Hello from the other side
I studied anthropology at UC Berkeley. The Department of Anthropology and the Law School sit on opposite sides of a central open space on the Berkeley campus, close enough that you could shout between them. This article is a friendly hello from an alumnus on one side of that square to colleagues on the other.
One of the things Berkeley anthropology is proud of is its role in dismantling a particular kind of scientific error: the reading of hidden moral and legal status from surface physical features. That error has a name and a history. Its author was the Italian physician Cesare Lombroso, and the history is one of the most embarrassing in modern science. In the 1870s, Lombroso proposed that criminals were a distinct anthropological type whose criminal nature could be read directly from physical features.[1] He measured thousands of skulls. He catalogued facial angles, ear shapes, forehead contours. He developed a list of what he called stigmata of degeneration: surface features that, he argued, reliably indicated the hidden legal-moral status of their possessor. You could identify the born criminal, he claimed, by looking carefully enough at their face.
The inference chain ran: observable surface features, to hidden inner nature, to legal-moral status. The features Lombroso catalogued were real enough as observable phenomena. What was false was the inference from surface to status. The features did not exclusively indicate criminality. They described perfectly well many innocent people. The more the system was applied, the more it fell disproportionately on groups who were already suspect for other reasons: the poor, immigrants, the racially marginalised. Generations of physical anthropologists spent their careers dismantling this apparatus.
The comparison I am about to draw has obvious limits. Lombroso's project was rooted in pseudoscience and racial ideology. Berkeley Law's AI policy stems from understandable pedagogical anxiety about assessment integrity. No one is accusing the law school of eugenics. The structural similarity in the inferential move, however, is exact: observable surface features of submitted text, to the inference of a hidden status (AI use), to legal-moral sanction (misconduct proceedings, exclusion from resitting). It is that inferential structure, and not the motivational story behind it, that makes the comparison analytically precise rather than merely rhetorical.
In the spring of 2026, Berkeley Law revised its AI policy. The new rule prohibits students from using AI for 'conceptualizing, outlining, drafting, revising, translating, or editing any work submitted for credit.' The rule has several enforcement mechanisms. The most revealing reads the presence of AI use from the surface features of submitted text. Across the square, the anthropologists notice.
What the rule actually says
The 2026 rule has five operative provisions, and their structure reveals the policy's implicit theory of what cognition is and what legal education is for. The first and most sweeping provision prohibits AI 'for aid in conceptualizing, outlining, drafting, revising, translating, or editing any work submitted for credit.' Six verbs, collectively exhaustive of the cognitive lifecycle of academic writing. The rule is not targeting a specific category of output. It is targeting an entire cognitive process.
The second provision bans AI 'for any use for any purpose in any exam situation.' The triple iteration of 'any' forecloses every possible exception. This is the rule's most defensible provision, because examination conditions are the closest analogue to supervised professional performance.
The third provision bars students from uploading course materials, including assignments, readings, slides, and class recordings, into AI systems.
The fourth provision is the rule's only exception: students may use AI to identify sources, specifically cases, statutes, or secondary materials.
The fifth provision establishes the enforcement mechanism: 'Citations to sources that do not exist will raise a presumption of prohibited AI use.'
Notice a logical problem running through the rule's basic structure. To use AI to find relevant sources, a student must first tell it what the paper is about. To do that, the student must have already thought about the argument, the legal framework, and the direction of inquiry. That is conceptualizing. The one permitted use presupposes the central prohibited activity. The exception swallows the prohibition.
As a matter of legal policy design, these five provisions implicate at least three well-established concerns in academic regulation. First, vagueness: a rule governing conduct whose consequences include potential expulsion and bar admission difficulties must give those subject to it fair notice of what compliance requires. Second, overbreadth: restrictions must be tailored to the interests they serve and must not sweep in legitimate educational activities alongside genuinely problematic ones. Third, arbitrary enforcement risk: where a rule's operative terms cannot be applied consistently, enforcement becomes discretionary, and discretionary enforcement in high-stakes contexts is a structural injustice, not merely an administrative inconvenience. Each of these concerns is examined in detail below.
