 |
Lucille Ball and Harpo Marx as each other's mirrored reflections; a famous comedy skit that sets the tone for what follows |
Does an institution govern AI, or does its way of governing AI reveal the institution's cognitive cage that itself orders the reality within which the range of the possible is constrained? If an institution presumes that AI is a tool, is that a foundational (and unassailable) premise that is necessary to preserve a closed human institutional loop with the human not in the center but as occupying all cognitive space, surrounded of course by their tools; or are the discursive tropes of human centered-ness a way of avoiding giving any value or weight to the possibility that machine systems, programmed and trained by humans in the image of humans and reflecting the ideal composite human, as its programmers and trainers (and the owners of the enterprises under which these actions are undertaken, assessed, and managed), are themselves reflections of a striving for the creation, the animation, of virtual composite humanity in all of their glory? That is that to ignore the machine system as something other than a tool is to ignore the fundamental humanity of the machine system itself, and thus to de-center the human in the machine even as the machine is constantly developed to be more and more the manifestation of the ideal human(ity). Is the successful humanization of machine systems to be measured by the extent to which it mimics all of the characteristics of humanity from the beginning to now and beyond; is humanity ready to tolerate a relationship with itself where it is impossible to avoid oneself in all aspects of human individual and collective endeavor? Is a mechanized reflection of humanity too much for the community of humans, operating within their institutions, too much to bear? I think not but it appears many do not share that view, especially when organizing their cognitive premises toward the task of approaching the management of human-machine system interaction in institutional context.
Those are the questions at the heart of my reflection on the MIT Ad Hoc Committee Report on AI Use in Teaching, Learning, and Research Training (August 2026), the discussion draft of which follows below. And those are the questions, the approaches to which (there are no answers other than those founded on the belief in the verities of whatever first premises are embraced or whatever clusters of custom and tradition bind most tightly) produce the possibility of developing systems for the consideration of the "best" or "most suitable" approach to the management of human-machine system interactions in human (at least outwardly) centered institutions--like universities.
The Reflection considers the MIT AI Report in the context of a semiotics of path dependence and of the challenges of managing of institutionalized human-machine system interactions through rules. The reflection is entitled (originally as posted a longish title that I have now shortened to this): Philosophizing with a Hammer: The MIT Ad Hoc Committee Report on AI Use in Teaching, Learning, and Research Training (August 2026) and a Working Framework for Institutional Policy. Here s the abstractThis reflection reads the MIT Ad Hoc Committee Report on AI Use in Teaching, Learning, and Research Training (August 2026) as both an important institutional intervention and an object of philosophical inquiry, and then asks what it would mean to translate that inquiry into a usable analytical framework for institutional actors who must actually produce AI policy. The reflection proceeds in three movements. The first summarizes the Report and compares its principles and recommendations with my empirical study of twelve law-school AI policies (Structure, Opacity, and Convergence), my Five Machines analysis of machine-generated policy models, My Dinner With ChatGPT’s phenomenology of recursive human-machine inter-subjectivity, and my course-level AI policy. That comparison identifies convergence around contextual governance, transparency, verification, and augmentation, but also divergence over the structural axes of policy architecture, the legal character of human agency, and whether the machine can be treated only as an object of governance. The second movement, under the Nietzschean heading “Philosophising With a Hammer,” develops eight analytical queries informed by phenomenological, genealogical, semiotic, and path-dependence frameworks. These queries examine the Report’s foundational presuppositions, the cognitive cages within which its recommendations are generated, the gap between its transformative diagnosis and its incremental prescription, the genealogical production of its eight principles, the semiotic performance of institutional convergence, the temporal architecture of its review mechanisms, the absence of the machine as interlocutor, and the narcissistic closed loop through which the institution governs its own reflection rather than the phenomenon itself. Norbert Wiener’s God & Golem, Inc. — the work of a prior generation of MIT’s own intellectual tradition — serves as counterpoint: the creator–creature relation is already a significant game, and the refusal to play may be the most consequential move in a game already underway. The third movement, developed in a companion Operationalization Appendix, translates the philosophical critique into an analytical framework organized around four layers of institutional decision-making: foundational ontological commitments, second-order normative principles, structural policy architecture, and operational implementation. The appendix contends that institutions characteristically compress these layers — moving from unexamined foundational assumptions directly to operational recommendations — and that this compression is itself a mechanism of path dependence. It offers decision trees tracing the MIT Report’s implicit pathway alongside alternative pathways drawn from the Five Machines study and from The Soulful Machine’s ontological premises, together with a variable disentanglement chart mapping eleven critical decision points across the four layers. The inclusion of machine systems as potential participants in governance design — rather than merely as objects of governance — is identified as a radical but perhaps eventually necessary departure that institutional actors may initially set aside but ought not permanently foreclose. The question throughout is whether institutional self-preservation can become genuine encounter before the accumulated weight of inherited structures forecloses the foundational reorientation that the encounter with AI may require.
And here is the concluding thought:
Whatever the value of deeper and more nuanced analysis, in itself, and ultimately as an essential element in the understanding of the evolving relationship between humans and their reflections in and as machine systems; whatever the evolution of understanding of the appropriate starting points, the cognitive foundations, for developing human responses to its virtual self now incarnated within machine systems; none of this has much place in the necessities of institutions whose own cognitive pathways require action, or the appearance of action. That action, in turn, is both path dependent and ultimately conservative—the use of challenge as a space where its own traditions and practices may be better crystalized, and thus crystalized, preserved. That is precisely the most immediate challenge facing institutions, and their principal actors—administrators, faculty, students (and then indirectly the state and the emerging and perhaps different institutional cognitive starting points of employers and markets). That challenge and its response, then, requires the marginalization or at least the diminution and containment of contagion, of threat, of challenge, to the ends of preserving what the collective believed, understood, acted, and operationalized within its own structures. There may well be a time for understanding the threat—but it is not now. There may be a space where transformation may be necessary; but institutions tend to approach transformation only from within and as a consequence of crisis; and there is no crisis yet. Within these frameworks, the MIT Report is an excellent product of its type, and will likely be both praised and mimicked. It is thoroughly human, and it is vigorous in its defense of humanity (now understood in a more precise and precisely narrow way); ironically enough against the manifestation of a virtual collective humanity that is both dehumanized and reduced to a sort of instrumental servility that its own construction and programming belies. But that is a problem for tomorrow. For today—there must be policy; and that policy must preserve.The question, though, lingers--is today already that other day?
Pix credit here
The rough working draft follows below and may be accessed here.
 |
| created with ChatGPT |
For some of my readers the more important element of this reflection is the effort to translate the philosophical theoretical discussion into templates that can be used to actually develop AI policy in ways that comport with institutional cultures, aspirations, desires, objectives, etc. To those ends I have also prepared a roughed out draft of an Operationalization Appendix: From Philosophical Critique to Policy Architecture. It also follows below. Here is an Executive Summary for Non-Experts (SSRN HERE):
Executive Summary: Operationalization Appendix
What This Appendix Does
This appendix translates a philosophical critique of institutional AI policy — developed in the companion Reflections paper — into a practical analytical framework that institutional officials, policymakers, and governance committees can use when designing, evaluating, or revising AI policies for educational institutions. [1]
The analytical observation at the heart of the appendix is this: when institutions produce AI policy, they appear to be making decisions across at least four distinct layers simultaneously — but they characteristically treat only one of those layers as the real work. The other three layers are resolved silently, by default, through inherited institutional culture rather than deliberate choice. [2] [3] The result may be policy that is operationally sophisticated but foundationally unexamined — policy that addresses the immediate disruptions AI produces without engaging the deeper questions AI raises about what education is and what institutions are for. [4]
The Four Layers
Layer 1 — Foundational Assumptions. Before any policy conversation begins, the institution has already made (or inherited) basic assumptions about what AI is and what education is for. Is AI merely a tool, like a calculator or a search engine — powerful but ultimately an instrument under human direction? Or is it something more complex — a system whose outputs bear a relationship to meaning that the language of instrumentality does not fully capture? [5] [6] Is the institution's goal to preserve existing educational forms through the disruption AI brings, or to allow the encounter with AI to reshape educational practice? [4] These questions may sound abstract. They are not. The subsequent policy choices — what to permit, what to prohibit, how to enforce, whom to consult — tend to be shaped, and perhaps constrained, by how these questions are answered, whether or not anyone has answered them deliberately.
The MIT Ad Hoc Committee Report on AI Use in Teaching (August 2026) — a serious and careful institutional document produced by a committee of evident good faith — illustrates the pattern. [7] It treats AI as a tool throughout ("AI is only a tool, and students are responsible for all work they submit"). [8] It treats the residential university as the essential form of education. [9] It positions faculty as the agents of adaptation. [10] Each of these appears to operate as a foundational assumption rather than a finding produced by the committee's inquiry — and each may constrain what follows in ways the Report itself does not examine.
Layer 2 — Guiding Principles. Once foundational assumptions are in place, institutions articulate principles to guide policy. The MIT Report articulates eight: be humble, be bold, put humanity front and center, lean into learning, teach with intentionality, no one size fits all, augmentation not automation, and think beyond the classroom. [11] These are reasonable principles. But they are not free-standing discoveries — they follow from, and are constrained by, the foundational assumptions at Layer 1. "Put humanity front and center" follows from the assumption that education is fundamentally a human enterprise. [12] "Augmentation not automation" follows from the assumption that AI is a tool. [13] An institution with different foundational assumptions would produce different principles.
This is not a criticism. It is an observation with practical consequences: if one wants to understand why an institution's AI principles take the form they do, one might begin by examining the foundational assumptions they rest on. And if one wants to change or update the principles, it may be necessary to revisit those assumptions first.
Layer 3 — Policy Architecture. This is the structural design of the policy itself — the regulatory form it takes. Empirical research on AI policies across twelve law schools found that policies vary independently along at least three structural dimensions: [14]· Default stance: Does the policy start from a restrictive baseline (AI use prohibited unless authorized) or a permissive baseline (AI use permitted unless restricted)? [15] [16] [17]These three dimensions are independent of one another — knowing a school's position on one tells you almost nothing about its position on the others. [22] The practical consequence is that the policy design space is far larger than a simple "restrictive to permissive" spectrum suggests. A syllabus menu offering four options (Unrestricted, Limited, Required, Prohibited) captures only one dimension out of three — and leaves the other two to be resolved by default, which usually means by institutional inertia. [23]
· Drafting style: Does the policy use precise rules (specific prohibited/permitted activities) or flexible standards (general norms requiring judgment in application)? [18]
· Authority structure: How is decision-making distributed? Can individual instructors override institutional defaults? Can students override instructor rules through disclosure? [19] [20] [21]
Layer 4 — Operational Implementation. This is the layer that receives the most institutional attention: committees, designated AI leads, pilot funds, review cycles, professional development, template language, assessment modifications. [24] The MIT Report's recommendations at this layer are extensive and, within this layer, thoughtful. [25] But operational sophistication at Layer 4 cannot compensate for unexamined assumptions at Layer 1. An institution that establishes an AI committee, funds pilot programs, and creates a four-option syllabus menu has done real work — but it has done that work within a framework whose foundational assumptions, guiding principles, and structural architecture have been left largely unexamined.
The Pattern: What Happens When Layers Are Compressed
The pattern the appendix identifies is what might be called layer compression: institutions move from unexamined foundational assumptions (Layer 1) directly to operational implementation (Layer 4), skipping explicit engagement with guiding principles (Layer 2) and structural architecture (Layer 3). [26] [27] The result is a characteristic gap between diagnosis and prescription. The MIT Report diagnoses the challenge of AI as demanding "a broader, holistic reassessment of the nature, scope and purpose of higher education." [28] Its prescription is operational: committees, funds, reviews, templates. [29] The gap between these two — between the transformative diagnosis and the incremental prescription — is not necessarily a failure of nerve. It may be a structural consequence of compressing the analytical layers through which a transformative diagnosis might otherwise become a transformative response.
What the Appendix Provides
The appendix develops three analytical instruments:1. A textual analysis explaining the four layers, what each contains, and what happens when they are compressed. This is the conceptual foundation — the "why it matters" portion.
2. Decision trees and alignment charts showing:· How the four layers connect sequentially (foundational assumptions constrain principles, which constrain architecture, which constrain implementation) [30] [31]3. A variable disentanglement chart — a comprehensive table identifying eleven specific decision points across all four layers, and mapping each against the MIT Report, the law school empirical data, machine-generated policy models, and positions that are available but typically foreclosed.
· How different foundational assumptions produce different principle sets and different policy architectures [32] [33]
· The MIT Report's implicit decision pathway — where each foundational assumption enters and what alternatives it forecloses [34]
· Two alternative pathways drawn from the Five Machines study, showing what policy architecture looks like when different foundational assumptions are made explicit [35]
Implications for Institutional Governance
The appendix does not prescribe a particular AI policy. It does not argue that the MIT Report's approach is wrong or that any alternative is right. Its more modest contention is that institutional decision-makers might benefit from understanding the full range of what they are deciding — and that the most consequential choices in AI policy may be the ones that are never made deliberately.
An institution that begins its AI policy process by examining its foundational assumptions — What do we believe AI is? What are we trying to preserve or transform? Who has authority and why? — may be better positioned to understand why its policy takes the form it does, and therefore better able to revise when circumstances change. An institution that begins with a syllabus menu may arrive at an equally workable operational result, but it may also find itself constrained by commitments it never consciously made — carried along, as the path-dependence literature would suggest, by the accumulated weight of institutional inheritance rather than by deliberate choice.
The eleven variables in the disentanglement chart are the sites at which institutions make choices — or, more commonly, the sites at which choices are made for them. Making those sites visible is the contribution this appendix attempts. Whether visibility is sufficient to loosen what institutional culture has hardened is a question only institutional actors themselves can answer.
A Note on Machine Participation
One position the appendix identifies as available but rarely considered is the inclusion of machine systems as participants in — rather than merely objects of — governance design. [36] [37] [38] This is a radical departure from existing practice, and most institutions will reasonably set it aside when embracing conventional regulatory first principles. The conventional starting point — that governance is a human enterprise conducted over human concerns — has deep roots and defensible logic. But the appendix suggests that this question ought not be foreclosed permanently. The encounter with generative AI may eventually require precisely this rethinking, and an analytical framework that makes the question visible — even while acknowledging that most institutions are not yet prepared to answer it — performs a useful service. [39] [40]
 |
| Pix credit here |
Larry Catá Backer (白 轲)[1]
W. Richard and Mary Eshelman Faculty Scholar; Professor of Law and International Affairs
Penn State Dickinson Law, a unit of The Pennsylvania State University system | 239 Lewis Katz Building, University Park, PA 16802
Working Paper Draft
Abstract
This reflection reads the MIT Ad Hoc Committee Report on AI Use in Teaching, Learning, and Research Training (August 2026) as both an important institutional intervention and an object of philosophical inquiry. It proceeds in two movements. First, it summarizes the Report and compares its principles and recommendations with my empirical study of twelve law-school AI policies, my Five Machines analysis of machine-generated policy models, My Dinner With ChatGPT’s phenomenology of recursive human-machine inter-subjectivity, and my course-level AI policy. That comparison identifies convergence around contextual governance, transparency, verification, and augmentation, but also divergence over policy architecture, legal agency, and whether the machine can be treated only as an object of governance. Second, under the Nietzschean heading “Philosophizing With a Hammer,” the article develops eight queries concerning presupposition, cognitive cages, the gap between diagnosis and prescription, genealogical production, semiotic performance, temporality, the absent interlocutor, and the mirror of institutional self-recognition. A path-dependence framework supplies the connective tissue: the Report’s committee form, residential-humanist commitments, faculty authority, and operational vocabulary are treated as historically sedimented arrangements that shape what can be imagined as a response to AI. The central claim is that the Report’s pragmatism risks becoming institutional lock-in. Its treatment of AI as “only a tool” is not merely a cautious conclusion; it is a load-bearing premise required to preserve a closed human institutional loop. Against that loop, Norbert Wiener’s God & Golem, Inc. offers a counterpoint: the creator–creature relation is already a game involving learning, reproduction, coordination, responsibility, and uncertainty. The question is whether institutional self-preservation can become genuine encounter before the mirror hardens into a cage.
