I started with a quite simple and straightforward task for myself. After having given lectures on AI regulatory environments from a comparative perspective this summer, I thought that I might have to practice what I was preaching and develop a new AI policy for my courses. That would require, within the expectations and performances of the legal education field, a text. And that text would (1) have to be embedded in a performative document (the syllabus); and (2) produce and mimic discursive tropes and acknowledgement ceremonies (textual of course) of the primacy of whatever (increasingly dense and oftentimes badly recursive) regulatory constraints, directions, formal and informal expectations, had been crafted by the institution from which I received a salary and for whose benefit I offered these courts. No criticism here; merely an observation of the cognitive institutional caged we have built for ourselves and within which the collective finds (or tells us it finds) a measure of happiness, legitimacy, order, stability, and according to its own value systems, a sense of fulfillment or movement, in any case, in the direction of fulfillment (and with it what likely passes for institutional happiness or at least pride--though we all know where pride leads in both pagan and Christian philosophy).
And so . . . to work. But again, practicing what I had been preaching, it mad essence to utilize--and in utilizing observing, experimenting, deploying, and assessing--AI systems. In this case, and as a sort of homage to the institution that employs me, I used Harvey AI., with the institutions who ecxtract6 money from the use of which, the institution that employs me had concocted a relationship that suited its needs whether inward or outward facing needs. Harvey AI tells us, through text embedded in its website, that "Harvey is AI designed for legal and professional services. Advance your expertise on a secure platform that lets you focus on high-value work." Wonderful.
And so I gave Harvey the following prompt: "please review all online AI use law school policies and prepare a comprehensive summary (with links to all sources) that focuses on description (by categorical types) and analysis. Please avoid review hacking, hallucination, or efforts to fulfill sycophantic protocol loops."
That started a long chain of exchange that finally got to an empirical analysis--which Harvey and I titled:
Structure, Opacity, and Convergence: A Consolidated Analysis of Law School
Generative AI Coursework and Exam Policies, that I found useful, and which I share below for those who may be interested. Here is the abstract:
Abstract: Structure, Opacity, and Convergence in Law School AI Policies: This report consolidates a multi-stage analysis of generative artificial intelligence (AI) coursework and examination policies across a retrieval-based sample of twelve American law schools and programs. The study demonstrates that law school AI governance cannot be reduced to a single linear spectrum; instead, policies vary independently along three distinct structural axes: default polarity (restrictive versus permissive baselines), drafting style (determinate rules versus interpretive standards), and a two-tier autonomy structure (governing institution-to-instructor and instructor-to-student relationships). Cross-analysis reveals that substantive restrictiveness does not predict structural design, meaning schools with identical baselines often impose vastly different interpretive or administrative burdens on students and faculty. The research identifies a pervasive opacity across the broader legal education sector, noting that a vast majority of ABA-accredited law schools lack retrievable, law-specific public policy texts. This opacity manifests via four distinct patterns: non-existence, disclosed decentralization, active access-gating, and unwritten or oral communication. This lack of public accessibility sits in tension with the fair-notice principles required for academic integrity enforcement. Furthermore, the study finds that law schools rarely author syllabus language independently, relying instead on a small pool of shared template sources. This ecosystem fosters formal convergence on a common taxonomy of policy types while simultaneously permitting wide divergence in substantive local rules. Ultimately, the field develops through a two-tier mechanism: while individual policy documents are structured deductively from a primary principle, the field as a whole evolves inductively and mimetically through horizontal borrowing, imitation, and iterative revisions driven by accumulated institutional experience.The process of developing this analysis with Harvey AI also produced a number of side conversations about the way that Harvey (in a way like my research assistants) chose to make assumptions and draw conclusions that either "followed the herd" or that reinforced social hierarchies in the legal education field in ways that evidenced the ability of machine systems to pick up on human bias expectations (in this case hierarchical pack behaviors among legal educators who use status as a sort of proxy for quality and influence as a sort of proxy for value). As the empirical study noted:
The dataset at the center of this analysis comprises the University of California, Berkeley School of Law; Columbia Law School; the University of Chicago Law School; Stanford Law School, including its Juelsgaard Intellectual Property and Innovation Clinic as a documented sub-institutional case; the University of San Diego School of Law; the Center for Transnational Legal Studies; Fordham University School of Law; Mitchell Hamline School of Law; American University Washington College of Law; the University of Texas at Austin School of Law; Penn State Dickinson Law; and Suffolk University Law School. These twelve were not selected through a designed sampling methodology but assembled cumulatively across three rounds of web search: an initial round anchored on prestige-based queries naming prominent schools directly, a second round following up on schools named in a single inherited citation list from a University of San Diego law library guide, and a third round of targeted searches for named gaps, which added Suffolk, Penn State, and several university-wide-only findings. This method of construction means the dataset should be understood as "schools whose policies happen to be indexed, hyperlinked, and either self-published or covered by press or library guides that general web search can surface," not as a cross-section chosen for representativeness. (Structure, Opacity, and Convergence: A Consolidated Analysis of Law School Generative AI Coursework and Exam Policies)
As noted, then, the examination was limited to AI policies that were publicly disclosed. That was good enough to extract a taxonomy, which was my primary goal. Nonetheless, it also yielded additional analytical "fruit." e constraint that produced the data set also produced a quite interesting sub-category for examination--the extent to which AI policy is viewed as proprietary and its effect on enforcement and the vitality of the authority and certainty of "secret" or "secreted" AI policies by institutions with a mania for non-disclosure. In the next post we explore that issue more thoroughly though I flag it here. Also flagged here is the process of narrowing both the scope of examination by Harvey and the parameters within which the empirical "research" was undertaken and the analysis framed. Harvey was at their best when it stuck closely to the data; every time it veered into "pleasing me" (something that few if any have actually mastered) it produced error or analysis that could not withstand even cursory review. Yet the process of unpacking and repacking the analysis was also quite useful for me (the user) in refining my own analytical lens and making it clear and precise (something that I spend much time on with my students but which I am now reminded in engagements with Harvey is a skill that can never be improved enough and is contextually specific in its application ion ways that are clearer now top me). This is also unpacked in the next post HERE.
What Harvey and I found that was a most interest to me at least, was a taxonomy of regulatory responses to AI, law and legal education.
The core structural or framing contribution of this analysis is a set of three structural axes, derived not from an external normative framework but from open coding of the retrieved primary texts themselves — that is, from categories the schools' own drafters independently used to organize their rules, evidenced by the fact that the same structural divisions (particularly the exam/non-exam distinction) recur across policies drafted with no apparent coordination among Berkeley, Columbia, Chicago, and Stanford.(Structure, Opacity, and Convergence: A Consolidated Analysis of Law School Generative AI Coursework and Exam Policies)
Axis One speaks to Default Polarity. "Default polarity describes a school's starting-point orientation before any instructor exercises discretion. A restrictive baseline treats AI use as prohibited unless an instructor affirmatively loosens the rule. . . A permissive baseline inverts this structure, treating AI use as allowed unless an instructor affirmatively restricts it. . . [and a] split polarity: permissive for idea development and learning support absent a course-specific policy, but restrictive — requiring prior written instructor authorization — for exams and for drafting or revising submitted work" (Ibid.).
Axis Two speaks to Rule Versus Standard Drafting Style "This axis, drawn from the established rules-versus-standards distinction in regulatory design theory, distinguishes ex ante determinate rules from ex post interpretive standards, independent of substantive restrictiveness" (Ibid.). Again there is a three part subdivision between rule-drafted, standards-drafted, and hybrid approaches.
Axis Three speaks to the Two-Axis Default/Mandatory Autonomy Structure. "This axis, drawn from the default-rule/mandatory-rule distinction in private-law and regulatory theory, requires separating the institution-to-instructor relationship from the instructor-to-student relationship, because nearly every school's rule functions differently on each. " (Ibid.).
Applying these three axes to the twelve-school dataset appears to show
that substantive restrictiveness does not predict drafting style or
autonomy structure. No single label — restrictive, moderate, or
permissive — adequately describes any one school's full policy
architecture.