The legal framework: three standards the rule must meet
Before examining each provision, it helps to have the relevant analytical standards clearly in view. Berkeley Law is not directly bound by the constitutional rules that govern state actors. But Berkeley Law is committed through their own policies and accreditation obligations to principles of procedural fairness that track, in most respects, the constitutional standards applicable to public institutions. Applying that framework here is not a category mistake. It is the appropriate vocabulary for evaluating whether the institution's rule meets the standards the institution has set for itself.
The vagueness standard requires that rules be stated with sufficient clarity to give fair notice of prohibited conduct and to constrain the discretion of those who enforce them. As the Supreme Court articulated in Papachristou v. City of Jacksonville (1972), a rule 'which does not give a person of ordinary intelligence a reasonable opportunity to know what is prohibited, so that he may act accordingly' fails basic rule-of-law requirements.[2] Adapted to academic regulation, the standard asks: can a student, reading the rule in good faith, determine with reasonable confidence whether specific conduct is prohibited? Where the answer is no, enforcement becomes arbitrary, and students who seek in good faith to comply are placed in a worse position than students who calibrate strategically to the institution's detection capabilities.
The overbreadth principle requires that restrictions be tailored to the interests they serve. In Broadrick v. Oklahoma (1973), the Supreme Court explained that where a regulation's overbreadth is 'substantial, judged in relation to the statute's plainly legitimate sweep,' it cannot stand without more justification.[3] Applied to academic policy, a rule that prohibits activities functionally indistinguishable from long-accepted educational practices, without a principled account of why those activities are now problematic, fails the tailoring requirement.
Rational basis review, the most deferential standard available, requires only that a rule be rationally related to a legitimate purpose. As the Supreme Court explained in FCC v. Beach Communications (1993), the connection cannot be 'entirely imaginary.'[4] Where a rule's enforcement mechanism systematically fails to identify the conduct it prohibits while reliably identifying innocent conduct instead, the rational relationship dissolves.
Procedural fairness requires that where institutional sanctions carry consequences beyond the immediate academic context, the evidentiary basis for those sanctions meet some basic standard of reliability. For law students, an academic misconduct finding is not merely a matter of grades. A finding of dishonesty can affect bar admission through character and fitness review in every United States jurisdiction.[5] The stakes of misconduct proceedings at a law school are therefore qualitatively different from those at most undergraduate institutions, and those higher stakes require a correspondingly higher evidentiary standard.
The three claims that are not argued
The rule's public defence makes three claims. First, that activities like brainstorming and drafting are essential cognitive skills. Second, that students will not develop these skills if AI helps them. Third, that 'AI should not be used for anything that students should learn to be able to do autonomously and by their own lights.' These are starting points, not arguments. When examined, each turns out to be weaker than it sounds.
The second claim, that AI prevents skill development, is stated as though self-evident. It is not. The philosophers Andy Clark and David Chalmers demonstrated in their foundational work on the extended mind that human cognitive capacities have always been partly constituted by the tools through which we operate, making the line between 'my thinking' and 'my tools' contribution' philosophically unstable from the start.[6] Calculators changed mathematical practice without destroying mathematical ability. Word processors changed drafting without eliminating the capacity to write. Statistical software changed empirical research without eliminating analytical judgment. The claim that AI necessarily prevents the development of legal reasoning, rather than changing its character, requires evidence. None is cited.
A thoughtful proponent of the rule might respond that AI represents a categorically different kind of assistance because it can generate near-complete work product, making it qualitatively distinct from study aids, peer feedback, or professor guidance. This is a fair point and deserves a direct response. It is true that a student could ask an AI to write a complete legal essay and submit it unchanged. That is a genuine problem. But the rule does not address that problem specifically. It addresses all AI involvement in the cognitive process of writing, from an initial brainstorm through a grammar check. Under an intermediate scrutiny-style analysis, the rule fails the narrow tailoring requirement: it is not crafted to address the specific harm of work-product substitution but instead imposes a categorical prohibition across a continuous spectrum of uses, most of which do not involve the submission of unreviewed AI output.