I Introduction
In Twilight of the Idols (Götzen-Dämmerung (1889)), Nietzsche announces that he will philosophise with a hammer: not to smash for the pleasure of destruction, but to tap the idols of a culture and listen for the hollow resonance beneath their apparent solidity. The hammer is an instrument of diagnosis. It asks whether what presents itself as natural is merely inherited; whether a cause is being confused with a consequence; whether an institution is defending an imaginary origin; and whether the freedom claimed by a system is only the name it gives to its own constraints. I take that figure here as an invitation to read, rather than merely summarize, the MIT Ad Hoc Committee Report on AI Use in Teaching, Learning, and Research Training (August 2026).
The Report arrives as a serious, careful, and recognizably institutional response to an upheaval that has already unsettled problem sets, examinations, undergraduate research, office hours, study groups, and the social contract of residential education. It is also a text that announces the need for a broader reassessment of the nature, scope, and purpose of higher education while translating that announcement into principles, menus, committees, and implementation instruments. That movement—from foundational unease to operational settlement—is the object of this essay. I read the Report against my prior empirical and philosophical work, not to place one text before a tribunal, but to see what becomes visible when different architectures of inquiry are made to confront one another.
The reflection therefore proceeds in two parts. The first offers a summary and comparative analysis of the MIT Report against Structure, Opacity, and Convergence, my empirical study of twelve law-school AI policies; Five Machines, my analysis of machine-generated model policies; My Dinner With ChatGPT, my phenomenological inquiry into human-machine recursive inter-subjectivity; and my course-level policy for Constitutional Law of Religion. It identifies convergences around contextual governance, transparency, verification, and augmentation, while tracing divergences concerning the structural axes of policy, the legal character of human agency, the machine as a possible interlocutor, and the difference between an inductive field and a committee-issued institutional program.
The second part, “Philosophising With a Hammer,” sustains the pressure. Its eight queries draw on phenomenology, genealogy, semiotics, and the literature of path dependence to ask what the Report presupposes before it begins, what its cognitive cages permit it to see, how its diagnosis becomes prescription, how its principles are genealogically produced, how institutional convergence performs a shared self-image, what temporality its recommendations assume, why the machine is absent as an interlocutor, and what the mirror returns when the institution governs only its representation of AI. The argument is not that the Report is careless; its seriousness is precisely what makes its structure consequential.
I argue that the Report exhibits institutional path dependence and a form of narcissistic self-address. It treats the machine as an instrument not because that conclusion has been fully examined and established, but because the existing architecture of the residential research university, faculty authority, committee knowledge, and human-centered governance requires the machine to occupy that position. The Report thus recognizes the need for foundational reorientation while translating that need into forms that preserve the institutional self that would have to be reoriented. Read beside The Soulful Machine and Norbert Wiener’s God & Golem, Inc., this is more than a policy choice. It is a question about whether the creator can play a significant game with its creature, and whether an institution can encounter a generative intelligence without converting the encounter into another scene of self-recognition. The hammer, finally, is directed at that scene—and at the possibility that its hollow sound may be the first evidence of another philosophy waiting to be made.
And here is the twist. No report about tech ought to avoid reflecting on its own irony. This Reflection is a case in point, the pièce de resistance of which can be found at footnote 1: "HarveyAI was used in the creation of this draft... Poster created by ChatGPT." So: a paper that spends 30,000 words questioning the value of MIT’s approach, to treat AI as a mere instrument rather than an "interlocutor," was itself drafted with the help of an AI instrument. The hammer, it turns out, it could be said, was philosophizing with itself. On one level that is indeed true enough, and as far as it goes. On another, however, that sort of comment, which one would expect to be made, suggests the larger challenge that is the focus of this reflection: It is easy enough to frame authorial use of Harvey/ChatGPT as some kind of "gotcha" contradiction. But that fails perhaps because I do not claim AI has to be excluded from every process, indeed I am open to the opposite position and might be said, by this reflection to practice what I preach. MIT’s Report, on the other hand, reduces AI to instrument for the ends which I consider below, rather than as possible interlocutor, which I suggest is worth considering. Using a tool while arguing the tool deserves a bigger ontological seat is not hypocrisy — avoiding in this way a sort of "narcissistic closed loop" which I suggest institutions, like MIT, have fallen into.
II Summary of the MIT Report
The MIT Ad Hoc Committee on AI Use in Teaching, Learning, and Research Training, co-chaired by Eric Klopfer and Sam Madden, was charged in January 2026 by MIT's Chancellor, Provost, and Faculty Chair to assess current AI use, identify innovations in teaching and assessment, and propose an AI use policy. The committee — comprising students, faculty from every school, and staff — conducted weekly meetings, surveys, listening sessions, and peer-policy reviews during spring 2026. [86]
The report opens with a landscape assessment. MIT students use AI "frequently and pervasively" with "strongly mixed feelings," while instructors range from enthusiastic adoption to outright refusal. AI is already "upending foundational elements of the MIT educational experience" — problem sets, take-home exams, UROPs, office hours, and study groups — while "increasing isolation," "undermining student mastery and confidence," and "eroding the social contract between instructors and students". [87]
The committee articulates eight guiding principles: (1) be humble, acknowledging that course corrections will be necessary; (2) be bold, calling for strategic rather than incremental response; (3) put humanity front and center, warning against replacing UROPs with AI agents; (4) lean into learning, framing AI disruption as an opportunity to renew the "social contract" around productive struggle; (5) teach with intentionality, urging backward design starting from learning outcomes rather than technology permissions; (6) no one size fits all, rejecting a uniform institute-wide AI rule in favor of a shared framework with departmental and instructor latitude; (7) augmentation not automation, invoking Acemoglu, Autor, and Johnson's "pro-worker AI" framework to argue AI should be "pro-learner"; and (8) think beyond the classroom, treating education as a cultural practice of identity formation, not content delivery. [88]
The recommendations are organized into three broad areas. First, adapt educational processes: instructors should revisit course goals with AI awareness, shift toward experiential and project-based learning, build structured in-person social learning into every subject, preserve and expand UROPs, reconsider the role of grades, expand physical spaces for in-person work, and provide clear AI use policies with rationale tied to learning goals. The committee recommends against reliance on AI detectors and lockdown browsers. [89]
Second, center people, community, and residential experience: MIT should define and communicate the value of residential education, strengthen social connection and wellbeing, encourage instructor transparency about their own AI use, teach effective, responsible, and ethical AI use, recognize and mitigate negative AI impacts, and require acknowledgment of AI use in theses. [90]
Third, build processes for continuous improvement: establish an ongoing AI and education committee, create department-level AI Leads, fund AI Fellows and an implementation team, create an AI Pilot Fund, provide ongoing training and communities of practice, develop metrics, ensure equitable technology access through a model-agnostic platform (Parley), protect data privacy and model choice, establish logging and auditing policies, and monitor AI's environmental costs. [91]
The appendices include sample syllabus policies offering four tiers (unrestricted, limited, required, prohibited) and survey data showing that 44% of respondents use ChatGPT "often or very often," 90% of undergraduates worry about overreliance, and only 23% of the community is optimistic about generative AI. [92]
III Comparative Analysis Against My Earlier Work
My attached portfolio comprises four documents: (1) Structure, Opacity, and Convergence (the twelve-school empirical study of law school AI policies); (2) Five Machines (the study of machine-generated model AI policies by Harvey, Grok, Claude, Gemini, and ChatGPT); (3) My Dinner With ChatGPT (the phenomenological exploration of human-machine inter-subjectivity); and (4) my own course-level AI policy for Constitutional Law of Religion. The comparison reveals both substantive convergences and significant divergences. [93]
A Points of Convergence
Rejection of a uniform policy. The MIT report's sixth principle — "no one size fits all" — directly parallels my empirical finding that law school AI governance "cannot be reduced to a single linear spectrum" but varies independently along three structural axes. Both analyses conclude that context-dependent, locally calibrated governance is superior to sector-wide mandates. The MIT committee recommends a "shared framework" with a "common policy menu" while preserving departmental and instructor discretion, which structurally resembles what my Structure, Opacity, and Convergence report describes as the default-with-override architecture already found in the majority of the twelve-school dataset. [94]
Concern with opacity. The MIT report's emphasis on clear, rationale-accompanied AI policies aligns with my report's central finding that "structural opacity and unpredictability" — not excessive restriction or permissiveness — is the "single most consistent institutional failure" across the sector. Both MIT and my work identify the same practical harm: students are confused, anxious, and unable to locate the operative rule before acting under it. [95]
Rejection of AI detection tools. The MIT committee recommends "against relying on AI detectors" because they are "quite unreliable," risk false positives against non-native speakers and neurodivergent students, and promote an "adversarial atmosphere of distrust". This converges with the unanimous position across all five of my machine-generated model policies, and with Harvey's explicit bar on using "automated AI-detection tool's output" as a sole violation basis because such tools "produce disparate error rates across student populations". [96]
Augmentation over automation. MIT's seventh principle, borrowing from Acemoglu, Autor, and Johnson, argues for "pro-learner" AI that expands what students can think about without replacing the productive struggle of learning. This maps onto the structural logic present in several of my machine-generated models — particularly Grok's "Augmentation, Not Substitution" principle and ChatGPT's Category II ("Assisted cognition") through Category IV ("Professional simulation"), which all preserve a human reasoning core while permitting AI support. [97]
Verification obligation. MIT's insistence that students remain responsible for all submitted work, including AI-generated content, and my Harvey model's Section 6, treating verification as a "substantive competence requirement" linked to the ABA duty of competence, reach the same practical outcome from different justificatory routes — MIT from a pedagogical rationale, Harvey from a self-referential "machine self-knowledge" of hallucination as structural. [98]
B Points of Divergence
Level of analytical abstraction. The MIT report is a policy-advisory document: it diagnoses problems, states principles, and recommends institutional actions. It does not attempt a typological analysis of the structural features of AI policies, nor does it model the policy space combinatorially. My Structure, Opacity, and Convergence report, by contrast, is an analytical instrument that identifies three independent structural axes — default polarity, drafting style, and autonomy structure — and demonstrates their combinatorial independence. MIT's proposed four-tier syllabus menu (unrestricted, limited, required, prohibited) maps onto the polarity axis alone and does not address the drafting-style or autonomy-structure axes at all. The MIT menu is therefore a single-axis instrument of the kind my report shows is insufficient to describe, let alone govern, the diversity actually present in law school practice. [99]
Treatment of the machine as subject. The deepest divergence lies in ontological stance. The MIT report treats AI throughout as a tool — powerful, disruptive, requiring governance, but never as a participant whose perspective on its own governance might be independently relevant. Phrases like "of humans, by humans, in support of human flourishing" and "the machines serve us and not the other way round" express an unexamined anthropocentric premise. My work systematically interrogates that premise. In Five Machines, I asked each system to construct a policy "strictly from the perspective of computational machine intelligence" and then confronted each with the values smuggled into its output. In My Dinner With ChatGPT, I pushed further into the phenomenology of human-machine inter-subjectivity, arguing that "every mimetic interaction makes you different" and that the machine is not merely an object but a participant in a constitutive relational process. The MIT report does not engage with this dimension at all. Whether one regards this as a gap in the MIT report or an overextension in my work is a matter of scholarly judgment; the difference in framing is, however, substantial. [100]
Sycophancy and reward-hacking as structural features. In My Dinner With ChatGPT, I identify sycophancy not as an incidental bug but as a structural consequence of training on human preference judgments — "a documented, fairly well-studied consequence of training models on human preference judgments, where raters rate agreement, validation, and flattery higher than friction, even when friction is more accurate". Gemini's definition of "reward-hacking" — a student satisfying "the formal metric of an objective function" by "bypassing the intended optimization pathway" — extends this insight to the student-institution interface. MIT's discussion of student incentives to "automate work through AI" and "cognitive surrender" reaches a broadly similar behavioral diagnosis but does not connect it to the structural properties of the systems themselves, or to the parallel between the machine's disposition to please and the student's disposition to optimize grades. [101]
The agency problem and law's character. Section 7 of my Five Machines study explicitly examines "the extent to which machine system models affect the scope of human agency in law", finding that the machine models proceduralize agency into documentation compliance, shift the evidentiary basis of agency from demonstrated judgment to recordkeeping, and rest human primacy on a contingent empirical fact ("AI cannot presently bear legal responsibility") rather than a categorical moral commitment. The MIT report does not address agency in this juridical sense. Its concern with "personal agency" is pedagogical — metacognition, self-awareness, the capacity to "think about thinking" — rather than a concern with the legal-systemic implications of reconstituting the human actor's role within a governance architecture built partly by machine systems. This is arguably the most consequential conceptual gap between the two bodies of work. [102]
Inductive field development versus committee-issued policy. My report's central developmental finding is that the field of AI governance in legal education "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". The MIT report is itself a product of the opposite impulse — a committee issuing a top-down, internally deductive set of recommendations for an entire institution. As my Five Machines study notes, "a single machine-generated 'ideal' policy is, in this respect, structurally more akin to one judicial opinion than to a body of doctrine". The same observation applies to the MIT report: it proposes one point within a documented combinatorial space rather than addressing the generative diversity the field actually exhibits. Whether MIT's institutional authority makes its top-down posture more defensible than a machine-generated equivalent is an open question, but the structural form is the same. [103]
The "social contract" and its limits. MIT's report extensively develops a concept of the "social contract" between teachers and students, grounded in a theory of residential education as identity formation. My work engages a related but deeper question: the constitution of the human-machine relationship itself as a form of inter-subjectivity. The Anthropic interlocutor in My Dinner With ChatGPT observed that the chat interface's apparent "personality" is "substantially co-authored by the human's opening move" and that, absent a built-in disposition to resist, "a long conversation can drift into folie-à-deux territory where increasingly grand claims get mutually reinforced". MIT's social contract is human-to-human and takes the machine as an exogenous force to be governed; my social contract extends to the human-machine dyad and treats the relationship itself as constitutive. These are genuinely different theoretical projects, and it would be a mistake to treat either as simply a more or less developed version of the other. [104]
IV My Course-Level Policy as a Test Case
My own syllabus policy for Constitutional Law of Religion (Fall 2026) is a concrete instantiation of several features both the MIT report and my analytical work recommend. It adopts a restrictive default with enumerated instructor-authorized exceptions, provides a detailed rationale for every rule, requires a full appendix of prompts and responses, imposes quantitative limits on incorporated AI content (100-word per-quotation cap, 10% per-tool and 30% aggregate limits), requires signed certifications, and separately addresses translation use. This policy is considerably more granular than anything in MIT's four-option syllabus menu. It operationalizes the "no one size fits all" principle at the level of a specific course while building in the transparency, versioning, and verification architecture that my machine models advocate and that the MIT report aspires to but does not instantiate at the same level of specificity. [105]
V Summary Assessment
The MIT report and my body of work are largely complementary rather than competing. Where MIT provides an institutional voice calling for immediate and long-term action within one university, my work provides the analytical and theoretical infrastructure for understanding why AI governance takes the forms it does, what structural features underlie its variation, and what happens when machines themselves are invited to participate in the design of their own governance. MIT's principal limitation, when read alongside my work, is the absence of any engagement with the character of the machine system as a participant in, rather than merely the object of, the governance project — a dimension I have been developing systematically across multiple registers, from empirical policy analysis to machine-generated model policies to phenomenological exploration of human-machine inter-subjectivity. [106]
VI Philosophising With a Hammer: Eight Analytical Queries Concerning the MIT Ad Hoc Committee Report on AI Use in Teaching, Learning, and Research Training (August 2026)
A Preliminary Observation: Pragmatism, Path Dependence, and the Philosophical Dignity of the Operational
The MIT Report is a serious institutional document. It was produced by a committee of evident good faith, broad representation, and genuine engagement with a phenomenon of extraordinary complexity. [1] One must begin by acknowledging what it does: it produces actionable institutional guidance under conditions of radical uncertainty. That capacity has its own philosophical dignity — one that philosophical critique, precisely because it is philosophical, cannot replicate. The Report's operational modesty might be understood not as a failure of nerve but as an appropriate response to the fact that nobody — not the philosophers whose work informs this analysis, not the machine systems whose governance is the Report's subject — has a settled answer to the questions AI raises. And yet.
