sprint. Both administrators and faculty agonized over how to balance consistent student experience with academic freedom, leadership imperatives with shared governance, and a rapidly evolving job market with calcified disciplinary skills. Standing on the front lines of this issue has exposed me to just how deliberative and collaborative a university can be when facing an existential threat. Our resulting approach may not be perfect, but it is nothing if not consciously aware of the need for flexibility.
Before critiquing Berkeley Law’s latest AI policy, let’s first examine what an AI policy should seek to accomplish. At a minimum, an AI policy should:
Identify the nature of the issue at hand (including its social and/or ethical dimensions);
Scope out the kinds of educational activities subject to the policy;
Articulate the conditions under which AI is permitted and/or prohibited;
Demonstrate feasibility in terms of implementation; and
Explain the consequences of non-compliance.
In the years since ChatGPT arrived, many universities have issued well-designed AI policies. While there exists no single, silver bullet solution to the creeping use of AI in the classroom, there are better and worse examples of thoughtful efforts to craft meaningful AI policy (you can find an indicative collection of university AI policies here).
For instance, Penn State’s Ethical and Responsible Use of AI Guidelines (15 pages long) offers a terrific exemplar. It distinguishes between different types of AI, describes how it builds on an existing framework for ethical decision-making, enumerates various ethical issues associated with AI use, and then lists a series of 22 questions that should be answered before arriving at a decision about whether to sanction or disallow AI in that context. Best of all, it is applicable to faculty, staff, and students. While perhaps it could be argued that this isn’t a policy per se, functionally it operates as a resource for determining whether or not the AI application in question runs afoul of university values (ergo, it is policy-adjacent).
By contrast, the AI policy promulgated by the faculty at Berkeley Law is a minimalist nightmare that is incoherent, unrealistic, and unworkable. Following on the heels of the law school’s pithy first iteration of an AI policy, which was adopted in 2023 and permitted AI for research purposes but not for use during exams, this new version begins by acknowledging that AI may become essential to lawyers in the “future” while implying that, at present, it erodes the cognitive capabilities of its users and thus is antithetical to a proper legal education, which demands the very kind of thinking AI strips from us. Then the policy proceeds by itemizing a list of prohibited uses long enough to make a teetotaler blush, so I am including it here in its entirety:
Asking an AI tool to brainstorm a paper topic or thesis (prohibited conceptualizing)
Asking an AI tool to propose an organizational structure for a paper (prohibited outlining)
Asking an AI tool to compose a paragraph summarizing a legal rule for use in a paper (prohibited drafting)
Asking an AI tool to identify repetitive passages in a paper that should be cut (prohibited revising)
Asking an AI tool to polish a paper by correcting grammatical mistakes (prohibited editing)
Asking AI to generate an exam outline, elements of which are then used on the exam (prohibited exam use)
Asking AI to translate a paper originally written in another language into English (prohibited translating)
There are at least three problems with the restrictions indicated above. First, there is the issue of detection. Unless all papers and exams are written under the kind of controlled environment one might expect from a surveillance state (i.e., a computer lab with lock down browsers and human proctors who check student eyeglasses upon entry), it will be practically impossible to assess whether the policy has been violated. If abrogations of this policy are not readily observable, this regulatory tool either relies on an honor system or is simply impotent at both catching and deterring unwanted behavior. In the language of game theory, the prospects of defection remain high and the undesirable actions are likely to continue unabated.
Second, there is a disjuncture between the prohibited uses and the reality of the legal field in the age of AI. For one, the policy conflates mundane applications of AI that virtually no one would construe as academic dishonesty (i.e., “correcting grammatical mistakes”) with those that could be considered plagiarism (i.e., “compose a paragraph summarizing a legal rule”). Next, since presumably students arrive at Berkeley Law able to faithfully engage in the craft of writing, it remains unclear how excluding AI from being deployed in some of these tasks preserves the development of the cognitive skills necessary to become an effective lawyer. Are they afraid that the attrition of intellectual capabilities will be so pronounced that the students they train will lose the ability to form sentences and arguments? If so, Berkeley may want to revisit their admissions criteria. Finally, as is the case with higher education writ large, the ironclad grasp faculty wish to maintain on certain analog skills belies the extent to which people use AI at work. For instance, a 2025 Clio study of over 1,700 legal professionals finds that 79% utilize AI in their firms (the proportion swells to 87% among those working for firms with over 50 employees) and 65% believe that AI improves the quality of their work.