The third claim is the most revealing. 'Students should learn to do these things autonomously and by their own lights.' This is an entire theory of how the mind works, compressed into one subordinate clause. But take it seriously and notice what follows. Discussing your paper with a professor is not autonomous. Getting feedback on your outline from a classmate is not by your own lights. Using a research assistant to gather sources is not independent. None of these is banned. The autonomy standard is applied to AI and AI alone, with no principled account of why AI is categorically different from all the other forms of cognitive assistance that legal education has always involved and always will involve.
A rule no one can actually follow
As a matter of fair notice, the rule must give a student reading it in good faith a reasonable opportunity to determine whether specific conduct is prohibited. It does not meet this standard, and the failure is not a minor drafting imprecision. It is a structural defect that makes consistent enforcement impossible and selective enforcement inevitable.
What does 'conceptualizing' mean? Consider a concrete example from the current law school curriculum. A second-year student is writing a note on administrative law following Loper Bright Enterprises v. Raimondo (2024), the Supreme Court decision that overruled the long-standing Chevron deference doctrine.[7] She asks an AI to explain the practical difference between the old Chevron framework and the new standard. Has she engaged in prohibited conceptualizing? She has used AI to build her understanding of the doctrinal landscape before forming her own argument. The rule provides no workable standard for answering this question, and the answer determines whether she faces an academic misconduct proceeding.
'Translating' raises concerns that go beyond vagueness into questions of equal treatment. Under Title VI of the Civil Rights Act and its institutional analogues in accreditation standards, universities must not apply requirements in ways that impose discriminatory burdens on students on the basis of national origin. A rule that prohibits using AI to translate legal concepts from a student's primary language into English legal prose imposes a specific cognitive burden on international students that does not apply to native English speakers, without any pedagogical justification for the differential burden. The rule appears not to have considered this implication, which is precisely the kind of unexamined consequence that narrow tailoring analysis is designed to surface.
'Editing' creates a direct conflict with the 2023 version of Berkeley Law's own policy, which explicitly permitted grammar correction. The 2026 rule prohibits editing without qualification and without acknowledging that anything has changed. Has grammar correction been silently prohibited? If not, the rule requires a workable distinction between permitted grammar correction and prohibited editing. No such distinction is provided. A Berkeley Law student asked to evaluate a statute for vagueness would identify exactly this problem within minutes of reading the text.
A rule that bans far too much
Even accepting the rule's stated purpose, protecting the development of genuine legal reasoning skills, the rule prohibits substantially more than is necessary to achieve it. This is textbook overbreadth, and the standard remedy is to narrow the rule to its legitimate core rather than to enforce a broad prohibition that sweeps in protected educational activity alongside genuinely problematic conduct.
Consider the prohibition on uploading course materials to AI systems. A student who uploads a difficult judicial opinion and asks AI to explain a hard passage is engaged in AI-assisted reading comprehension. This is functionally indistinguishable from consulting a commercial study aid, asking a teaching assistant to explain a passage, or using office hours to clarify a doctrinal difficulty. The rule bans the AI-assisted version while leaving all other forms of the same activity untouched, without any principled account of what AI does to reading comprehension that makes it categorically different.[8]
The ban on revising with AI is equally overbroad. Revision is the stage of legal writing where judgment is most actively exercised. A student who writes a draft, asks AI to identify weaknesses in the argument, evaluates those identified weaknesses critically, decides which are real and which are not, and rewrites accordingly is doing exactly the kind of analytical reasoning that legal education claims to develop. Treating this as equivalent to submitting an AI-written document, which is what the rule requires, is an analytical error. The activities are not equivalent. The rule's failure to distinguish them undermines the pedagogical purpose it purports to serve.
Here is the distinction the rule ignores but that any coherent account of AI and legal education must make. There is AI use that bypasses intellectual formation: the student receives a finished product and submits it with minimal engagement. And there is AI use that exercises and extends intellectual formation: the student uses AI as a thinking partner, evaluates its output critically, identifies its errors and limitations, and develops judgment through that engagement. A rule crafted to protect the development of legal judgment would distinguish between these two categories with precision. This rule prohibits both with equal severity. That is a reliable sign that it is not, in fact, designed to do what it claims.