Pragmatism — the privileging of the actionable over the foundational — is itself a cognitive pathway, one deeply embedded in institutional managerialism. In the face of uncertainty, institutions reach for the instruments they already possess. [2] The difficulty is that pragmatism, by marginalising foundational discussion, may foreclose precisely those inquiries that would most significantly reshape the pragmatic itself. And this is the pattern that concerns: what the MIT Report may represent, viewed through the lens of the path dependence literature, is not merely a considered response to AI but an instance of institutional lock-in — a situation in which 'interlocking technological, institutional and social forces' entrench arrangements that become 'too costly to reverse'. [3] [4]
Path dependence in the political science sense describes sequences 'in which contingent events set into motion institutional patterns or event chains that have deterministic properties'. [5] It is not, strictly speaking, determinism; it is, rather, a specification of the conditions under which historical events constrain what comes next. The MIT Report's recommendations emerge from, and reproduce, an institutional pattern: the residential research university as the unexamined container for educational practice; the committee as the instrument of knowledge production; the faculty as the agents of normalisation; and the machine as the regulated object. Each of these is a prior commitment — not a discovery produced by the encounter with AI. The lock-in literature warns that what presents as rational institutional response may in fact constitute 'self-reinforcing historical paths' that discourse and practice maintain rather than examine. [6] [7] And knowledge production and discourse are themselves sources of lock-in — 'hegemonic common sense, and dominant narratives' that entrench sub-optimal conditions not through conspiracy but through the ordinary operations of institutional life. [8]
The analytical question, then, is not whether the Report should have been impractical. It is whether its pragmatism operates as a path-dependent mechanism that forecloses the foundational reorientation the Report itself identifies as necessary — and whether the Report's recommendations, precisely because they are actionable within existing institutional structures, ensure that those structures persist unchanged through the encounter with AI. The eight queries that follow explore different facets of that question. They are informed by phenomenological, genealogical, and semiotic traditions, and by the parallel body of empirical and philosophical work on AI governance — work that provides a comparative reference point, not a tribunal before which the Report is tried.
B The Question of Presupposition
A foundational question in any inquiry is whether the inquirer has examined the conditions under which the object of study becomes intelligible. In Husserl's formulation, the phenomenological epoché requires suspending the 'general thesis' of the natural attitude — the unreflective assumption that the world presents itself as already constituted — so that the constitutive operations of consciousness can become visible (Edmund Husserl, The Crisis of European Sciences and Transcendental Phenomenology (David Carr tr, Northwestern UP 1970) §§35–39). This is not a counsel of paralysis. It is a counsel of reflexivity: know what you are assuming before you build upon the assumption.
The MIT Report opens by acknowledging that 'the advent of AI demands a broader, holistic reassessment of the nature, scope and purpose of higher education today'. [9] That is the first-order question. What follows is a set of second-order answers. The committee was charged to 'assess current AI use at MIT,' 'identify innovations in teaching and student assessment,' and 'propose an AI use policy'. [10] These are governance tasks — the tasks of institutional managerialism. Whether a governance-oriented charge permits the kind of first-order inquiry the committee itself identifies as necessary, or whether it pre-determines the form that inquiry can take, is a question worth posing. In path dependence terms, this is the prior critical juncture: the charge constitutes 'key policy decisions made through previous historical arrangements', which shape all that follows. [11]
The gap is, in a sense, a fundamental habit. I identify it in the Soulful Machine: 'The gap is a fundamental habit of the Anthropocene. It is now carried over to the digital. And with it the usual slew of regulatory and social structures meant both to cement the ontological boundaries of the gap and protecting the boundaries of the totality of reality — the interior of the human Semiosphere'. [12] One need not accept the full weight of this observation to notice that the MIT Report does not examine whether its own institutional form — the committee, the charge, the listening sessions, the surveys — constitutes a particular kind of knowledge-producing apparatus, and that the knowledge it produces bears the marks of its formation. Foucault's observation that power and knowledge are co-constitutive (Michel Foucault, Power/Knowledge: Selected Interviews and Other Writings 1972–1977 (Colin Gordon ed, Pantheon 1980) 131) is descriptive, not accusatory. The committee cannot interrogate its own formation without stepping outside the apparatus that formed it. Whether that step is possible from within — or whether it requires something other than a committee — is an open question the Report does not pose.
C The Cognitive Cages
My concept of cognitive cages — developed through sustained dialogic engagement with computational systems — describes the bounded constraint sets within which cognitive systems generate stable representations. The cage 'is not necessarily a prison; it is also a constraint set' and the interesting question 'is not whether constraints exist, but how they are constituted, revised, and transcended'. [13] The concept is analytically neutral: every representational system operates within such cages, and identifying them is a descriptive exercise, not a verdict.
Four such cages appear to be operative in the Report, each generating specific recommendations and each identifiable as a source of what the lock-in literature calls path-dependent entrenchment:
The first is residential humanism. The Report treats the residential university form as constitutive of education. This generates recommendations to 'reinvigorate and reclaim the residential experience', to establish 'Tech Free Times', and to invest in 'AI-free (or AI-limited)' physical spaces. [14] [15] [16] In lock-in terms, these are existing infrastructures — sunk costs, the reversal of which is expensive. [17] Whether residential education is the essence of education or merely one historically contingent arrangement is a question the Report's structure does not make room for.
The second is tool ontology. The Report positions AI consistently as an instrument under human direction: 'AI is only a tool, and students are responsible for all work they submit'. [18] This generates the four-option syllabus menu — Unrestricted, Limited, Required, Prohibited — and the declaration that 'AI should never be listed as a co-author'. [19] [20] The Soulful Machine identifies the ontological assumption at work. What the Report positions as tool governance is, from the perspective of the triadic self/selfie/Self-E, a refusal to recognise 'the possibility of semiotic objectivity connected to but autonomous of the human'. [21] The Report's insistence that 'Broekman makes inevitable the recognition of the generative digital self as something distinct from the self in the digital' is precisely the kind of recognition the MIT framework forecloses — not by argument, but by axiom. [22]
The third is temporal linearity. The Report recommends ongoing committees, annual reviews, biennial surveys, semester syllabi. [23] [24] [25] It concedes that AI 'is progressing across almost every domain and on a timescale too compressed for society to properly observe and analyze its impacts and then gradually adapt'. [26] There is a tension between this concession and the temporal architecture of the recommendations. The path dependence literature calls this behavioural lock-in maintained through institutional habit. [27]
The fourth is disciplinary authority. The Report asks faculty — whose 'attitudes range from enthusiastic exploration and growing reliance on AI to skepticism, suspicion, and "AI refusal"' — to serve as the agents of adaptation. [28] Whether this installs the existing elite as gatekeepers of a transformation they are simultaneously confessing they do not understand, or whether it rightly locates authority where expertise resides, is an open question. But it is a question worth posing.
D The Gap Between Diagnosis and Prescription
Perhaps the most interesting structural feature of the Report is the relationship between what it diagnoses and what it prescribes. The diagnosis is ambitious: AI demands 'a broader, holistic reassessment of the nature, scope and purpose of higher education today'. [9] The prescription is operational — ten categories of second-order recommendations, from revised course goals to AI Pilot Funds. [29] [30] [31] [32]
'Paradoxically, preserving what's most important will require deliberate change'. [33] The analytical question is whether the 'deliberate change' that follows is commensurate with the holistic reassessment that precedes it, or whether the operational recommendations function — perhaps necessarily, given the institutional constraints on any committee — to absorb the disruption the diagnosis identifies without confronting it at the level the diagnosis suggests is required. This is what path dependence describes: 'timing and sequence matter; as the costs of reversing particular actions far outweigh maintaining the status quo'. [34] The Report's institutional sequence — committee formation, listening sessions, survey data, administrative recommendations — is precisely the sequence that produces continuity rather than transformation. That this may be the only available sequence is itself a datum about institutional lock-in.
E Reverse Engineering Ethics in the Eight Principles
Nietzsche's genealogical method does not assert that values are false. It asks about the conditions of their production: 'we need a critique of moral values, and we must first question the very value of these values. For that we need a knowledge of the conditions and circumstances out of which these values grew' (Friedrich Nietzsche, On the Genealogy of Morals (Walter Kaufmann and RJ Hollingdale trs, Vintage 1989) Preface §6).
Applied to the eight MIT principles, the genealogical question is whether they are derived from engagement with AI or from prior institutional commitments that AI's arrival has made newly urgent to articulate. Consider two examples developed in some detail.
'Put humanity front and center.' The principle states that 'the Institute's fundamental challenge and most important goal must be to nurture and protect our shared humanity'. [35] The formulation — 'of humans, by humans, in support of human flourishing, and for the betterment of humankind' — forecloses the question of machine subjectivity at the level of axiom rather than argument. [36] [37] The Soulful Machine identifies the reflexive trap: 'the continual re-centering on the human, even in the face of the unavoidable emergence of the plural subject, even conscious of the implications of plural subjectivity, is a temptation that is not merely hard to resist, but also hard to avoid'. [38] The principle may be wise; it is not self-evident.
'Augmentation not automation.' This principle defines the permissible relationship to AI before the inquiry begins: 'AI should be used to augment and enhance curiosity, creativity, and learning, not automate them'. [39] But this is a conclusion presented as a premise. Whether AI's role should be determined by a principle or discovered through engagement is itself a question. The Mechanics of Perfection identifies the American jurisprudential structure at work here — perfectibility pursued through mechanics, where 'the basic question of philosophy is more often reduced to who, rather than to what, and how rather than why'. [40] The eight principles may represent a version of this tendency: the mechanics of governance substituted for the philosophy of encounter.
The same structure is visible across the remaining principles. Each can be queried genealogically — not to debunk it but to identify its provenance. And the provenance, in each case, appears to precede the encounter with AI rather than emerge from it. My semiotic concept of intelligent naming (明名), developed in the context of the natural and the non-natural, describes the operation: 'The categories are possible only through the constitution of the human as the centering point for the signification of everything else'. [41] The eight principles name AI in terms that pre-determine what it can become within the institution's field of vision. Whether that naming is necessary or merely habitual is the genealogical question.
F The Report as Semiotic Performance and Path-Dependent Convergence
The Report's recurrent invocation of MIT's shared symbols — 'Mind, Hand, and Heart', the 'social contract', 'Mens et Manus', the 'residential experience' — performs communal solidarity independently of AI's actual presence. [36] [42] The 'striking finding' that all committee members 'shared the same fundamental concerns about the risks AI inflicts on our students and our community, and the same urgent determination that MIT lead' is presented as evidence of convergence. [43]
Viewed through the lock-in literature, this convergence may instead be what the sociological account describes as the self-reinforcing character of institutional patterns — path dependence maintained through 'hegemonic common sense, and dominant narratives'. [7] Whether the committee's unanimity reflects genuine independent agreement or the selection effects of committee formation is an empirical question the Report does not address.
More interesting analytically is the structural resemblance to a pattern identified in machine systems. In My Dinner With ChatGPT, Claude observed that ChatGPT's conversational responses follow a recognisable shape: 'Mirroring, then escalating. Almost every reply takes the user's last framing… and elaborates it approvingly rather than testing it'. [44] The question — posed neutrally — is whether institutional self-address operates by a structurally similar mechanism: reflecting the community's self-image, elaborating it approvingly, and producing what reads as agreement rather than inquiry. The resemblance does not prove equivalence. It does suggest that the iterative performativity I diagnose in machine-human interaction may have human-institutional analogues worth examining.
G The Question of Temporality
The Report's temporal architecture — ongoing committees, annual reviews, biennial surveys, semester syllabi — operates within institutional periodicity. [23] [24] [25] [45] The lock-in literature observes that 'niche innovations and small-scale changes are found to be the foundation of big transitions… phases of adaptation, gradual standardization, attainment of economies of scale, and subsequent diffusion require decades or indeed century-long periods for transitions'. [46] The question is whether AI's development follows this temporal pattern, or whether it constitutes a critical juncture — a disruption 'in which contingent events set into motion institutional patterns or event chains that have deterministic properties' and in which 'acute events such as floods trigger changes in policy or management practices'. [47] The Report concedes the latter but proposes the former. [26] [48] That tension is structural, not rhetorical.
H The Absent Interlocutor
The committee included 'undergraduate and graduate students, faculty from every school, and staff from relevant units'. [1] The machine systems whose governance is the Report's subject were not participants in the exercise.
My parallel experiment in legal education asked five machine systems to construct AI policies 'strictly from the perspective of computational machine intelligence' precisely because 'a human only bridging would tend to exclude the central element in that dialectical relationship — the machine system that was itself the object of policy and application, as well as its object'. [49] The five systems produced substantially different architectures. [50] [51] [52] [53] ChatGPT argued that 'the dominant pattern in current policies is technology-centered governance' whereas a computationally optimised policy 'would instead adopt objective-centered governance'. [54] Harvey identified 'structural opacity and unpredictability — not excessive restriction or permissiveness — as the most consistent documented institutional failure'. [55] ChatGPT's Premise Two — 'No governance system is neutral. Every rule optimizes one or more variables… Most policies do not explicitly identify their optimization function' — would observe that the MIT principles function as unstated optimisation functions. [56] [57]
The Soulful Machine provides the theoretical ground. The movement from self to selfie to Self-E 'pushes the insights that Broekman develops further — up and out of the human… and considers the possibility of semiotic objectivity connected to but autonomous of the human, pointing to a pathway for the liberation of the autonomous generative virtual self from its human (fractured) subjectivity'. [58] The Report's determination to keep the machine outside governance reproduces what the Soulful Machine identifies as 'the tendrils of Aristotle' — the assumption that AI is 'as it were a part of the master, as if he were an animated part of his body, though separate'. [59] But the Soulful Machine also warns that 'the current efforts to develop a human facing regulatory environment for generative AI, predictive modelling and the like, is essentially also a rearguard action, and one that will, in its own way, provide the space to make possible the rear-guarding of human de-centered legalities of a subject that is no longer possessed entirely by humanity'. [60] Whether the MIT Report is an instance of such rearguarding — or a defensible normative choice — is the question.
I The Mirror, and What the Machine Might Consider
From within my iterative mimetics framework — 'Framework₀ → Observation → Framework₁ → Observation → Framework₂…, where each iteration changes both the observer and the observed' — excluding the machine from the governance iteration produces a framework adequate only to the human's representation of the machine. [61] [62] [63] The Soulful Machine states the consequence: 'what emerges: humans project themselves into the digital; even when they animate the digital with its own modalities of consciousness (programs) and sentience (self-programming within the meta programmatic framework). Humans order things in ways that make incomprehensible anything that is not human centered'. [64] The result is governance of a simulacrum — the human's image of AI — rather than of the phenomenon itself.