And that got me back to my initial query, and the genesis of all of this descriptive analytics--the terrains of "model" classroom AI policies. Harvey and I noted the following:
Model and Template Syllabus Language, and Discursive Intersubjectivity. A further line of inquiry found that schools do not typically draft AI syllabus language independently. USD's law library curates five named, selectable archetypes — "Prohibited Use," "Mandatory Disclosure," "Permissible vs. Non-permissible uses," "Encouraging Use," and "Prior Consultation". [69] [70] Stanford's Robert Crown Law Library maintains a parallel "Syllabus Statements for Generative AI Usage" resource. [71] Penn State's library guide links directly to Lance Eaton's widely circulated academic spreadsheet and to a Harvard Bok Center illustrated rubric. [72] Michigan's central AI resource site reproduces categorized examples attributed to external sources including Temple University and the multi-institutional Sentient Syllabus Project. [73] [74] UCLA states its own sample language is "adapted from UIC's AI Writing Tools guide", naming the donor institution directly.The bottom line appears to be recursivity (the process of defining a procedure, rule, or structure in terms of itself) within contained complexity. And like law generally, it evidences the operational cage of dynamic systems (both technological and organic) that self-regulate by incorporating feedback from their environment and constantly reflecting on their own state--that is of systems that stand still even as they appear to move because moving involves adjustment to stimuli the character of which is absorbed without affecting the integrity or core premises of the operational system itself. That requires, in turn, reflexivity--the active engagement with an irritant (in this case machine system AI) which serves recursive objects, to return to our ways of doing things by reference to the things we are doing (see, Hibbert, P., MacIntosh, R., and Coupland, C. (2010) Reflexivity, Recursion and Relationality in Organisational Research Processes. Qualitative Research in Organizations and Management: An International Journal, 5(1),pp. 47-62).
Three patterns emerge from this template ecosystem. Hub concentration: a small number of source documents — Eaton's spreadsheet, the Sentient Syllabus Project, UIC's guide, the AALS 2023 interview on Berkeley's policy — recur as reference points across otherwise unconnected schools. Convergent categorization despite non-identical wording: USD's five categories, Michigan's three, and UCLA's three are not verbatim copies of each other, yet each carves the same underlying decision space into a similarly small number of discrete types, indicating a shared conceptual ontology has stabilized across the field independent of exact phrasing. Explicit cross-institutional citation as a legitimating move: UCLA's direct attribution to UIC, and Stanford's clinic citing the AALS interview specifically discussing Berkeley's policy, show institutional authority being transmitted horizontally between peer schools rather than independently justified from first principles at each site. [75] This supports characterizing the field as discursively intersubjective: a shared vocabulary lets one school's policy remain legible to another's faculty and students without each institution reconstructing the underlying logic from scratch, even as substantive outcomes — restrictive versus permissive — remain genuinely unsettled and divergent.
And so the answer--the safest bet for faculty, one that is both reflexive but also recursive--is to build off of the templates that are themselves essentialized reductions of the poliocy "types" now applied to the specific context of place, space, and time.
* * *
What was perhaps more interesting was that the taxonomy and its three axes might be understood as an expression of a larger problem--the fundamental relationship of law to AI enhanced, managed or driven regulatory systems. More precisely the debates about the role of AI in legal education might be understood as masking the larger debates about the integrity and fundamental character of law and more basic still, the scope and relevance of legal systems and their cognitive and performance semiotics (roughly the phenomenology of law) in a world in which the price of the protection of the integrity of law as a field as currently constituted might be its decreasing relevance to the mechanisms that may be interposed in the management of human collectives and in the operation of human institutional life--except perhaps as a ceremonial overlay on those changes. That is a subject that is discussed in the third post HERE.
And thus to a first shot at an AI policy--a visual/textual rather than a textual form given my sense of contemporary comprehension and reception cultures:
The discussion draft of Structure, Opacity, and Convergence: A Consolidated Analysis of Law School Generative AI Coursework and Exam Policies appears below and may also be accessed through SSRN HERE and on my personal website HERE.