Lombroso in the examination hall
From a due process perspective, the enforcement apparatus is where Berkeley Law's AI policy reproduces, with instructive precision, the epistemological error that physical anthropology spent a century and a half learning to recognise and reject.
When Berkeley Law tries to detect AI use, it relies on two instruments. The first is algorithmic detection software, tools such as Turnitin's AI detector and GPTZero, which promise to identify AI-generated text with sufficient reliability for institutional use. The second is what researchers have called stylistic physiognomy: the informal identification, by academic staff, of prose textures that 'feel' machine-generated. Certain surface features have attained a quasi-official status as AI indicators: the em dash used with uncharacteristic frequency, repetitive transitional phrases, excessive summarising at the end of paragraphs, a generic enthusiasm applied uniformly across heterogeneous content, a smoothness of surface that sits oddly against thin substantive engagement.
Recognise the structure of this inference: observable surface features of text, to the inference of a hidden status (AI use), to a legal-moral conclusion (deception, misconduct, exclusion from resitting). This is Lombroso's inference chain, reproduced in a new domain. The surface features being catalogued are real. The inference from those features to hidden deception is the error.
The surface features that detection systems treat as AI indicators describe perfectly well the prose of weak human writers, of second-language students exercising deliberate caution, and of students trained into formulaic academic registers by years of institutional feedback demanding exactly the kind of structured, hedged, transitional prose that AI also produces. The correlations are not stable: models change rapidly, and what read as AI-typical in 2024 is unremarkable by 2026. The inferential move from 'this text has these features' to 'this student used AI' to 'this student committed misconduct' involves two additional leaps, each requiring independent justification, that stylistic observation cannot supply.
The algorithmic detection tools are worse. Research published in peer-reviewed journals has documented that these tools flag essays written by non-native English speakers at substantially higher rates than essays written by native speakers, with false positive rates exceeding fifty percent for international student populations in some studies.[9] Under principles analogous to those the Supreme Court articulated in Griggs v. Duke Power (1971) regarding employment tests with disparate impact, instruments used to make high-stakes decisions must be shown to be valid for the purpose for which they are used and must not produce discriminatory outcomes without adequate justification.[10] The AI detection apparatus used by Berkeley Law meets neither requirement.
The stakes of getting this wrong are particularly high in a law school context. A finding of academic misconduct at a law school carries consequences that extend well beyond course grades. Every United States jurisdiction requires bar applicants to demonstrate good character and fitness, and academic misconduct findings are specifically interrogated in that process. A student who receives a misconduct finding based on the output of a detection tool with a fifty percent false positive rate for international students may face consequences affecting the entire trajectory of their legal career. The evidentiary standard applied to decisions with these stakes should be considerably higher than 'this text has surface features that a piece of software associated with AI.'
Lombroso's stigmata of degeneration fell disproportionately on people who were already marginalised: the poor, immigrants, the racially marginalised. The current detection apparatus falls disproportionately on students who are already at risk: international students flagged at higher rates by detection tools, students trained in formulaic institutional styles, students whose native linguistic patterns happen to resemble what AI produces. The system does not merely fail to detect what it targets. It reliably harms the wrong people while the students it most needs to catch escape unnoticed.
In the Department of Anthropology, across the square, generations of anthropologists built their scholarly identities on the critique of exactly this inferential structure.[11]
The trap that catches the wrong people
The rule's citation presumption is a distinct problem that raises concerns at the level of basic evidentiary logic. 'Citations to sources that do not exist will raise a presumption of prohibited AI use.' In legal terms, a rebuttable presumption is a device that shifts the burden of proof: once the triggering fact is established, the presumed conclusion is treated as true unless the party against whom it operates can rebut it. For such a presumption to be legitimate, there must be a rational and, in stronger formulations, a necessary or highly probable connection between the triggering fact and the presumed conclusion.