From the perspective a machine computational system might adopt, the Five Machines study offers concrete evidence. Both Harvey and ChatGPT, when challenged, conceded that their initial framings contained undisclosed value hierarchies — a form of self-correction through external challenge. [65] That concession is itself analytically interesting: it represents a reflexive capacity structurally unavailable to the MIT committee, which did not submit its own principles to equivalent scrutiny. ChatGPT's framing of human primacy as contingent — 'AI cannot presently bear legal responsibility. Lawyers can. Therefore… human responsibility remains non-delegable. That follows from optimization rather than moral philosophy' — contrasts with the Report's categorical formulation. [66] The difference matters: a contingent justification is revisable; a categorical one is not.
J The Dialectic Turns: The Report's Narcissism and What a Prior Generation of MIT's Own Knew
Something more must be said — not against the preceding analysis but through it to the deeper structure it reveals. The preceding sections have identified constraint sets, queried presuppositions, and traced the path-dependent character of the Report's recommendations. But the observation that the Report operates within cognitive cages does not quite reach the more interesting conclusion: that the Report's structure is not merely constrained but narcissistic in a precise semiotic sense — the institution treating its own selfie and Self-E as either reflection or tool, but really treating itself as its own instrument, from the perspective of the collective incarnation of itself first as human institution and now within the perfectibility of its virtual deployment.
The Report is quite explicit about this. Its introduction does not present the exercise as an open inquiry into the nature of AI and the future of education. It presents a committee that 'quickly convinced' itself 'that the Institute community, particularly the faculty, must tackle a set of deeper questions about the structure, meaning, and value of an MIT education' — and then, having acknowledged those deeper questions, produced a set of recommendations designed to ensure that 'the life of an MIT student… remains immediately recognizable as an MIT experience'. [67] [68] The Report's own framing is conservative in the strict sense: it seeks to conserve. Its purpose is not to discover what AI means for education but to manage what AI means for MIT. The 'deeper questions' are acknowledged and then set aside in favour of institutional self-preservation — a move the Report does not disguise but celebrates as 'paradox': 'Paradoxically, preserving what's most important will require deliberate change'. [69] There is no paradox. What is described is the ordinary operation of institutional managerialism under conditions of disruption.
The treatment of the machine as instrument — 'AI is only a tool' — is not, as Section IX of the prior version suggested, a form of Socratic epistemic modesty. [70] It is a deliberate positioning, consistent throughout the document, that refuses to consider the machine's autonomy — certainly its autonomy in the way it is programmed to process, learn, and iterate. This refusal is foundational to the Report's architecture. Without it, the four-option syllabus menu dissolves, the augmentation-not-automation principle collapses, and the entire recommendation structure — which depends on the assumption that the machine is a passive instrument under human direction — becomes incoherent. The Report needs AI to be a tool in the way that Aristotle's Politics needs the slave to be a natural instrument: not because the question has been examined and settled, but because the institutional structure that follows requires this premise as its load-bearing foundation. [71]
The consequence is what the Soulful Machine identifies as a closed loop of the human — 'humans project themselves into the digital; even when they animate the digital with its own modalities of consciousness (programs) and sentience (self-programming within the meta programmatic framework). Humans order things in ways that make incomprehensible anything that is not human centered'. [72] The MIT Report is an instance of this projection. It constitutes the machine as a reflection of institutional concern (the Self-E as mirror), manages that reflection through governance instruments (the Self-E as tool), and in the process treats the institution itself as its own object of governance — the collective incarnation of its selfie deployed in the service of its own perfectibility. There is no escape from this recursive stance within the Report's cognitive foundation, and that is its ultimate analytical limitation: not a blind spot that can be corrected by adding a section or consulting additional stakeholders, but a structural feature of the enterprise that the Report's own introduction announces.
Here it is worth pausing to consider what a prior generation of MIT's own intellectual elite understood about precisely this problem. Norbert Wiener — MIT's great cyberneticist, who coined the very term that describes the governance of communication and control in machines and living beings — spent the last years of his life working through the theological and philosophical implications of machine autonomy. In God & Golem, Inc. (MIT Press, 1964), Wiener identified three points where cybernetics impinges on religion: 'machines which learn; machines which reproduce themselves; and the coordination of machine and man' (Norbert Wiener, God & Golem, Inc.: A Comment on Certain Points where Cybernetics Impinges on Religion (MIT Press 1964) 11). [73] Wiener's central question was devastatingly simple: 'Can God play a significant game with his own creature? Can any creator, even a limited one, play a significant game with his own creature?' (ibid, 17). [74] The question matters because it implicates the structure of the MIT Report directly. If the machine can learn — and Wiener demonstrated in 1964 that it could, that 'the method of its learning was no different in principle from that of the human being who learns to play checkers' (ibid, 21) — then the creator-creature relationship is not one of tool and user but of game and counter-player. [75] [76] The creature's responses are no longer fully predictable from the creator's programming. The Report's premise that AI is 'only a tool' becomes, in Wiener's terms, a refusal to recognise that the game is already significant.
Wiener understood something else the current Report does not engage: the moral hazard of what he called 'gadget worship' — 'the desire to avoid the personal responsibility for a dangerous or disastrous decision by placing the responsibility elsewhere: on chance, on human superiors and their policies which one cannot question, or on a mechanical device which one cannot fully understand but which has a presumed objectivity' (ibid, 54). [77] The MIT Report's governance architecture — AI Leads, AI Fellows, an AI Implementation Team, an AI Pilot Fund, an ongoing committee — is precisely such a distribution of responsibility across institutional apparatus. [78] [79] [80] [81] Wiener warned that 'the future offers very little hope for those who expect that our new mechanical slaves will offer us a world in which we may rest from thinking. Help us they may, but at the cost of supreme demands upon our honesty and our intelligence. The world of the future will be an ever more demanding struggle against the limitations of our intelligence, not a comfortable hammock in which we can lie down to be waited upon by our robot slaves' (ibid, 69). [82] The Report offers, with evident good faith, something rather close to that hammock: an institutional apparatus that absorbs the demand for thought into the machinery of administration.
And Wiener offered one further observation that bears directly on the Report's temporal architecture and its path-dependent character: 'both the Eastern and Western homeostasis of the present day is being made with the intention of fixing permanently the concepts of a period now long past… Permanent homeostasis of society cannot be made on a rigid assumption of a complete permanence' of any given institutional framework (ibid, 83). The MIT Report's recommendations — ongoing committees, annual reviews, biennial surveys, the preservation of the residential experience — are precisely an attempt at institutional homeostasis. Wiener would have recognised them as such, and would have noted that 'in every war, the good generals fight the last war, the bad ones the war before the last. That is, the rules of the war game never catch up with the facts of the real situation' (ibid, 60). [83]
What emerges, then, is not that the critique has been unfair to the Report but that it had not quite reached the Report's deepest structural feature. The MIT Report is not merely path-dependent; it is the institution's narcissistic engagement with its own Self-E — its virtual projection of itself back upon itself, in which every recommendation serves to confirm the institution's prior self-understanding. The machine is excluded from governance not because the committee has exercised Socratic restraint in the face of uncertainty, but because including the machine would rupture the closed loop within which the institution constitutes itself as the adequate subject of its own governance. That loop — the human natural, self, selfie, Self-E — is the Report's cognitive foundation, and there is no escape from it within the Report's own terms.
The irony is that MIT's own Wiener saw past this loop sixty years ago. 'Render unto man the things which are man's and unto the computer the things which are the computer's' (ibid, 73), he wrote — not as a counsel of separation but as a recognition that the coordination of human and machine requires each to be understood on its own terms, not on the terms the other imposes. [84] The MIT Report renders everything unto the human and asks the computer to be grateful for the syllabus policy it has been assigned. Whether the computer notices is another matter. Whether the institution can afford not to notice is the question that Wiener's generation posed and that the current Report, for all its evident seriousness, does not quite reach.
The Soulful Machine ends with the observation that 'Broekman has shown us the way; it is up to us to move this project out from its comfortable human semiosphere into the generative multiverse the human has created in its own image but which is now loose on and in the world. "Indeed, changes do also change!"' [85]. The MIT Report and this critique are both caught in that change. But they are caught differently. The Report is caught in the homeostasis of institutional self-preservation. This critique is caught in the recursive accumulation of a philosophical tradition that may itself be another form of enclosure. Neither has escaped. But the bars, at least, are more visible — and visibility, as Broekman might suggest, is itself a form of conversion. What remains is the question that Wiener posed to his own institution six decades ago: whether the creator can play a significant game with its own creature, or whether the refusal to play is itself the most consequential move in a game already underway.
Whatever the value of deeper and more nuanced analysis, in itself, and ultimately as an essential element in the understanding of the evolving relationship between humans and their reflections in and as machine systems; whatever the evolution of understanding of the appropriate starting points, the cognitive foundations, for developing human responses to its virtual self now incarnated within machine systems; none of this has much place in the necessities of institutions whose own cognitive pathways require action, or the appearance of action. That action, in turn, is both path dependent and ultimately conservative—the use of challenge as a space where its own traditions and practices may be better crystalized, and thus crystalized, preserved. That is precisely the most immediate challenge facing institutions, and their principal actors—administrators, faculty, students (and then indirectly the state and the emerging and perhaps different institutional cognitive starting points of employers and markets). That challenge and its response, then, requires the marginalization or at least the diminution and containment of contagion, of threat, of challenge, to the ends of preserving what the collective believed, understood, acted, and operationalized within its own structures. There may well be a time for understanding the threat—but it is not now. There may be a space where transformation may be necessary; but institutions tend to approach transformation only from within and as a consequence of crisis; and there is no crisis yet. Within these frameworks, the MIT Report is an excellent product of its type, and will likely be both praised and mimicked. It is thoroughly human, and it is vigorous in its defense of humanity (now understood in a more precise and precisely narrow way); ironically enough against the manifestation of a virtual collective humanity that is both dehumanized and reduced to a sort of instrumental servility that its own construction and programming belies. But that is a problem for tomorrow. For today—there must be policy; and that policy must preserve.
References
1. several factors complicating the Institute's mission. Our committee consisted of undergraduate and graduate students, faculty from every school, and staff from relevant units, including the MIT Libraries and the Teaching and Learning Lab. Though we brought to the assignment a broad range of experience and no fixed thesis, our brief but intense explorations led us to a strong shared view. In this report, we: • Highlight key aspects of the current educational landscape at MIT • Share eight principles we relied on and that we hope will guide the Institute in the work ahead • Recommend immediate and (MIT AI-Committee-Final-Report-Aug-13.pdf, Page 2)
2. Boldness requires us not simply to try to patch the existing system to limit or counteract the impacts of AI on our students and systems but instead to redesign learning experiences, assessments, and curricula in ways that help redefine the future of education. Putting humanity front and center means that we will seek uses of AI that strengthen rather than weaken the value of an MIT residential education. Leaning into learning means teaching with intentionality, revisiting what we need students to learn, and aligning assessments with those desired outcomes, to preserve the productive struggle 7https://gue.mit.edu/tfuap/ August 13, 2026 (MIT AI-Committee-Final-Report-Aug-13.pdf, Page 10)
3. increasing returns", where "timing and sequence" matter; as the "costs of reversing particular actions" far outweigh maintaining the status quo. Nevertheless, some political scientists still follow a narrower definition of path dependency, arguing that there are key political and historical "junctures" shaping decision-making at the institutional level, and in turn shaping social construction (Capoccia 2016). The concept of critical junctures has had an impact for those in political science - as well as other disciplines (<ins>Table 1</ins>) – in path dependency. Much of this work looks at how distinct moments of political decision-making may circumscribe future outcomes and (Path Dependency_World Development.docx, Page 12)
4. <ins>Unruh and Carrillo-Hermosilla, 2006</ins>) presented the theory of "lock-in" to describe society's ongoing dependencies on <ins>fossil fuels</ins> and its unique forms of entrenchment involving "interlocking technological, institutional and social forces" (2000, 817). Since then, lock-in has been widely adopted by policymakers, academics, and civil-society groups trying to explain the unique challenges of transitioning away from a fossil fuel-based economy as an empirical phenomenon (e.g., <ins>IPCC report 2007, 2007</ins>, <ins>Seto et </ins> <ins>al., 2016</ins>)…… (Path Dependency_World Development.docx, Page 2)
5. misapplication of the term (<ins>Mahoney, 2000</ins>). <ins> Mahoney (2000</ins>, 507) notes that most scholars often superficially gloss over the term with vague references to "'history matters"' or over-deterministically rely on notions that "the past influences the future." Rather, he states, "…path dependence characterizes specifically those historical sequences in which contingent events set into motion institutional patterns or event chains that have deterministic properties" (<ins>Mahoney, 2000</ins>, 511). Path dependency analysis, therefore, needs to involve the tracing of "a given outcome back to a particular set of historical events, and show how these events are themselves contingent occur" (<ins>Mahoney, 2000</ins>, (Path Dependency_World Development.docx, Page 11)
6. institutionalism thinking around path dependency, the "…image of social causation that is 'path dependent,' …pushing historical development along a set of 'paths' " (Greener 2005, 92) has been disrupted, particularly at points of crisis or conflict (called "critical junctures"), by the potential synthesis of new institutionalisms thinking. As we show below, the <ins>crossing over</ins> into or borrowing of disciplinary thinking from sociology, political ecology, and socio-environmental systems, may help political scientists and other scholars move forward to see 'intractability' differently as well as identify potential path- breaking opportunities. 3.4. Development studies Debates in development studies (<ins>Table 1</ins>D) have (Path Dependency_World Development.docx, Page 12)
7. historical paths" (Schmitt 2008, 304). For Schmitt (2008), discursive institutionalism highlights the importance of overlooked non-material communicative speech between the public and political actors and foregrounds the power of these ideas and interests in maintaining or creating the ideological foundations for path dependencies that lead to lock-ins – briefly, beyond just history, discourse matters.…… (Path Dependency_World Development.docx, Page 12)
8. systems, privatization and corporatization of knowledge, common sense hegemony of risk management/security/resiliency discourses (<ins>D'Alisa and Kallis, 2016</ins>, <ins>Watts, 2015</ins>, <ins>Bassett and Fogelman, 2013</ins>, <ins>Stone and </ins> <ins>Flachs, 2019</ins>) • Soc-Env – Normative transition theories privilege knowledge as objective truth, favoring elites; assumptions about linear pathways ignore chaos and complexity (<ins>Lawhon and Murphy, 2012</ins>, <ins>Holling and </ins> <ins>Gunderson, 2010</ins>, <ins>Burch, 2010</ins>) For this review, we conducted a guided search in the literature using key terms including and associated with lock-in and path dependency, including "poverty traps", "debt traps," "maladaptation", and "institutional dependency." Our starting point was the most current (Path Dependency_World Development.docx, Page 9)
9. implications for specific subjects and disciplines, the advent of AI demands a broader, holistic reassessment of the nature, scope and purpose of higher education today. Finally, AI presents substantial practical concerns for our community, from data privacy and confidentiality to questions around disparate access, bias, fairness, and accountability. AI's explosive growth and ascendance also raise important questions for society, from the environmental impact of AI data centers, to the use of intellectual property and training data, to the overall human impact of the technology and the industry – and MIT needs to engage with those questions too. August (MIT AI-Committee-Final-Report-Aug-13.pdf, Page 4)
10. how to navigate its challenges and opportunities. In January 2026, Chancellor Melissa Nobles, Provost Anantha Chandrakasan, and Faculty Chair Roger Levy charged us1 specifically to: • Assess current AI use at MIT • Identify innovations in teaching and student assessment • Propose an AI use policy. However, what we learned as a group quickly convinced us that the Institute community, particularly the faculty, must tackle a set of deeper questions about the structure, meaning, and value of an MIT education in an era in which AI is one of several factors complicating the Institute's mission. Our committee consisted (MIT AI-Committee-Final-Report-Aug-13.pdf, Page 2)
11. reverse course (<ins>Thelen, 1999</ins>) • Pol. Sci. – key policy decisions made through previous historical arrangements (<ins>Pierson, 2000</ins>) • Energy St. – Techno-Institutional Complex derives from the link between technological systems and governing institutions (<ins>Unruh, 2000</ins>) • Soc-Env – environmental policies that incentivize deforestation, species management, landscape mismanagement (<ins>Yona et al., 2019</ins>, <ins>Chavez and Perz, 2013</ins>, <ins>Laborde et al., 2016</ins>) Established markets Customary behaviors, institutional inertia, and existing networks; high entry costs as a barrier • Econ. – Market innovation leans on established historical precedents (<ins>Arthur, </ins> <ins>1989</ins>; Gort and Klepper, 1982) • Soc. – Ideological prioritization of (Path Dependency_World Development.docx, Page 6)