Third, the policy is internally inconsistent to the institution. As the document states, “Instructors may deviate from the default rule for courses designed intentionally to teach AI fluency (or for other courses for which the instructor decides a distinct rule is pedagogically appropriate).” If AI is really that corrosive to the cognitive flourishing of future lawyers, why would it be allowed in some classes that emphasize AI literacy and others that find AI pedagogically valuable? Furthermore, for a tool that is alleged to be so detrimental to legal education, AI seems strangely prominent in Berkeley Law’s library website. Among the informational items found on the library’s page for “Generative AI Resources,” students can locate a list of generative AI tools, a “Generative AI Guide for Law Students” (the horror!), a worksheet on AI prompts for the legal profession, and access to Lexis+ AI, which allows “users to perform a variety of tasks from complex legal analysis to everyday brainstorming without switching between different AI tools.” For a law school that treats AI like the bane of legal education, it sure offers a lot of resources designed to help students…use AI for legal writing and research.
There are good reasons why institutions might want to enact an AI policy. Such rules, when written carefully, can offer guidance, clarity, and expectations regarding AI use for people internal and external to the organization. Many colleges and universities have struggled on this account, but some have successfully devised AI policy that is consistent with their values, comprehensive, and actually useful. Berkeley Law’s new AI policy has only managed to evince fidelity to its values, which we can only infer based on the strength of its conviction that AI is generally bad (while also being permissible in some contexts). To be sure, the policy is neither comprehensive nor useful.
It’s hard to believe that brilliant legal minds actually assembled this jumbled and ineffectual mess. Perhaps they should have asked an AI system like Claude whether their approach was likely to bear the desired fruit?1
I did just this. See the prompt and Claude’s resulting output below.
Prompt: You are a law professor at one of the best law schools in the United States. You understand that while the legal profession is increasingly relying on AI, you may not want to allow it in every circumstance in the legal education offered by your institution, as you are hoping to cultivate skills and practices that are foundational to the practice of law. Please review this draft of an AI policy, identify its shortcomings, and suggest alternative language that overcomes these deficits.
Claude (Sonnet 4.6): This is a substantive, well-drafted policy with real pedagogical seriousness behind it. Let me give you a thorough critique organized by deficiency, followed by suggested replacement language for each.
I. Structural and Definitional Deficiencies
1. “AI” Is Undefined
Problem: The policy uses “AI” and “generative AI systems” somewhat interchangeably, but never defines either term. This creates significant administrability problems. Does “AI” include:
Grammarly or spell-check?
Westlaw’s AI-assisted research tools (including its AI Brief Analyzer)?
Google Scholar’s citation alerts?
A word processor’s autocomplete?
A professor’s own AI-generated hypotheticals used in course materials?
The policy purports to promote fairness and administrability as an explicit purpose, but an undefined central term directly undermines both.
Suggested language:
Definitions. For purposes of this policy, “Generative AI” means any software system that uses machine learning models to generate text, images, code, audio, or other content in response to user prompts, including but not limited to large language model-based tools such as ChatGPT, Claude, Gemini, Copilot, and similar products, as well as AI-assisted drafting features embedded in other software platforms. “AI-assisted research tools” means tools offered by established legal research platforms (e.g., Westlaw AI, Lexis+ AI) in their source-identification functions, as distinguished from their drafting or summarization functions. Passive spell-check and grammar-flag features in standard word processors (e.g., Microsoft Word’s red underline) are not “Generative AI” for purposes of this policy; AI-powered rewriting and suggestion tools (e.g., Grammarly’s full-sentence rewrites, Copilot in Word) are.
2. The Research Carve-Out Is Dangerously Vague
Problem: The policy permits AI use “for the limited purpose of identifying sources, such as cases, statutes, or secondary sources.” This carve-out is both the most practically important provision and the most poorly drafted. It fails to address:
Hallucination risk: AI tools notoriously fabricate citations. The policy says students are “responsible for the accuracy” and that nonexistent citations “raise a presumption” of violation — but this is a remedial provision, not a preventive one. The policy effectively permits a use it knows to be unreliable without requiring any safeguards.
The line between identifying and summarizing: What happens when a student asks an AI “What cases discuss the nondelegation doctrine?” and the AI responds with a list plus a one-sentence description of each? Has the student used the summary in drafting?
Verification requirement: Students newly to AI should be taught that verification is mandatory, not merely that they bear responsibility. These are different emphases pedagogically.
Which tools? Westlaw’s AI functions are architecturally different from ChatGPT’s. The policy draws no distinction.
Suggested language:
Permitted Research Use. Students may use Generative AI solely to identify potentially relevant legal sources — cases, statutes, regulations, or secondary sources — subject to the following conditions: (a) Students must independently verify the existence and accuracy of every source so identified before citing it, using an authoritative legal research platform (e.g., Westlaw, Lexis+, HeinOnline); (b) Students may not use any AI-generated summary, description, or characterization of a source’s content as a basis for analysis, argument, or writing — the student must read the primary source; (c) Students must disclose in a footnote or end note any source first identified through Generative AI; (d) A citation to a source that does not exist, or that does not stand for the proposition for which it is cited, raises a rebuttable presumption of prohibited AI use and may constitute an academic integrity violation independent of this policy.