The presumption fails this requirement. Citations to non-existent sources arise from many causes entirely independent of AI use: misremembered case names under exam pressure, confusion between two similarly named cases, reliance on a secondary source that itself misrepresented a primary, transcription errors, and citation format errors that make real materials appear non-existent. None of these causes has any connection to AI use, and all are independently common in student work. The triggering fact is ambiguous, and the presumption flows from that ambiguous fact to a specific conclusion about cognitive process without any showing that the two are rationally connected.
In evidentiary terms, the presumption is simultaneously underinclusive and overinclusive. It is underinclusive because the most extensive AI users, those who use AI carefully throughout every prohibited stage and verify their citations by hand, will never trigger it. It is overinclusive because every student who makes a citation error through ordinary human fallibility, without any AI involvement, may be required to prove a negative about their own internal thought processes. A presumption that fails simultaneously on both dimensions implicates serious questions of rational basis under any evidentiary standard.
There is also a professional responsibility dimension that the rule's authors appear not to have considered. Berkeley Law trains students in the rules of professional conduct, which impose on lawyers an obligation of competence that includes citation accuracy. A rule that creates a presumption of serious misconduct on the basis of citation error, without discriminating between error attributable to AI and error attributable to ordinary fallibility, risks treating as academic fraud exactly the kind of mistake that professional responsibility instruction is designed to prevent through competence development rather than punishment. The rule conflates the problem it targets, AI-assisted deception, with a different problem, citation error, in ways that Berkeley Law's own professional responsibility curriculum would not endorse.
The double bind: how uninhabitable rules cause harm
Up to this point, the analysis has been largely technical: the rule is vague, overbroad, its detection apparatus is epistemologically invalid, its presumption is backwards. But there is a human dimension that the technical critique does not fully capture.
Academic integrity rules are not just compliance mechanisms. They are supposed to sustain genuine norms: about honesty, about the relationship between intellectual effort and intellectual product, about what it means to actually know something. When those norms become practically uninhabitable, when the rules cannot be followed in good faith because the environment they assume no longer exists, something worse than non-compliance happens. The connection between the norm and the student's capacity to orient themselves by it is severed.
Students are caught in what the philosophical literature calls a double bind. They are simultaneously required to submit independent work and required to operate in an environment where independence, as the rule defines it, cannot be reliably enacted. AI assistance is not an exotic option that students choose to pursue. For many students, in many working environments, it is ambient: woven into the basic infrastructure of how academic work now happens. If they use it in the ways their environment has normalised, they risk catastrophic sanction. If they avoid it entirely, they face competitive disadvantage relative to peers who do not, and relative to the efficiency expectations that contemporary study practice has built in. There is no costless option.
The predictable consequence, and this is a structural outcome rather than a moral failing of individual students, is that ethical orientation gets replaced by strategic calibration. Students do not learn to think clearly about when AI use is appropriate and when it is not. They learn to manage the surface of their work so as not to trigger suspicion. The unit of attention shifts from the intellectual quality of the work to the stylistic plausibility of its claimed authorship. This is the precise inverse of what assessment is supposed to produce.
The ABA's accreditation standards require that law schools prepare students for 'effective, ethical, and responsible participation as members of the legal profession.'[12] A rule that produces strategic calibration rather than genuine ethical orientation undermines this requirement. Students who leave Berkeley Law having learned to manage surfaces rather than to think clearly about the appropriate use of new tools are not being prepared for effective and responsible professional participation. They are being trained in a form of compliance performance that does not reflect the actual demands of modern legal practice.
We have always thought with tools
The rule's demand that students work 'autonomously and by their own lights' rests on a picture of cognition that sounds natural but falls apart under scrutiny. The picture holds that real thinking happens inside the individual, independently of external aids, and that tools are supplements you might use after the real thinking is done.