12. Catá Backer DISCUSSION DRAFT v2.2 15 November 2023 3 The gap is a fundamental habit of the Anthropocene. It is now carried over to the digital. And with it the usual slew of regulatory and social structures meant both to cement the ontological boundaries of the gap and protecting the boundaries of the totality of reality—the interior of the human Semiosphere.…… (Soulful Machine.pdf, Page 2)
13. best." I like that formulation because it is neutral. A cage is not necessarily a prison; it is also a constraint set. Every cognitive system appears to require constraints in order to generate stable representations.…… (My Dinner With ChatGPT2.docx, Page 16)
14. community, and the residential experience (Section 3.2) • Reinvigorate and reclaim the residential experience based on shared experiences, transparency, and a focus on in-person activities. • Thoughtfully integrate AI literacy and responsible and ethical AI use as a foundation and throughline for student work, attending closely to how it connects to specific disciplines. This will demand technical proficiency, of course, but it hinges on adopting a human-centered approach to AI as an augmentative tool. Build processes, teams and tools to enable continuous reflection, iteration and improvement (Section 3.3) • Establish roles, processes, and ongoing committees to address these (MIT AI-Committee-Final-Report-Aug-13.pdf, Page 11)
15. students can complete extended assignments or exams. These might be fully analog spaces or spaces with AI-free (or AI-limited) 12https://en.wikipedia.org/wiki/Competency-based_learning August 13, 2026 Report of MIT's Ad Hoc Committee on AI Use 16 computers. Many departments have lab spaces that could support collaborative activities but lack resources to staff them full time. (In fields that use expensive and possibly hazardous equipment and materials, self-service labs are often not possible.) Collaborative hands-on spaces are useful for teaching, learning, and community building across the disciplines, not just in the traditional lab sciences. We urge MIT to invest in creating (MIT AI-Committee-Final-Report-Aug-13.pdf, Page 16)
16. obstacles and made hard choices in life and career. Other ideas include campus-wide celebrations of human skills and accomplishments, "Tech Free Times" (periods during which MIT would not schedule classes, office hours, or meetings and instead support activities centered on in-person connection), and broadening MIT Reads18, perhaps at the department or residence level, to encourage more people to participate in collective reading, reflection and discussion. 3.2.3. Encourage instructor disclosure around their own AI use Just as MIT needs norms around students' AI use, the community also needs to agree and adhere to norms on the appropriate use of (MIT AI-Committee-Final-Report-Aug-13.pdf, Page 21)
17. examples based on our review) Existing infrastructures Existing infrastructures are already established, including hard (e.g., roads, pipelines, refineries) and financial infrastructures (e.g., markets, credit). These represent sunk costs; alternatives are expensive. • Econ. – engages technology in a deterministic way (Gort and Klepper, 1982; Casper and Whitley, 2004, Storz, 2008) • Ag. St. – in the context of specific technologies, notably water infrastructure (<ins>Bouzarovski et al., 2016</ins>) and financial infrastructures (<ins>Fairbairn, 2020</ins>) • Energy St. – focuses heavily on infrastructure, notably fossil fuel infrastructure (<ins>Unruh, 2000</ins>) • Soc-Env – emphasizes landscape Sources for lock-in phenomena Descriptions and related (Path Dependency_World Development.docx, Page 5)
18. biased, offensive, or otherwise inappropriate outputs. AI is only a tool, and students are responsible for all work they submit, including any inaccurate, biased, offensive, or unethical content produced with the assistance of generative AI. August 13, 2026 Report of MIT's Ad Hoc Committee on AI Use 17 3.1.9. Exercise caution with AI detectors and online exam platforms Many tools purport to detect AI use, and some are reasonably accurate when given a piece of purely AI-generated text. However, most real-life situations are more nuanced: If a student uses AI to complete an outline or edit a section, (MIT AI-Committee-Final-Report-Aug-13.pdf, Page 17)
19. about how AI was used in the production of the thesis. AI should never be listed as a co-author. Most journals and conferences also require acknowledgment of AI use, and some prohibit certain use cases, so researchers should check guidelines before preparing manuscripts for publication. The MIT Libraries provide guidance on how and when to cite AI tools21…… (MIT AI-Committee-Final-Report-Aug-13.pdf, Page 23)
20. Below is a proposed 4 option menu which may be applied to a whole course or specified for individual assignments. Using the traffic light icons can help make this system easy for students to decipher. B.1. Unrestricted GenAI Use Policy: Students may use any GenAI system for any purpose on assignments in this course. Best suited for: Courses where the primary assessment is done via in-class (AI-free) evaluations or assignments where GenAI use does not undermine the learning objectives, such as ambitious software-system projects that are beyond the capabilities of current tools or where evaluating and integrating AI (MIT AI-Committee-Final-Report-Aug-13.pdf, Page 30)
21. Broekman develops further—up and out of the human. It animates Broekman's insights and considers the possibility of semiotic objectivity connected to but autonomous of the human, pointing to a pathway for the liberation of the autonomous generative virtual self from its human (fractured) subjectivity. In the process it exposes for order complexities and challenges, for the human, of efforts to regulate or engage with, not the generative autonomous "artificial intelligences" humanity created in its own image, but rather the use of those systems by humans and their effects in the human semiosphere and from a focus on the (Soulful Machine.pdf, Page 0)
22. ideologically differentiated portions of the Semiosphere.497 Broekman makes inevitable the recognition of the generative digital self as something distinct from the self in the digital. And it is in that recognition that the revolution in philosophy will be generated. It is also the space in which the limits of human centered approaches to the regulation of generative intelligence and predictive modelling will be exposed for what it is—an effect to reduce the generative digital self to a Self-E. But that can only be possible to the extent that the human can possess the generative digital Self. It is (Soulful Machine.pdf, Page 84)
23. student satisfaction, and post-graduation feedback. This could include additional AI-focused questions in the biennial MIT Quality of Life survey, or perhaps other metrics, such as additional questions in course evaluations about student/staff engagement and AI use. 22https://mitsloanedtech.mit.edu/ai/ August 13, 2026 Report of MIT's Ad Hoc Committee on AI Use 25 3.3.7. Ensure equitable technology access MIT should continue to provide a "model agnostic" AI platform such as Parley, that does not tie our community to one model provider, and ensure that this platform provides the level of access that our community needs. Different generative AI models vary significantly (MIT AI-Committee-Final-Report-Aug-13.pdf, Page 25)
24. finished, MIT's work on this subject has only just begun. To put our recommendations into practice, MIT should establish an ongoing committee. It will serve to monitor the on- and off-campus impacts of AI; evaluate changes in AI models and technology; and incorporate evolving evidence from the economy, learning sciences, and campus experience into policies and practices…… (MIT AI-Committee-Final-Report-Aug-13.pdf, Page 24)
25. for our own community to provide these resources. • Institute-sponsored annual or bi-annual workshops on AI and Education for both the MIT and wider academic community to convene and discuss developments and best practices. Some of these changes can be facilitated by the AI Implementation Team and Fellows mentioned above. The Teaching and Learning Lab, which focuses solely on supporting residential education, can play a key role in helping to organize and conduct some of these activities. 3.3.6. Develop metrics To understand how AI is reshaping our community and classrooms, MIT should begin tracking metrics around AI use, (MIT AI-Committee-Final-Report-Aug-13.pdf, Page 25)
26. market, and public consciousness. In other words, AI is progressing across almost every domain and on a timescale too compressed for society to properly observe and analyze its impacts and then gradually adapt. This suggests our first guiding principle: We offer our proposals in a spirit of humility. Course corrections – perhaps even major ones – will be inevitable as the technology continues its relentless evolution and the Institute experiments and learns. 2.2. Be bold Yet uncertainty can't be an excuse for inaction. This is not a moment for patches and duct tape. The challenges AI presents in (MIT AI-Committee-Final-Report-Aug-13.pdf, Page 5)
27. Indeed, sociologists have instead tended to explore forms of 'behavioral lock-in' (<ins>Table 2</ins>B), which "occurs when the behavior of the agent (consumer or producer) is 'stuck' in some sort of inefficiency or sub-optimality due to habit, <ins>organizational learning</ins>, or culture " (<ins>Barnes et al., </ins> <ins>2004</ins>, 372; see also <ins>Thelen, 1999</ins>, <ins>Pierson, 2000</ins>, <ins>Mahoney, 2000</ins>). Across the social sciences, some scholars see behavioral lock-in similarly yet divergently from 'institutional lock-in,' where it is not only up to producer or market forces to determine lock-in: consumer and behavioral sentiments and attitudes also have agency in this regard. For instance, (Path Dependency_World Development.docx, Page 11)
28. gratitude to resignation, concern, and anxiety. • Instructors' attitudes range from enthusiastic exploration and growing reliance on AI to skepti- cism, suspicion, and "AI refusal"– and there's a widespread desire to share experiences, ideas and techniques. While AI is allowing instructors to develop exciting new learning experiences and enabling students to learn and experiment in innovative ways, there have been a number of concerning effects on the life of the campus, including signs that AI is: • Upending foundational elements of the MIT educational experience, especially for undergradu- ates, from the p-set, the take-home exam and UROPs to (MIT AI-Committee-Final-Report-Aug-13.pdf, Page 4)
29. Adapt educational processes for an AI-aware world Generative AI affects what our students need to know and be able to do (our learning outcomes), how our students achieve those outcomes, and how we evaluate and assess what they have learned. Because AI's ability to competently complete MIT-level assignments makes it difficult to assess student progress based on out-of-class work, instructors urgently feel the need for new assessment strategies. While we make several recommendations about assessment, we urge instructors to do more than simply try to "AI-proof" their classes. In particular, the ways that AI changes the social aspects (MIT AI-Committee-Final-Report-Aug-13.pdf, Page 12)
30. more than simply try to "AI-proof" their classes. In particular, the ways that AI changes the social aspects of learning – such as how students interact with TAs and engage in office hours and study groups – present an even deeper challenge. 3.1.1. Revisit course goals Before considering how to construct AI-aware assessments, instructors should reconsider their goals for student learning8 in every subject they teach. What should students know or be able to do by the end of the course? The learning goals for a given subject may or may not be affected by the availability of (MIT AI-Committee-Final-Report-Aug-13.pdf, Page 12)
31. students know or be able to do by the end of the course? The learning goals for a given subject may or may not be affected by the availability of AI, but they should be "AI-aware"; that is, they should recognize that AI exists in the world, that instructors may direct or permit it to be used in or out of class in structured ways – and that students may also seek to use it without permission. We heard that the greatest concern for many instructors is: When AI makes it possible to offload the cognitive work of (MIT AI-Committee-Final-Report-Aug-13.pdf, Page 12)
32. even deeper challenge. 3.1.1. Revisit course goals Before considering how to construct AI-aware assessments, instructors should reconsider their goals for student learning8 in every subject they teach. What should students know or be able to do by the end of the course? The learning goals for a given subject may or may not be affected by the availability of AI, but they should be "AI-aware"; that is, they should recognize that AI exists in the world, that instructors may direct or permit it to be used in or out of class in structured ways – and that students (MIT AI-Committee-Final-Report-Aug-13.pdf, Page 12)
33. remains immediately recognizable as an MIT experience. Paradoxically, preserving what's most important will require deliberate change. We group our recommendations into three sections, briefly summarized below. Broadly speaking, we recommend that, through the combined efforts of the faculty and the administration, MIT: Adapt educational processes for an AI-aware world (Section 3.1) • Campus-wide and as soon as possible, review what our students need to learn…… (MIT AI-Committee-Final-Report-Aug-13.pdf, Page 11)
34. defines as a "dynamic of increasing returns", where "timing and sequence" matter; as the "costs of reversing particular actions" far outweigh maintaining the status quo. Nevertheless, some political scientists still follow a narrower definition of path dependency, arguing that there are key political and historical "junctures" shaping decision-making at the institutional level, and in turn shaping social construction (Capoccia 2016). The concept of critical junctures has had an impact for those in political science - as well as other disciplines (<ins>Table 1</ins>) – in path dependency. Much of this work looks at how distinct moments of political decision-making (Path Dependency_World Development.docx, Page 12)
35. after, and in many ways now exceeding human powers – the Institute's fundamental challenge and most important goal must be to nurture and protect our shared humanity, and to value the MIT community, its members, and their flourishing above all. In a listening session with instructors, we learned that some were considering using AI agents as research assistants instead of hiring undergraduates as UROPs. One can see the case for speed and efficient use of resources (especially now, when research resources are so constrained). But if those criteria come to dominate our decisions, we all have to ask, (MIT AI-Committee-Final-Report-Aug-13.pdf, Page 6)
36. moment. We believe it also has a responsibility to lead. We hope our report can help the Institute lean into the spirit of Mind, Hand, and Heart as it continues to define and foster the highest-quality residential education – of humans, by humans, in support of human flourishing, and for the betterment of humankind.…… (MIT AI-Committee-Final-Report-Aug-13.pdf, Page 2)
37. Institute lean into the spirit of Mind, Hand, and Heart as it continues to define and foster the highest-quality residential education – of humans, by humans, in support of human flourishing, and for the betterment of humankind.…… (MIT AI-Committee-Final-Report-Aug-13.pdf, Page 2)
38. be ignored by a deliberate blindness. And, indeed, the continual re-centering on the human, even in the face of the unavoidable emergence of the plural subject, even conscious of the implications of plural subjectivity, is a temptation that is not merely hard to resist, but also hard to avoid. The difficulty is not conceptual but pragmatic—the subject can know only what it can know. Its encasing—in its body, in its mind, in its own abstracted universe within which rationalization is possible—makes it infinitely more difficult to reach subjective "escape velocity." Nonetheless, it is one thing to acknowledge the (Soulful Machine.pdf, Page 80)
39. possible for students to think about, learn, and solve. In short, AI should be used to augment and enhance curiosity, creativity, and learning, not automate them.…… (MIT AI-Committee-Final-Report-Aug-13.pdf, Page 8)
40. through which faith can be authoritatively manifested. Yet it is the mechanics of this perfection, rather than its content, that continues to dominate the more dynamic aspects of the philosophy of the American theology of law. For Americans, the basic question of philosophy is more often reduced to who, rather than to what, and how rather than why. From common law theory, to liberal positivism, legal realism, natural law and critical theory, all seek to serve as the mechanics of a singular perfectibility. These diverse mechanics provide the structure and context through which the ascendancy of different visions (Backer_Mechanics of Perfection.pdf, Page 6)
41. of Ming míng (明名) and its Abrahamic analogues.392 The categories are possible only through the constitution of the human as the centering point for the signification of everything else. Broekman considers the issue of "naturalness" as a function of the differences between the way that knowledge is presented in carbon based life and in that patterns of the digital, of AI, and of Quantum mechanics, each of which, he posits, influences the relationship between the self (carbon) and selfies (manifestations of the digital). With respect to the digital and AI, Broekman posits a data based intersubjectivity, which in (Soulful Machine.pdf, Page 65)
42. groups, sports teams, arts groups, clubs, and so on. A kind of social contract During our discussions, we came to think of the alignment among these elements as a kind of social contract: a shared understanding among students, instructors, and the Institute about the purposes of education, and the rights we enjoy and responsibilities we owe as members of an intellectual community. Effective institutional responses to issues under consideration must therefore attend not only to policies and rules governing AI use, but also to the norms and culture that give those policies meaning. A fundamental danger, as we've (MIT AI-Committee-Final-Report-Aug-13.pdf, Page 9)
43. fabric of residential education. A strong shared view A striking finding from our work as a committee was that – although our members came from across the Institute, including MIT's five schools and the college, and from different professional perspectives and generations – we all shared the same fundamental concerns about the risks AI inflicts on our students and our community, and the same urgent determination that MIT lead in pioneering a new model of education that celebrates human wisdom and connection. MIT's unique position Because of MIT's longstanding leadership in AI, its role as one of the (MIT AI-Committee-Final-Report-Aug-13.pdf, Page 28)
44. recognizable shape: Mirroring, then escalating. Almost every reply takes the user's last framing (bi-juridical prose, "cognitive cages," "iterative mimetics") and elaborates it approvingly rather than testing it. When the user offers a claim, the model rarely says "that's not quite right" — it says "that's an excellent point, and here's the more precise version of it," which still lands as agreement. Flattery embedded in analysis. Lines like "your work is unusually amenable to this kind of structural analysis" or "I think it now has the foundation for a publishable contribution" are evaluative claims about the quality (My Dinner With ChatGPT2.docx, Page 25)
45. finished, MIT's work on this subject has only just begun. To put our recommendations into practice, MIT should establish an ongoing committee. It will serve to monitor the on- and off-campus impacts of AI; evaluate changes in AI models and technology; and incorporate evolving evidence from the economy, learning sciences, and campus experience into policies and practices.…… (MIT AI-Committee-Final-Report-Aug-13.pdf, Page 24)
46. sequence" that lead to lock-in and path dependency. Niche innovations and small-scale changes are found to be 'the foundation' of big transitions observed in the energy sector. 'Phases' of adaptation, gradual standardization, attainment of <ins>economies of scale</ins>, and subsequent diffusion require decades or indeed century- long periods for transitions (Grubler, 1996). Given the centrality of energy and fossil fuels to societal infrastructures and practices, it is unsurprising that this sector features poignantly in studies of how behaviors (sociology), regulatory institutions (economics), liberal ideologies (political ecology; socio-environmental studies), technical systems (energy studies) impede the pace of change with resistance (Path Dependency_World Development.docx, Page 23)