II. Gaps in Coverage
3. The Exam Provision Is Underspecified
Problem: “AI use is prohibited for any use for any purpose in any exam situation” is absolute and probably the right default rule, but it fails to define “exam situation.” Does this include:
Take-home exams?
Open-book in-class exams?
Timed writing competitions run under course credit?
Oral examinations with written preparation components?
The policy also does not address what happens when an instructor’s deviation from the default rule intersects with an exam — can an instructor permit AI on a take-home exam? The policy implies yes (instructor discretion) but doesn’t say so explicitly, creating a conflict with the absolute exam prohibition.
Suggested language:
Exam Prohibition. Generative AI use is prohibited for any purpose in any examination, including in-class examinations, take-home examinations, timed writing assessments, and any other evaluative exercise designated as an exam by the instructor, unless the instructor has expressly authorized specific AI use in writing as part of an approved deviation from this policy. An instructor’s general deviation authorization for a course does not extend to examinations unless the instructor explicitly so states.
4. No Treatment of Agentic or Embedded AI
Problem: The policy was drafted with the prompt-and-response paradigm in mind (a student asks an AI tool to do X). It does not address increasingly common scenarios:
AI embedded in research platforms that automatically summarizes a case when the student opens it
AI “copilot” features that suggest sentence completions as a student types
AI that a student uses to evaluate their own argument (”Is my reasoning here sound?”) — which isn’t brainstorming, outlining, drafting, revising, editing, or translating, so may not be covered
Suggested language:
The prohibitions in this policy apply regardless of whether AI assistance is actively solicited or is offered passively through embedded product features. Students who use research platforms, word processors, or other tools with embedded AI features must disable those features for work submitted for credit, or must confirm in writing to the instructor that such features were not used. Querying an AI tool about the quality, persuasiveness, or legal sufficiency of any portion of submitted work constitutes prohibited use under this policy, whether or not the query resulted in direct revision.
5. No Treatment of Pre-Submission vs. Process Use
Problem: The policy focuses entirely on what is submitted. But a student could use AI extensively throughout their process and then rewrite the output in their own words before submission. The policy does not address this, creating a significant loophole and, more importantly, allowing students to use AI in a way that defeats the policy’s stated pedagogical purpose without technically violating its text.
Suggested language:
The prohibitions in this policy apply to the process of preparing work for credit, not merely to the final submitted product. A student who uses Generative AI at any prohibited stage of the work — including conceptualization, outlining, drafting, revising, or editing — and then rewrites or paraphrases the AI’s output before submission has violated this policy. The measure of compliance is the process, not the product.
III. Enforcement and Procedural Deficiencies
6. The Presumption Mechanism Is Underdeveloped
Problem: The policy says nonexistent citations “raise a presumption of prohibited AI use,” but provides no procedural framework: Who investigates? What is the standard of proof? How is the presumption rebutted? Without this, the presumption is legally meaningless and practically unenforceable. It also creates the odd implication that only citation errors trigger the presumption — what about AI-characteristic writing patterns, hallucinated legal standards, or fabricated quotations?
Suggested language:
Enforcement Presumptions. The following circumstances raise a rebuttable presumption that a student has engaged in prohibited AI use, which may be considered by the Academic Integrity Officer in any proceeding under the Honor Code: (a) Citation to a case, statute, or secondary source that does not exist or that is materially misrepresented; (b) A quotation attributed to a court, statute, or author that does not appear in the cited source; (c) A legal standard or rule stated with specificity that does not correspond to current doctrine in the relevant jurisdiction. A student may rebut any such presumption by producing contemporaneous evidence of their research and writing process (e.g., research notes, drafts, search histories). The burden of production shifts to the student upon the triggering of a presumption; the burden of persuasion remains with the institution.
7. The Disclosure Architecture Is Insufficient
Problem: Disclosure is only required when an instructor authorizes a deviation from the default rule. But disclosure serves important functions even under the default rule: it creates accountability, supports enforcement, and habituates students to the professional norms of AI transparency that bar rules will increasingly require. The policy misses an opportunity here.
Suggested language:
Disclosure. Any student who uses Generative AI for the permitted purpose of source identification must disclose that use in their submitted work as provided above. Where an instructor has authorized deviation from the default rule, students must attach to any submitted work a disclosure statement identifying: (a) the AI tool(s) used; (b) the specific tasks for which AI was used; and (c) the student’s description of how they evaluated and verified the AI’s work product. Instructors may prescribe a specific disclosure format. Failure to disclose authorized AI use when disclosure is required constitutes a separate violation of this policy.