This view has been argued before, and it has been wrong in the same way every time. Plato's Socrates argued in the Phaedrus that writing would destroy memory: by recording things externally, people would stop holding them internally, producing the appearance of knowledge without its substance. The concern was not ridiculous. Writing does change how memory works. But the catastrophe did not arrive. Writing made possible forms of reasoning that no purely oral culture could achieve. The same pattern has repeated with calculators, word processors, and internet search. Each time, the new technology changed the character of the cognitive practice it touched. Each time, the doom scenario failed to materialise. And each time, what was being protected was not pure cognition, because there is no pure cognition, but a specific set of institutional arrangements for recognising and credentialling competence.
The anthropologist Edwin Hutchins studied navigation teams on naval vessels in the 1990s and showed something fundamental: expert performance in complex professional domains characteristically involves the distribution of cognitive work across people, instruments, and representational systems.[13] None of these elements is merely external to the cognitive process. They are constitutive parts of it. Legal cognition has always been distributed in exactly this sense. Case reporters extended legal memory beyond what any individual could hold. Digest systems made doctrinal synthesis tractable. Shepard's Citations, Westlaw, LexisNexis: each became constitutive of what legal reasoning is, not merely a supplement to reasoning that happened independently.
When Westlaw and LexisNexis became commercially available in the late 1970s, legal educators expressed concern that electronic research would produce lawyers incapable of the deep engagement with legal materials that professional formation required.[14] Law schools adapted. They learned to teach students how to use these tools well, how to exercise judgment in navigating electronic research, and how to maintain critical engagement with materials surfaced by algorithm rather than discovered through manual search. The current rule represents a departure from that adaptive tradition without explanation of why this particular technology requires categorically different treatment.
The existing rules of academic integrity are not timeless moral principles. They are practical compressions of a specific material environment: one in which assistance could be individuated, identified, and excluded, and in which the distinction between independent work and assisted work was stable enough to guide action. That environment no longer exists. Assistance is now ambient and layered into the basic infrastructure of access itself. The rules are not merely outdated. They are ontologically misfitting: they presuppose a configuration of the world that no longer obtains and continue to claim authority over a domain whose actual structure they no longer track. Stating such rules more loudly and enforcing them more severely does not fix them. It produces more confusion, more injustice, and more strategic compliance.
The strongest case for the other side
It is worth engaging honestly with the strongest version of the case for a restrictive policy, because the case is not without substance.
Three concerns genuinely motivate the rule. The first is the signaling and credentialing function of the law degree. Legal employers and clients rely on the J.D. as a signal of certain baseline competences. If AI assistance makes it possible to complete legal education without developing those competences, the degree loses the function it serves, and everyone who relies on that signal, employers, clients, and ultimately the public, is worse off.
The second concern is the practical difficulty of scaling assessment alternatives. This article recommends oral examination, process portfolios, and supervised writing as replacements for surface-based assessment. These are more resource-intensive than written submission. At a law school with hundreds of students across dozens of courses, the transition from surface-based to inhabitation-based assessment carries real costs, and the institution's concerns about feasibility deserve acknowledgment.
The third concern is fairness between students. If some students use AI and others do not, those using AI may be systematically advantaged in the production of polished written work, shifting assessment curves in ways that disadvantage students who comply with the prohibition. This is a legitimate equity concern that any proposed alternative must address.
These concerns are real but are not adequate justification for this rule, and each is better addressed through a different response than the one the rule provides. The credentialing concern is better addressed by redesigning assessment to test the competences the credential is supposed to signal: a regime that tests inhabitation rather than surface would credential the right competences more reliably than the current one, not less. The feasibility concern is genuine but does not justify the specific design of the rule. There is no obvious reason that the response to scaling constraints should be a categorical prohibition on brainstorming rather than, for example, structured AI disclosure requirements or a shift to more frequent low-stakes oral components integrated into existing seminar formats. The fairness concern cuts in an unexpected direction: the current prohibition already creates serious fairness problems of its own, by falling asymmetrically on risk-averse students who over-comply and on international students who are disproportionately flagged by detection tools. A well-designed AI literacy framework with clear disclosure requirements and inhabitation-based assessment would be more equitable than the current regime, not less.
What a good rule would look like
A rule genuinely designed to develop legal reasoning skills would do four things this rule does not do.