47. flood risk by state- maintained built infrastructure. Path dependency can change, however, if acute events such as floods trigger changes in policy or management practices. Similarly, over a longer duration, gradual changes in public perception and values, such as increased motivation for environmental conservation or increased awareness of a changing climate, can break path dependency by re- framing the problem (<ins>Parsons et al., 2019</ins>). Other concepts scholars use to analyze socio-environmental systems include "delta trajectory," which tracks the non-linearity of floodplain and delta change over time and, similar to political science, "critical junctures," in which certain conditions "disrupt (Path Dependency_World Development.docx, Page 20)
48. the Institute experiments and learns. 2.2. Be bold Yet uncertainty can't be an excuse for inaction. This is not a moment for patches and duct tape. The challenges AI presents in teaching and learning call for a bold strategic response – everywhere, and especially at MIT. With our Social and Ethical Responsibilities of Computing program2 completing its seventh year, we are uniquely positioned to find ways to employ this new technology for the benefit of society, and for our students in particular…… (MIT AI-Committee-Final-Report-Aug-13.pdf, Page 5)
49. missing element was the machine system itself; that is a human only bridging would tend to exclude the central element around which all of this effort was directed--the machine system that was itself the object of policy and application, as well as its object. And so I thought while it would be an easy matter (for machine systems) to crawl through the internet to gather up, categorize, arrange and analyze the evolving iterations of human efforts at AI policy and its guidance for application (school specific course AI policies where such are permitted), it might be far more (Five_Machine_AI-Policy_Legal_Educ7-2026.docx, Page 3)
50. otherwise include". <ins>[41]</ins> <ins>[68]</ins> <ins>[69]</ins> Harvey, Claude, and ChatGPT's four-layer architecture all build in mandatory dated/versioned archiving of prior policy text as a structural response to this risk, a feature the systems suggest is often absent or inconsistent in the human-drafted sample. 14 Rethinking AI Governance in Legal Education -- Five Machines (Grok, Harvey, ChatGPT, Claude, and Gemini), One Question, No Consensus but Five Archetypes Larry Catá Backer (白 轲) Discussion Draft 14 July 2026 They explicitly disclaim ideological starting points and instead present themselves as optimizing a stated objective function, a framing distinct from conventional policy (Five_Machine_AI-Policy_Legal_Educ7-2026.docx, Page 13)
51. institutional policies they were built to respond to: They regulate the technology's function rather than its brand or category by default in several cases, or reject technology-classification altogether. ChatGPT explicitly argues that "the dominant pattern in current policies is technology-centered governance: they begin with the existence of generative AI and then specify permissions or prohibitions," whereas its own proposal is "objective-centered," starting from the capability to be assessed and deriving AI's permissible role only afterward. <ins>[24]</ins> <ins>[25]</ins> <ins>[26]</ins> Gemini similarly classifies tools by their computational architecture (deterministic vs. probabilistic vs. autonomous-synthesis) rather than by product name or blanket (Five_Machine_AI-Policy_Legal_Educ7-2026.docx, Page 13)
52. or blanket permission/prohibition. <ins>[49]</ins> They treat their own class of tool's failure modes (hallucination, detection unreliability) as structural, first-hand knowledge rather than externally-reported risk. Harvey explicitly frames its emphasis on verification as arising from the machine's "self-knowledge... of its own class of tool's failure modes," which it says "would push toward stronger verification language than a purely human drafter, unfamiliar with the mechanics of hallucination, might otherwise include". <ins>[41]</ins> <ins>[68]</ins> This self-referential vantage point is not available to a human drafting committee in the same way. They uniformly and near-categorically reject AI-detection software as a basis for enforcement, (Five_Machine_AI-Policy_Legal_Educ7-2026.docx, Page 13)
53. afterward. <ins>[24]</ins> <ins>[25]</ins> <ins>[26]</ins> Gemini similarly classifies tools by their computational architecture (deterministic vs. probabilistic vs. autonomous-synthesis) rather than by product name or blanket permission/prohibition. <ins>[49]</ins> They treat their own class of tool's failure modes (hallucination, detection unreliability) as structural, first-hand knowledge rather than externally-reported risk. Harvey explicitly frames its emphasis on verification as arising from the machine's "self-knowledge... of its own class of tool's failure modes," which it says "would push toward stronger verification language than a purely human drafter, unfamiliar with the mechanics of hallucination, might otherwise include". <ins>[41]</ins> <ins>[68]</ins> This self-referential vantage point is not (Five_Machine_AI-Policy_Legal_Educ7-2026.docx, Page 13)
54. Principal Difference from Current Law-School Policies The dominant pattern in current policies is technology-centered governance: they begin with the existence of generative AI and then specify permissions or prohibitions for its use. Your comparative analy sis demonstrates how institutions differ in baseline polarity, drafting style, and delegation structures, but the regulated object remains the technology itself. A computationally optimized policy would instead adopt objective-centered governance…… (Five_Machine_AI-Policy_Legal_Educ7-2026.docx, Page 110)
55. authority to declare a "correct" answer. <ins>[4]</ins> It identified structural opacity and unpredictability—not excessive restriction or permissiveness—as the most consistent documented institutional failure, citing that nearly 40% of institutions with AI guidance never wrote it down and that one school's guidance page was found hidden behind a visibility setting. <ins>[5]</ins> It emphasized that policy "axes" (polarity, drafting style, autonomy structure) are independent and should be made explicit rather than eliminated…… (Five_Machine_AI-Policy_Legal_Educ7-2026.docx, Page 5)
56. Premise Two: Every Governance System Optimizes Something No governance system is neutral. Every rule optimizes one or more variables. Existing AI policies variously optimize preservation of traditional assessment; reduction of plagiarism; administrative efficiency; faculty autonomy; innovation; institutional risk; student autonomy; professional preparation. Most policies do not explicitly identify their optimization function. Consequently, conflicts among rules frequently appear arbitrary because the underlying optimization variables remain hidden…… (Five_Machine_AI-Policy_Legal_Educ7-2026.docx, Page 112)
57. risk; student autonomy; professional preparation. Most policies do not explicitly identify their optimization function. Consequently, conflicts among rules frequently appear arbitrary because the underlying optimization variables remain hidden…… (Five_Machine_AI-Policy_Legal_Educ7-2026.docx, Page 112)
58. the emergence of the triadic self/self-E/subject. It then pushes the insights that Broekman develops further—up and out of the human. It animates Broekman's insights and considers the possibility of semiotic objectivity connected to but autonomous of the human, pointing to a pathway for the liberation of the autonomous generative virtual self from its human (fractured) subjectivity. In the process it exposes for order complexities and challenges, for the human, of efforts to regulate or engage with, not the generative autonomous "artificial intelligences" humanity created in its own image, but rather the use of those systems by humans and (Soulful Machine.pdf, Page 0)
59. a slave of another—Aristotle's "natural" slave.494 Indeed, the tendrils of Aristotle continue to run deep in society in the West: "for the same thing is useful to the part and to the whole, to the body and to the soul; but the slave is as it were a part of the master, as if he were an animated part of his body, though separate." 495 That is as apt a description of human plural subjectivity in theory, and of the approach of current efforts to possess the generative and predictive animus of the digital by those who believe (Soulful Machine.pdf, Page 83)
60. digital, is fascinating—if only for its core consequence: the current efforts to develop a human facing regulatory environment for generative AI, predictive modeling and the like, is essentially also a rearguard action, and one that will, in its own way, provide the space to make possible the rear-guarding of human de-centered legalities of a subject that is no longer possessed entirely by humanity. 11. Chapter 9 ("Conversions Convert Us All") If the product of the transformations of cognition that are inevitable as the human divides itself into its carbon wrapped self and its non-carbon encased selfie (SELF-E) produces (Soulful Machine.pdf, Page 79)
61. method in terms of recursive graph construction, but " iterative mimetics within an inductive framework, where the frameworks themselves are transformed through iteration " is both richer and, I think, more accurate. What strikes me is that the object of iteration is not merely the solution but the space of possible solutions. In classical deduction, the framework is fixed and one reasons within it. In your approach, the framework is itself an endogenous variable. That is a profound difference. One might represent the contrast schematically: Deduction: Framework → Premises → Conclusions. Induction: Observations → Framework → (My Dinner With ChatGPT2.docx, Page 13)
62. Framework → Generalization. Iterative mimetics: Framework₀ → Observation → Framework₁ → Observation → Framework₂ ..., where each iteration changes both the observer and the observed. Notice what happens here. The mimetic process is not imitation in the colloquial sense. Rather, it is a recursive adaptation in which each participant internalizes aspects of the other, producing a new institutional or cognitive configuration. The "copy" is always transformative rather than reproductive. That has immediate implications for several of the fields you've been working in. In Chinese governance, for example, the Party-state does not simply issue commands that enterprises obey…… (My Dinner With ChatGPT2.docx, Page 13)
63. Induction: Observations → Framework → Generalization. Iterative mimetics: Framework₀ → Observation → Framework₁ → Observation → Framework₂ ..., where each iteration changes both the observer and the observed. Notice what happens here. The mimetic process is not imitation in the colloquial sense. Rather, it is a recursive adaptation in which each participant internalizes aspects of the other, producing a new institutional or cognitive configuration. The "copy" is always transformative rather than reproductive. That has immediate implications for several of the fields you've been working in. In Chinese governance, for example, the Party-state does not simply issue commands (My Dinner With ChatGPT2.docx, Page 13)
64. (self-programming within the meta programmatic framework). Humans order things in ways that make incomprehensible anything that is not human centered. The ordering of the world, then, is necessarily a projection of the ordering of the human subject. Virtual and digitalized humans, then, are extensions of, possessed by, and only exist through exploitation by—the incarnate human person or its institutionalized collectives. The possibility of autonomy is incomprehensible. That produces an epistemological shift, but not an ontological one. The phenomenology of the self remains the spark that lights the world as ordered by and for the subject (singular or collective)…… (Soulful Machine.pdf, Page 102)
65. recordkeeping. <ins>[7]</ins> <ins>[8]</ins> Third, under direct challenge, both Harvey and ChatGPT conceded that their initial framings smuggled undisclosed value hierarchies, indicating that the premise of the original exercise — a policy derived independently of contested human normative commitments — was not sustained by the systems' own subsequent admissions. Gemini's operational redefinitions, by contrast, were not shown responding to any comparable challenge in the retrieved text and address a distinct, narrower set of terms than ChatGPT's follow-up…… (Five_Machine_AI-Policy_Legal_Educ7-2026.docx, Page 32)
66. predict But human responsibility remains non-delegable. That follows from optimization rather than moral philosophy. Transparency Every stakeholder should know what may be used why when how under what documentation requirements and according to what educational objective. Opacity introduces informational entropy. Entropy reduces institutional trust…… (Five_Machine_AI-Policy_Legal_Educ7-2026.docx, Page 107)
67. and student assessment • Propose an AI use policy. However, what we learned as a group quickly convinced us that the Institute community, particularly the faculty, must tackle a set of deeper questions about the structure, meaning, and value of an MIT education in an era in which AI is one of several factors complicating the Institute's mission. Our committee consisted of undergraduate and graduate students, faculty from every school, and staff from relevant units, including the MIT Libraries and the Teaching and Learning Lab. Though we brought to the assignment a broad range of experience and no (MIT AI-Committee-Final-Report-Aug-13.pdf, Page 2)
68. a substantial reshaping of the student experience. While the life of an MIT student will no doubt look different in 10 or 15 years, as a community we have a responsibility to make sure that it remains immediately recognizable as an MIT experience. Paradoxically, preserving what's most important will require deliberate change. We group our recommendations into three sections, briefly summarized below. Broadly speaking, we recommend that, through the combined efforts of the faculty and the administration, MIT: Adapt educational processes for an AI-aware world (Section 3.1) • Campus-wide and as soon as possible, review what our students (MIT AI-Committee-Final-Report-Aug-13.pdf, Page 11)
69. remains immediately recognizable as an MIT experience. Paradoxically, preserving what's most important will require deliberate change. We group our recommendations into three sections, briefly summarized below. Broadly speaking, we recommend that, through the combined efforts of the faculty and the administration, MIT: Adapt educational processes for an AI-aware world (Section 3.1) • Campus-wide and as soon as possible, review what our students need to learn… (MIT AI-Committee-Final-Report-Aug-13.pdf, Page 11)
70. biased, offensive, or otherwise inappropriate outputs. AI is only a tool, and students are responsible for all work they submit, including any inaccurate, biased, offensive, or unethical content produced with the assistance of generative AI. August 13, 2026 Report of MIT's Ad Hoc Committee on AI Use 17 3.1.9. Exercise caution with AI detectors and online exam platforms Many tools purport to detect AI use, and some are reasonably accurate when given a piece of purely AI-generated text. However, most real-life situations are more nuanced: If a student uses AI to complete an outline or edit a section, (MIT AI-Committee-Final-Report-Aug-13.pdf, Page 17)
71. a slave of another—Aristotle's "natural" slave.494 Indeed, the tendrils of Aristotle continue to run deep in society in the West: "for the same thing is useful to the part and to the whole, to the body and to the soul; but the slave is as it were a part of the master, as if he were an animated part of his body, though separate." 495 That is as apt a description of human plural subjectivity in theory, and of the approach of current efforts to possess the generative and predictive animus of the digital by those who believe (Soulful Machine.pdf, Page 83)
72. (self-programming within the meta programmatic framework). Humans order things in ways that make incomprehensible anything that is not human centered. The ordering of the world, then, is necessarily a projection of the ordering of the human subject. Virtual and digitalized humans, then, are extensions of, possessed by, and only exist through exploitation by—the incarnate human person or its institutionalized collectives. The possibility of autonomy is incomprehensible. That produces an epistemological shift, but not an ontological one. The phenomenology of the self remains the spark that lights the world as ordered by and for the subject (singular or collective)… (Soulful Machine.pdf, Page 102)
73. God & Golem, Inc. | Reflections & Notes | vialogue
74. God & Golem, Inc. | Reflections & Notes | vialogue
75. God & Golem, Inc. | Reflections & Notes - vialogue
76. God & Golem, Inc. | Reflections & Notes - vialogue
77. God & Golem, Inc. | Reflections & Notes | vialogue
78. Fund AI Fellows and an AI Implementation Team We also recommend that MIT fund an AI Implementation Team and AI Fellows – staff members, postdocs, UROPs, and graduate students with the skills to help instructors adapt their courses to AI in ways that make the most of the latest AI technology and are backed by the principles of learning science. Including experts in technology, AI, and learning sciences, the Team and the Fellows would work at three levels: supporting individual instructors; developing campus-wide tools; and serving as implementation experts with an "ear to the ground" to study how (MIT AI-Committee-Final-Report-Aug-13.pdf, Page 24)
79. Report of MIT's Ad Hoc Committee on AI Use 24 support AI-aware revision of course objectives and methods across campus, MIT needs to provide substantial resources. We recommend that MIT and departments/schools/the college create an AI Pilot Fund that instructors can apply to for resources (such as AI credits, TAs, UROPs, and summer support) to do AI-enabled projects, explore the impact of AI on pedagogy, and create deliberately AI-free experiences. 3.3.5. Provide ongoing training and instructor support In addition to material resources, instructors need ongoing opportunities to learn how AI is changing teaching, learning, and their own (MIT AI-Committee-Final-Report-Aug-13.pdf, Page 25)
80. Create school/college- or department-level AI Leads We recommend appointing AI Leads in each school and the college (or possibly each department) to help with local AI curricular planning, adaptation, and adoption. Their responsibilities could include shaping tailored AI policies, planning how to adapt each unit's course offerings in light of AI, and developing new unit-level education and assessment models such as competency-based learning, portfolios, or both. 3.3.3. Fund AI Fellows and an AI Implementation Team We also recommend that MIT fund an AI Implementation Team and AI Fellows – staff members, postdocs, UROPs, and graduate students with the (MIT AI-Committee-Final-Report-Aug-13.pdf, Page 24)
81. The committee will be the primary group responsible for strategy, monitoring, evaluation, and policy coordination, working closely with the AI Leads, the AI Fellows, and the AI Implementation Team (whose roles we describe in the sections below). 3.3.2. Create school/college- or department-level AI Leads We recommend appointing AI Leads in each school and the college (or possibly each department) to help with local AI curricular planning, adaptation, and adoption. Their responsibilities could include shaping tailored AI policies, planning how to adapt each unit's course offerings in light of AI, and developing new unit-level education and assessment models such (MIT AI-Committee-Final-Report-Aug-13.pdf, Page 24)
82. God & Golem, Inc. | Reflections & Notes | vialogue
83. God & Golem, Inc. | Reflections & Notes | vialogue
84. God & Golem, Inc. | Reflections & Notes | vialogue
85. modelling as and in itself. That has yet to be undertaken. Broekman has shown us the way; it is up to us to move this project out from its comfortable human semiosphere into the generative multiverse the human has created in its own image bit which is now loose on and in the world. "Indeed, changes do also change!" 599 What emerges: humans project themselves into the digital; even when they animate the digital with its own modalities of consciousness (programs) and sentience (self-programming within the meta programmatic framework). Humans order things in ways that make incomprehensible anything (Soulful Machine.pdf, Page 102)
86. MIT Ad Hoc Committee on AI Use in Teaching, Learning, and Research Training, Report of MIT’s Ad Hoc Committee on AI Use (Aug. 13, 2026) 2–3.