IV. Policy Architecture and Drafting Issues
8. The Purpose Statement Promises More Than the Rule Delivers
Problem: The Purpose section articulates two goals: (1) cultivating cognitive skills essential to excellent lawyering, and (2) fairness and administrability. But the two purposes are in tension, and the policy does not acknowledge this. A rigidly uniform rule may be highly administrable but pedagogically blunt — the policy’s own instructor-deviation mechanism implicitly concedes this. More importantly, the second purpose (”fairness”) is invoked without explanation. Fairness to whom, in what sense? Students who believe AI use gives peers an advantage? Students who want access to tools they’ll use professionally? The ambiguity weakens the policy’s normative foundation.
Suggested language:
Purpose. This policy rests on the judgment that the most valuable legal education is one that builds genuine cognitive capability — the ability to read, reason, argue, and write independently. AI tools, however powerful, do not yet substitute for these capabilities; they require them. A lawyer who cannot independently analyze a legal problem, construct an argument, or produce clear prose is not well-served by fluency with tools that perform those functions, because tool fluency without underlying competence produces work the lawyer cannot evaluate or stand behind.
This policy also serves fairness among students: uniform application of clear rules ensures that academic evaluation reflects each student’s own work and development. The policy acknowledges, however, that fairness also requires giving students experience with the tools of modern practice, which is why instructor-authorized deviation is available for courses intentionally designed to develop AI fluency.
9. The Instructor Deviation Mechanism Lacks Guardrails
Problem: The policy grants instructors effectively unlimited discretion to deviate from the default rule, requiring only that deviations be in writing with appropriate notice and require disclosure. But there is no review mechanism, no minimum standards for what constitutes a pedagogically appropriate deviation, and no coordination mechanism to prevent students from experiencing inconsistent rules across courses in ways that undermine the policy’s fairness purpose.
Suggested language:
Instructor Deviations. An instructor may deviate from this default rule for a course or assignment upon determining that a distinct AI use policy is pedagogically appropriate. Deviations must: (a) be submitted to the Dean of Students or designated Academic Affairs officer at or before the start of the relevant course or no fewer than two weeks before the relevant assignment; (b) be communicated to students in writing in the course syllabus or in a written assignment instruction; (c) specify with particularity which AI tools are permitted, for which tasks, and subject to what conditions; and (d) require students to disclose AI use as provided in this policy. Deviations become part of the official course record. The law school will publish a summary of approved deviations each semester so that students can understand the range of AI policies they may encounter across their curriculum.
V. One Additional Drafting Note: The Rule Provision
The Rule section largely restates the Purpose section, which is appropriate for a policy document, but the drafting of the prohibition is worth tightening. “AI use is prohibited for any use for any purpose” is redundant on its face (”any use ... for any use”). More substantively, the prohibition should cross-reference the definitions and carve-outs rather than standing as a freestanding absolute, since a reader encountering the Rule section first will not know about the permitted research use.
Suggested language:
Rule. Except as expressly permitted below and except where an instructor has authorized a deviation in accordance with this policy, the use of Generative AI (as defined herein) is prohibited in connection with any work submitted for academic credit at Berkeley Law. This prohibition covers all stages of the preparation of such work, including but not limited to conceptualizing, outlining, drafting, revising, editing, and translating. Generative AI use is categorically prohibited in all examination situations except where an instructor has expressly authorized specific use in writing. Students may not upload course materials — including assignments, readings, slides, class recordings, or other course content — to any Generative AI system, regardless of the purpose of such upload.
Summary of Deficiencies
# Issue Severity 1 “AI” undefined High — administrability failure 2 Research carve-out vague High — creates loophole and false assurance 3 Exam provision underspecified Medium — ambiguity in most high-stakes context 4 Embedded/agentic AI not addressed Medium — increasingly common scenarios 5 Process vs. product distinction absent High — defeats core pedagogical purpose 6 Presumption mechanism underdeveloped Medium — unenforceable as written 7 Disclosure architecture insufficient Medium — missed professional norm-building 8 Purpose statement internally inconsistent Low-Medium — weakens normative foundation 9 Deviation mechanism lacks guardrails Medium — fairness and coordination concerns
The policy is thoughtful in its instincts — it rightly identifies that the point is to protect the process of legal reasoning, not merely to police the final product. The revisions above are designed to honor that instinct while building the definitional precision, procedural structure, and enforcement architecture that a serious institutional policy requires.
This is an EMINENT LAW faculty which emanated this ill-formed slop in the first place?? What on earth does that say about the worth of a degree there.
Disagree.
Keep it simple, and stand on ethics.
It is wrong (always) to use an LLM for any portion of the writing process one engages in, particularly as a student.
Your usage of an LLM to analyze/edit their document is comically insulting.