It would define its operative terms with sufficient precision to give fair notice. Conceptualizing, translating, and editing need definitions that allow a student reading the rule in good faith to determine with reasonable confidence whether specific conduct is prohibited. Vagueness is not a drafting technicality. It is a fairness failure and, in this context, a serious due process concern.
It would distinguish between AI uses that replace judgment and AI uses that extend it. A structured disclosure and reflection requirement, under which students document what AI tool they used, at what stage, for what purpose, what they accepted, and what they rejected and why, would develop exactly the evaluative capacity that legal practice requires. It would also create a genuine paper trail of intellectual engagement, something the citation presumption entirely fails to produce.
It would replace stylistic detection with assessment that tests inhabitation. If the question is whether a student can actually perform legal reasoning, the answer is found by watching them perform it under conditions where it cannot be outsourced: oral examination, supervised writing, clinical work where the stakes of the argument are real. This is how the Administrative Court of Kassel in Germany, which examined an actual AI misconduct case through a proper evidentiary process, reached its findings: not by relying on detection software but on a documented mismatch between written and oral performance, supplemented by other contextual indicators. That is an epistemologically valid inference chain. Stylistic physiognomy is not.
It would prepare students honestly for the profession they are entering. Law firms are already training associates on AI-assisted legal research and drafting tools. Students who graduate having been taught to avoid these tools, under the impression that avoidance is a professional virtue, will enter a profession that has already moved on. A rule designed to develop genuine professional competence would treat AI literacy as a curricular goal, not a contamination to prevent.
Across the square
Let us come back to where we started: Lombroso, Berkeley anthropology, and what it means that these two institutions are looking at each other across a central open space at one of the best universities in the world.
Cesare Lombroso's criminal anthropology was not just wrong. It was wrong in a specific and instructive way. He had a genuine concern, the identification of criminal tendency, and he pursued it with the methodological tools available to him: measurement, observation, cataloguing of surface features. The features he identified were real. The inference from those features to hidden criminal nature was the error. And the consequences followed a predictable path. The system fell hardest on people who were already marginalised. It provided a procedural veneer for pre-existing institutional judgments about who was likely to cause trouble. It resisted correction because it was institutionally useful, not because it was epistemologically sound.
Berkeley Law's AI detection apparatus reproduces this structure with precision. The surface features being catalogued in submitted text are real: AI-generated prose does have characteristic tendencies, at least at any given moment in the technology's development. The inference from those features to hidden deception is the error. And the consequences follow the same groove: the system falls hardest on students who are already at risk, provides a procedural veneer for pre-existing institutional anxieties about academic honesty, and resists correction because it serves institutional interests in appearing to act vigorously.
The anthropologists across the square know what happens next. They have watched it happen before, in domain after domain where surface features were mistaken for reliable indicators of hidden moral or legal status. The system keeps generating findings with institutional confidence. The epistemological foundations continue to rot. The harm accumulates asymmetrically. And eventually, the whole apparatus is recognised for what it is: a performance of certainty conducted over an evidentiary void.
Berkeley Law's institutional identity is continuous with the tradition of critical inquiry that made Berkeley a great university: the refusal to accept surface features as reliable indicators of hidden moral or legal status, the insistence that inferences be validated against the evidence, the willingness to examine one's own methods as rigorously as one examines anyone else's. That tradition is alive in the law school's own classrooms. Students who learn to ask, of any legal instrument, whether its inference chain is valid, whether its enforcement mechanism actually targets the conduct it prohibits, whether its stated purpose is its real purpose: those students, if they turn that analysis on their own institution's AI rule, will have no difficulty finding the answers.
The good news is that Berkeley Law does not need to wait for recognition to arrive from outside. It already has everything it needs to diagnose its own policy. It teaches these tools every day. The question is whether it will use them. It will fail its mid-term exam if it doesn’t.
Notes
- ^ Cesare Lombroso, L'uomo delinquente (1876; 5th edition 1897). For critical analysis of Lombroso's methods and their consequences, see Mary Gibson, Born to Crime: Cesare Lombroso and the Origins of Biological Criminology (Praeger, 2002). The Berkeley anthropology department's institutional identity is directly continuous with the Boasian tradition that dismantled Lombrosian criminal anthropology in American scholarship.