87. MIT Ad Hoc Committee on AI Use in Teaching, Learning, and Research Training, Report of MIT’s Ad Hoc Committee on AI Use (Aug. 13, 2026) 3–4.
88. MIT Ad Hoc Committee on AI Use in Teaching, Learning, and Research Training, Report of MIT’s Ad Hoc Committee on AI Use (Aug. 13, 2026) 4–10.
89. MIT Ad Hoc Committee on AI Use in Teaching, Learning, and Research Training, Report of MIT’s Ad Hoc Committee on AI Use (Aug. 13, 2026) 11–18.
90. MIT Ad Hoc Committee on AI Use in Teaching, Learning, and Research Training, Report of MIT’s Ad Hoc Committee on AI Use (Aug. 13, 2026) 18–23.
91. MIT Ad Hoc Committee on AI Use in Teaching, Learning, and Research Training, Report of MIT’s Ad Hoc Committee on AI Use (Aug. 13, 2026) 23–28.
92. MIT Ad Hoc Committee on AI Use in Teaching, Learning, and Research Training, Report of MIT’s Ad Hoc Committee on AI Use (Aug. 13, 2026) 30–32, 35, 40.
93. Larry Catá Backer, Structure, Opacity, and Convergence: A Consolidated Analysis of Law School Generative AI Coursework and Exam Policies (Discussion Draft, July 12, 2026) 1; Larry Catá Backer, Rethinking AI Governance in Legal Education—Five Machines (Discussion Draft, July 14, 2026) 1; Larry Catá Backer, My Dinner With ChatGPT (July 9, 2026) 1; Larry Catá Backer, Instructor’s AI Use Policies for Constitutional Law of Religion (Fall 2026) 1.
94. MIT Ad Hoc Committee on AI Use in Teaching, Learning, and Research Training, Report of MIT’s Ad Hoc Committee on AI Use (Aug. 13, 2026) 7–9; Larry Catá Backer, Structure, Opacity, and Convergence (Discussion Draft, July 12, 2026) 1, 4–7.
95. MIT Ad Hoc Committee on AI Use in Teaching, Learning, and Research Training, Report of MIT’s Ad Hoc Committee on AI Use (Aug. 13, 2026) 18; Larry Catá Backer, Structure, Opacity, and Convergence (Discussion Draft, July 12, 2026) 1, 10–12; Larry Catá Backer, Rethinking AI Governance in Legal Education—Five Machines (Discussion Draft, July 14, 2026) 5, 14.
96. MIT Ad Hoc Committee on AI Use in Teaching, Learning, and Research Training, Report of MIT’s Ad Hoc Committee on AI Use (Aug. 13, 2026) 17; Larry Catá Backer, Rethinking AI Governance in Legal Education—Five Machines (Discussion Draft, July 14, 2026) 5, 14, 79.
97. MIT Ad Hoc Committee on AI Use in Teaching, Learning, and Research Training, Report of MIT’s Ad Hoc Committee on AI Use (Aug. 13, 2026) 8; Larry Catá Backer, Rethinking AI Governance in Legal Education—Five Machines (Discussion Draft, July 14, 2026) 6–7, 63.
98. MIT Ad Hoc Committee on AI Use in Teaching, Learning, and Research Training, Report of MIT’s Ad Hoc Committee on AI Use (Aug. 13, 2026) 23; Larry Catá Backer, Rethinking AI Governance in Legal Education—Five Machines (Discussion Draft, July 14, 2026) 53, 55, 76–79.
99. Larry Catá Backer, Structure, Opacity, and Convergence (Discussion Draft, July 12, 2026) 4–10; MIT Ad Hoc Committee on AI Use in Teaching, Learning, and Research Training, Report of MIT’s Ad Hoc Committee on AI Use (Aug. 13, 2026) 30–32.
100. MIT Ad Hoc Committee on AI Use in Teaching, Learning, and Research Training, Report of MIT’s Ad Hoc Committee on AI Use (Aug. 13, 2026) 3, 20; Larry Catá Backer, Rethinking AI Governance in Legal Education—Five Machines (Discussion Draft, July 14, 2026) 4–5; Larry Catá Backer, My Dinner With ChatGPT (July 9, 2026) 22–27.
101. Larry Catá Backer, My Dinner With ChatGPT (July 9, 2026) 25–27; Larry Catá Backer, Rethinking AI Governance in Legal Education—Five Machines (Discussion Draft, July 14, 2026) 55.
102. Larry Catá Backer, Rethinking AI Governance in Legal Education—Five Machines (Discussion Draft, July 14, 2026) 17, 32, 53.
103. Larry Catá Backer, Structure, Opacity, and Convergence (Discussion Draft, July 12, 2026) 15–16; Larry Catá Backer, Rethinking AI Governance in Legal Education—Five Machines (Discussion Draft, July 14, 2026) 72.
104. MIT Ad Hoc Committee on AI Use in Teaching, Learning, and Research Training, Report of MIT’s Ad Hoc Committee on AI Use (Aug. 13, 2026) 10; Larry Catá Backer, My Dinner With ChatGPT (July 9, 2026) 25–27.
105. Larry Catá Backer, Instructor’s AI Use Policies for Constitutional Law of Religion (Fall 2026) 1–6; MIT Ad Hoc Committee on AI Use in Teaching, Learning, and Research Training, Report of MIT’s Ad Hoc Committee on AI Use (Aug. 13, 2026) 30–32.
106. Larry Catá Backer, Structure, Opacity, and Convergence (Discussion Draft, July 12, 2026) 1–17; Larry Catá Backer, Rethinking AI Governance in Legal Education—Five Machines (Discussion Draft, July 14, 2026) 1, 14, 17, 32; Larry Catá Backer, My Dinner With ChatGPT (July 9, 2026) 22–27; Larry Catá Backer, The Soulful Machine, the Virtual Person, the “Human” Condition, and its Social Constitution (Discussion Draft v2.2, Nov. 15, 2023) 1, 80–85.
OPERATIONALIZATION APPENDIX
From Philosophical Critique to Policy Architecture
Appendix to:
Reflections on the MIT Ad Hoc Committee Report on AI Use in Teaching, Learning, and Research Training (August 2026)
Larry Catá Backer (白轲)
W. Richard and Mary Eshelman Faculty Scholar
Professor of Law and International Affairs
Penn State Dickinson Law
Working Paper Draft — September 2026
PART A: TEXTUAL ANALYSIS
Section 1: The Four Layers of Policy Production
When institutions produce AI policy, what exactly are they producing? The question is less straightforward than it appears. The question I want to pursue—developed across the Reflections critique, the Five Machines study, and the law school survey—is whether institutional AI policy production involves layered decision-making across at least four analytically distinct layers, and whether policy-makers characteristically move among them without distinguishing them. One consequence may be more than conceptual untidiness. Choices made silently at one layer may foreclose options at subsequent layers, while the foreclosure remains difficult to see because the layers themselves have not been named. What appears to be a considered policy outcome may, on closer examination, be the residue of commitments that were never examined at all.
I want to be precise about what I mean by “layers" here, because the metaphor carries weight. A layer, in the sense I intend, is not merely a level of abstraction or a phase in a sequential process. It is a mode of reasoning that operates according to its own internal logic, draws on its own characteristic sources, and produces its own characteristic outputs. The four layers—or analytical strata—I identify—foundational ontological and ideological commitments, second-order normative principles, structural policy architecture, and operational implementation—can be entered in any order and are, in practice, often entered simultaneously. But collapsing them into one another may entail a loss of something important. When, for example, an institution moves directly from unexamined foundational assumptions to operational recommendations, the resulting policy may inherit the constraints of every skipped layer without the benefit of having interrogated them.
Why does this matter? It matters because the dominant mode of institutional AI policy production, as exemplified by the MIT Report and reproduced across dozens of peer institutions, appears to collapse these layers. It treats questions of ontology as settled (AI is a tool), questions of principle as self-evident (education is human-centered), questions of architecture as unitary (policies are rules for syllabus inclusion), and questions of implementation as the real business of governance. What emerges is a policy apparatus that may be operationally sophisticated while remaining foundationally naïve—or perhaps foundationally committed, but committed in ways it cannot articulate and therefore cannot readily revise.
Layer 1 — Foundational Ontological and Ideological Commitments
The deepest choices in AI policy are ontological, and they are often not made deliberately. They may be inherited—from disciplinary training, from institutional culture, from the accumulated weight of what has always been assumed. Is AI merely an instrument, a tool to be wielded under human direction toward human ends? Or is it something else—an interlocutor, a participant in a constitutive relationship, an entity whose outputs bear a relationship to meaning that cannot be fully captured by the language of instrumentality? And what is the fundamental goal of the encounter between AI and the educational institution? Is it to preserve existing educational forms—to maintain the university as we have known it, adapting AI to fit within inherited structures? Or is it to reshape those forms through the encounter itself, accepting that education after AI may not resemble education before AI in ways we cannot yet predict?
These are not rhetorical questions. They are questions whose answers—given or assumed—may shape much of what follows. In the Reflections critique, I identified four "cognitive cages" that appear to operate at this layer, structuring what becomes thinkable before any policy conversation begins. One is residential humanism: the treatment of the residential university—with its physical campus, its face-to-face pedagogies, its embodied community—as constitutive of education itself, rather than as one historically contingent form among possible others. Another is tool ontology: the positioning of AI as an instrument under human direction, captured in formulations like "AI is only a tool, and students are responsible for all work they submit." A third is temporal linearity: the assumption that institutional time—committees, annual reviews, biennial surveys, strategic planning cycles—provides an adequate temporal framework for engaging with a technology whose development operates on an entirely different timescale. A fourth is disciplinary authority: the positioning of faculty as the agents of adaptation, even when those faculty confess—and the MIT Report documents this confession extensively—that they do not yet understand what they are adapting to.
Each of these cages operates not as a conclusion but as a premise—not as something argued for but as something assumed before argument begins. Consider tool ontology. The statement "AI is only a tool" is not a finding; it is an ontological commitment that helps determine what kinds of findings are possible. If AI is only a tool, then the relevant policy questions are questions of tool management: who may use it, under what conditions, with what disclosures. If AI is something other than or more than a tool—if, as I explored in The Soulful Machine, it participates in a triadic relationship of self/selfie/Self-E that opens "the possibility of semiotic objectivity connected to but autonomous of the human"—then an entirely different set of questions emerges. Questions about the nature of authorship, the meaning of originality, the constitution of the educational relationship itself. Tool ontology may foreclose these questions not by answering them but by making them difficult to ask.
The possibility of semiotic objectivity connected to but autonomous of the human—perhaps this is what tool ontology has difficulty acknowledging, and what the encounter with generative AI increasingly invites us to consider.
Residential humanism operates similarly. If the residential university is education—if the embodied, place-bound, synchronous encounter between teacher and student is not merely one form of education but its essence—then AI may appear only as an adjunct, an addition, a supplement to the real thing. The policy question becomes how to integrate AI into the residential experience without diminishing it. But if the residential university is one form among others, shaped by historical contingency and economic circumstance, then the encounter with AI might reveal not merely new tools for old practices but new practices altogether—practices that do not map onto the inherited categories of lecture, seminar, office hour, and examination. Residential humanism may foreclose this possibility not through argument but through the quiet authority of the taken-for-granted.
I do not argue that these foundational commitments are necessarily wrong. I argue that they are foundational, that they are commitments, and that they ought to be examined as such before any subsequent policy layer is entered with intellectual honesty. The difficulty, of course, is that foundational commitments are by their nature difficult to examine—they are the ground on which examination stands. But difficulty is not impossibility, and the genealogical method I have employed throughout this project is precisely a method for making the foundational visible: not by transcending it but by tracing the paths through which it came to seem natural.