- ^ Papachristou v. City of Jacksonville, 405 U.S. 156, 162 (1972); Grayned v. City of Rockford, 408 U.S. 104, 108 (1972). The Supreme Court held that a law is unconstitutionally vague when it 'does not give a person of ordinary intelligence a reasonable opportunity to know what is prohibited.' Private universities that voluntarily commit to procedural fairness in their own policies are bound by the same logic: a rule that cannot be followed in good faith cannot serve a legitimate governance function.
- ^ Broadrick v. Oklahoma, 413 U.S. 601, 612 (1973). The Court explained that a statute's overbreadth must be 'substantial, judged in relation to the statute's plainly legitimate sweep.' Applied to academic policy, a rule whose prohibitions sweep in activities functionally indistinguishable from long-accepted educational practices fails this standard.
- ^ FCC v. Beach Communications, Inc., 508 U.S. 307, 313 (1993). Even under rational basis review, the connection between a rule's provisions and its stated purpose cannot be 'entirely imaginary.' Where a rule's enforcement mechanism systematically fails to identify the prohibited conduct while reliably identifying innocent conduct, the rational relationship does not exist.
- ^ ABA Standard 510 requires that law schools maintain fair procedures for taking adverse action against students, including notice of charges and an opportunity to respond. For the bar admission dimension, see National Conference of Bar Examiners, Character and Fitness Review Guidelines (2024). Academic misconduct findings are specifically interrogated in character and fitness review in every United States jurisdiction.
- ^ Andy Clark and David Chalmers, 'The Extended Mind,' Analysis 58.1 (1998): 7-19. Clark and Chalmers argue that cognitive processes can extend beyond the individual when external tools reliably augment cognitive function in ways that would be credited as cognitive if performed internally. The philosophical implication is that the boundary between 'my thinking' and 'my tools' contribution' is not fixed and is not the right boundary around which to construct an academic integrity regime. The delusion of a clear boundary between ‘internal thought’ and ‘external aids’ is also a major argument in Living Value Theory (livingvaluetheory.org).
- ^ Loper Bright Enterprises v. Raimondo, 603 U.S. ___ (2024), overruling Chevron U.S.A., Inc. v. Natural Resources Defense Council, 467 U.S. 837 (1984). The shift from Chevron deference to the new framework for judicial review of agency interpretations of law is a central topic in current administrative law education and an ideal concrete example for vagueness analysis of the rule.
- ^ Weixin Liang et al., 'GPT Detectors Are Biased Against Non-Native English Writers,' Patterns 4.7 (July 2023). The study found that seven widely used AI text detectors classified over fifty percent of authentic essays written by non-native English speakers as AI-generated, while correctly classifying essays written by native English speakers at substantially higher rates. This disparate impact has not been addressed by the major detection tool providers.
- ^ Griggs v. Duke Power Co., 401 U.S. 424, 431 (1971). The Court held that employment selection instruments with a disparate impact on protected groups must be 'related to job performance' to survive legal challenge. The principle, that high-stakes decision instruments must be valid for the purpose for which they are used and must not produce discriminatory outcomes without adequate justification, applies by close analogy to academic assessment tools.
- ^ ABA Standard 303(a)(1) and (2) require that law school curricula include instruction in professional skills and that students be prepared for effective and responsible participation as members of the legal profession. A rule that produces strategic calibration rather than genuine professional judgment undermines this accreditation requirement.
- ^ Edwin Hutchins, Cognition in the Wild (MIT Press, 1995). Hutchins's study of naval navigation teams demonstrates that expert performance in complex professional domains is characteristically distributed across people, instruments, and representational systems, none of which functions as a merely external supplement to an otherwise self-sufficient cognitive agent.
- ^ Robert C. Berring, 'Full-Text Databases and Legal Research: Backing into the Future,' High Technology Law Journal 1 (1986): 27. Berring's early account of law school adaptation to electronic legal research illustrates the adaptive model that the current rule departs from.