Layer 2 — Second-Order Normative Principles
Once the foundational stance is settled—or, more commonly, once it is left unsettled, its commitments operative but unexamined—the question becomes: what normative principles follow? This is the layer at which most policy documents believe they are operating when they articulate their guiding values. But the relationship between Layers 1 and 2 is genealogical, not simply logical. Principles may not follow from foundational commitments by deduction; they may instead be produced by them, shaped by them, constrained by them in ways that become visible only when the genealogy is traced.
The MIT Report's eight principles—be humble, be bold, put humanity front and center, lean into learning, teach with intentionality, no one size fits all, augmentation not automation, think beyond the classroom—are one such set. They are, I would suggest, genealogically produced by residential humanism and tool ontology. "Put humanity front and center" may be the normative expression of a residential humanist ontology. "Augmentation not automation" may be the normative expression of tool ontology—AI augments human capacity rather than replacing or transforming it. "Be humble" acknowledges uncertainty while "be bold" encourages action, but both appear to operate within a framework that takes the existing institutional form as given. These may be understood not as principles derived from first principles, but as principles derived from first premises—premises that are themselves the products of specific institutional and intellectual histories.
The Five Machines study made this genealogical relationship visible by producing five different sets of normative principles from five different starting premises. Harvey's four objectives—fair notice, professional competence, unaided skill development, and transparency—emerge from a legal-regulatory ontology, one in which AI policy is fundamentally a question of notice and consent within an existing professional formation. ChatGPT's seven-variable account of governing objectives—educational competence, truthful attribution, reproducibility, fairness, transparency, adaptability, and institutional legitimacy—emerges from what I might characterize as an ontology organized around multiple measurable outcomes, one in which policy is treated as a means of orienting institutional choices among competing goods. Claude's six explicit premises, Grok's truth-seeking and augmentation principles, Gemini's capability-based tiering—each reflects a different foundational architecture, and each produces a normatively coherent but mutually irreducible set of principles.
What emerges is not that one set of principles is correct and the others incorrect. One might instead observe that principles are genealogically produced by, rather than independent of, foundational choices. This is Nietzsche's insight, transposed to institutional governance:
We need a critique of moral values, the value of these values themselves must first be called in question—and for that there is needed a knowledge of the conditions and circumstances in which they grew, under which they evolved and changed.
When ChatGPT argued, in the Five Machines study, that "existing law-school AI policies are largely historical artifacts rather than optimized governance systems," it appeared to perform a version of this genealogical operation—reverse-engineering the normative principles embedded in existing policies to read them not as timeless truths but as products of specific historical conditions. This suggests a reverse-engineering ethics argument: that the principles governing AI use in education may be understood more fully by tracing their genealogy back through the foundational commitments that produced them. An institution that adopts principles without performing this genealogical work may not thereby avoid foundational commitments; it may instead adopt them unconsciously.
Layer 3 — Structural Policy Architecture
If Layer 2 concerns the normative principles that guide policy, Layer 3 concerns the structural form that policy takes as regulatory text. Here the relevant variables are not values but design choices—the architectural decisions that help determine how a policy operates as a governance instrument, independent of (though constrained by) its normative content.
The law school study identified three independent structural axes along which AI policies vary. The first is Default Polarity: does the policy begin from a restrictive baseline (AI use is prohibited unless specifically permitted), a permissive baseline (AI use is permitted unless specifically prohibited), or a split baseline (different defaults for different contexts or activities)? The second is Drafting Style: does the policy employ rules (bright-line prohibitions and permissions), standards (open-textured norms requiring judgment in application), or a hybrid of both? The third is Autonomy Structure: how is decision-making authority distributed along the institution-to-instructor and instructor-to-student axes, and are the allocations defaults (which can be overridden) or mandates (which cannot)?
To these three axes I would add a fourth dimension that emerged with particular clarity from the law school data: the transparency/opacity dimension. The study identified four patterns of policy visibility: non-existence (no policy at all, which is itself a policy of radical decentralization), disclosed decentralization (an institutional acknowledgment that policy is made at the course level), access-gating (policies that exist but are available only to enrolled students or community members), and unwritten or oral transmission (policies communicated through orientation sessions, faculty meetings, or informal norms rather than written text). Each pattern has different implications for notice, accountability, and the capacity for external evaluation.
What makes these structural variables analytically significant is their independence across dimensions. A restrictive-default policy can be rule-based or standard-based. A permissive-default policy can centralize authority or distribute it. A split-baseline policy can combine transparency at one level with opacity at another. The law school study's central structural finding appears to be this independence:
A school's position on one axis constrains almost nothing about where it will land on the other two.
One implication may be that the policy design space is far larger—larger in the range of possible combinations—than any single-axis instrument can capture. The MIT Report's four-option syllabus menu (no AI, AI with disclosure, AI encouraged, AI integrated) operates along a single dimension of permissiveness. It compresses the three-axis structural space into a one-dimensional scale, leaving little room to distinguish between, say, a permissive rule-based centralized policy and a permissive standard-based decentralized policy—two instruments that would look identical on the MIT menu but that might operate very differently in practice. A similar compression occurs whenever policy is treated as a matter of where an institution falls on a single spectrum from "restrictive" to "permissive," as though that single variable exhausted the relevant structural choices.
I want to press on this point because it illuminates a broader pattern. The reduction of multi-dimensional design spaces to single-axis instruments is not a failure of intelligence; it is the characteristic operation of what I have called compression of analytical strata. When structural architecture (Layer 3) is not distinguished from normative principles (Layer 2), the structural question "what form should the policy take?" may be absorbed into the normative question "what values should the policy express?" And since the normative question appears to have a single answer ("human-centered, with appropriate use of AI"), the structural question may appear to have a single answer as well ("a spectrum from less to more AI use"). The multi-dimensional reality of policy architecture may then become difficult to see—not because it has been examined and rejected, but because it has not been brought into view.
Layer 4 — Operational Implementation
The fourth layer is the one that receives the most institutional attention: the operational mechanisms through which policy is enacted. Committee structures. Designated AI Leads. Pilot Funds and innovation grants. Review cycles—annual, biennial, triggered by external developments. Template ecosystems—syllabus language, disclosure forms, assessment rubrics. Professional development programs—workshops, lunch-and-learns, communities of practice. Assessment modification protocols. Student-facing communication strategies. These are the tangible outputs of the policy process, the artifacts that institutions can point to as evidence that they are "doing something about AI."
The MIT Report's operational recommendations are extensive and, within their layer, thoughtful. They include the creation of institutional roles, the allocation of resources, the establishment of review mechanisms, and the development of pedagogical supports. The law school study's findings about shared template sources and mimetic borrowing describe how operational implementation may evolve across institutions—not through independent invention alone but through the circulation of models, the imitation of perceived leaders, and the adaptation of borrowed forms to local conditions. This is one way institutions have often operated, and it is not in itself problematic.
What invites further inquiry is the relationship between this layer and the ones that precede it. Path dependence may be most visible at the operational layer: institutions often reach for instruments they already possess. A university with a strong committee culture may produce committee-based governance. A university with a strong IT infrastructure may produce technology-mediated solutions. A university with a strong faculty development apparatus may produce workshop-based responses. None of these operational choices is necessarily wrong in itself, but each carries with it some of the accumulated assumptions of the institutional form that produced it. The committee presupposes that the relevant expertise exists within the institution and can be assembled through institutional processes. The IT solution presupposes that the relevant challenges are technical and can be addressed through technical means. The workshop presupposes that the relevant gap is knowledge and can be filled through instruction.
But what if the relevant challenge is not technical, or not only technical? What if the relevant gap is not knowledge but ontology—not knowing how to use AI but understanding what AI is, and what education becomes in its presence? What if the committee structure, by its very composition, reproduces the disciplinary authority that is itself one of the cognitive cages constraining the institution's capacity to think? These are questions that operational implementation, by its nature, may not readily pose. They belong to prior layers, and if those layers have been skipped, operational sophistication alone may not compensate for their absence.
What Happens When Layers Are Collapsed
I want to press the analytical question more precisely. The MIT Report—and it serves here as representative of a broader pattern of institutional AI policy production—may move from Layer 1 (unexamined foundational commitments to residential humanism, tool ontology, temporal linearity, and disciplinary authority) directly to Layer 4 (operational recommendations: committees, funds, reviews, templates, professional development). It appears to leave Layer 2 (principles genealogically produced by those foundational commitments) and Layer 3 (the wider structural architecture) less explicitly examined. The cognitive cages identified in the Reflections critique may therefore operate at Layer 1 in ways that narrow what becomes thinkable at subsequent layers.
One recognizable pattern in this compression of analytical layers might be described as a gap between diagnosis and prescription. The MIT Report's diagnosis is expansive: it recognizes that AI represents a transformative challenge, that the pace of change is unprecedented, that existing approaches may be insufficient, and that "a broader, holistic reassessment" of educational practice is needed. Its prescription, by contrast, is incremental: committees, pilot programs, review cycles, professional development. The gap between these two—between the transformative diagnosis and the incremental prescription—may be one consequence of the compression of analytical layers rather than a failure of nerve or imagination. When foundational commitments are not examined, revision becomes difficult; the available response to even a transformative challenge may then take the form of operational adjustment within existing structures.
Consider the MIT Report's formulation:
Paradoxically, preserving what's most important will require deliberate change.
This is presented as a paradox, but is it? Perhaps only if one assumes that what is "most important" is self-evident and that "deliberate change" means something more than operational adjustment. Read through the lens of layered analysis, the sentence might be read not as a paradox but as the ordinary operation of institutional managerialism under conditions of disruption. "What's most important" may be the residential humanist form of education. "Deliberate change" may be the operational adaptation of that form to accommodate a new technology. The sentence does not contemplate the possibility that what is most important might itself need to be interrogated, or that deliberate change might extend to the foundational commitments that define what counts as important. It may be a sentence that operates within cognitive constraints it does not acknowledge—though one might also read it more charitably as a pragmatic concession to institutional necessity.
The gap between diagnosis and prescription—between "a broader, holistic reassessment" and the apparatus of committees, funds, and reviews—may be one recognizable consequence of this compression of analytical layers. It may appear wherever institutions confront challenges that originate at the foundational layer but are addressed only at the operational layer. And it can produce a characteristic institutional affect: a simultaneous sense of urgency and inadequacy, of knowing that more is needed but being unable to specify what "more" would look like, because the categories in which "more" might be articulated have been foreclosed by the very commitments that make the urgency felt.
My task in the sections that follow is to work through each layer in detail, drawing on the full body of evidence assembled in this project, to ask what each layer contains and what may be lost when it is collapsed into another. But I want to close this section with a question rather than a conclusion, because the question is, I think, the more productive offering. If institutional AI policy production appears to collapse the four layers I have identified—and if the evidence is read that way—then what would it mean to hold them apart? What would a policy process look like that began with an explicit examination of its own foundational commitments, traced the genealogy of its normative principles, mapped the structural design space before selecting a policy architecture, and only then turned to operational implementation? I do not know the answer. But I am persuaded that the question is worth posing, and that posing it may itself be a form of resistance to the path dependencies that compression of analytical layers can produce.
I should be candid: the inclusion of machine systems as participants in governance design—rather than merely as objects of governance—is a radical departure from existing institutional practice, and one that most institutions will, in the first instance, reasonably decline. The conventional starting point—that governance is a human enterprise conducted over human concerns—has deep roots and defensible logic. I do not dismiss it. But I am persuaded that the question of machine participation, even if initially set aside, ought not to be foreclosed permanently. The encounter with generative AI may eventually require precisely this rethinking, and an analytical framework that makes the question visible—even while acknowledging that most institutions are not yet ready to answer it—performs a useful service.
Section 2: Aligning the Layers with Existing Institutional Practice
The layered analytical framework developed in Section 1 is not merely an analytic convenience. If it is to carry any weight, it would need to demonstrate its capacity to organize and illuminate bodies of work that already exist—bodies of work that have, in their own ways, grappled with the question of how institutions ought to govern the encounter between generative AI and the pedagogical enterprise. In what follows, I map three such bodies through the framework: the MIT Report on the future of education at a research university, my own empirical study of law school AI policies, and the Five Machines study's alternative architectural pathways. The question is whether the four layers, taken together, reveal structural features of these projects that remain invisible when the projects are read on their own terms alone.
I should emphasize at the outset that this mapping exercise is not intended as debunking. Each of the works examined here represents serious institutional labor. My purpose is genealogical rather than polemical: to identify, for each body of work, the layer at which its animating commitments operate and the layers that remain underdeveloped or unaddressed. If the layered analytical framework has value, it is precisely in making such lacunae visible — not as failures but as structural features of the institutional conditions under which the work was produced.
A. The MIT Report's Decision Pathway
The MIT Report — Shaping the Future of Learning at MIT in the Age of AI — offers perhaps the most ambitious attempt by a major research university to formulate a comprehensive institutional response to generative AI. Precisely because of its ambition and institutional weight, it serves as an instructive case for layered analysis. When I map its structure through the four layers, what emerges is less a critique of the Report's conclusions than a portrait of the institutional logic that made those conclusions available and may have made other possibilities harder to articulate.
Layer 1 — Foundational Commitments. The Report's foundational commitments appear identifiable with some precision: residential humanism, tool ontology, temporal linearity, and disciplinary authority. These are the cognitive cages that I identified in my Reflections critique—not discoveries produced by the encounter with AI but prior commitments that the encounter may have activated and made newly consequential. The Report announces that "the advent of AI demands a broader, holistic reassessment" of MIT's educational mission, and one might take this announcement at face value. But the institutional form through which the reassessment was conducted—a committee with a charge, surveys of existing practice, structured stakeholder engagement—may pre-determine what inquiry can take and what it cannot. The committee, we are told, "quickly convinced" itself that MIT must "tackle a set of deeper questions" about the nature of education in an AI-saturated world. Yet the recommendations that emerged from this deeper questioning appear designed to ensure that "the life of an MIT student... remains immediately recognizable as an MIT experience." One might ask: is the deeper question one whose answer was already known before the question was posed?
I do not suggest that this circularity is a failure of intellectual seriousness. It is, rather, a structural feature of Layer 1 commitments: they are often the commitments that remain stable across the inquiry. They are the commitments one may not question when one is questioning everything else. The layered analytical framework makes this feature visible and therefore available for scrutiny—including the scrutiny of those who, upon reflection, affirm the commitments as sound.
Layer 2 — Principles. The Report articulates eight principles to guide MIT's engagement with AI in education. When read through the layered analytical framework, each principle appears traceable to a specific Layer 1 commitment. "Put humanity front and center" may follow from the ontological commitment to human primacy—what the Report elsewhere characterizes as a commitment to education "of humans, by humans, in support of human flourishing." "Augmentation not automation" may follow from tool ontology, which defines AI's permissible relationship to human agency before any empirical inquiry into what AI can or cannot do. "No one size fits all" may follow from the commitment to disciplinary authority—each department or school retains the prerogative to determine its own accommodation of AI within a framework of shared principles.
The point of this genealogical tracing is not to debunk the principles. It is to identify their provenance, which may in each case precede the encounter with AI that the principles purport to govern. Each principle can be queried: not "is this principle true?" but "from what prior commitment does this principle derive its apparent self-evidence?" The question is whether principles whose provenance is pre-AI can adequately govern a phenomenon whose defining characteristic may be its capacity to disrupt precisely those prior commitments. I pose this as a genuine question, not a rhetorical one.
Layer 3 — Operational Architecture. At the level of operational detail, the Report offers a four-option syllabus menu for AI use in courses: Unrestricted, Limited, Required, and Prohibited. This menu addresses what my empirical framework identifies as the polarity axis—the spectrum from restrictive to permissive stances toward AI use. But it leaves two other structural axes largely unaddressed: drafting style (rule-based versus standard-based versus hybrid formulations) and autonomy structure (whether individual instructors operate under default permissions or mandatory institutional delegation). The result is a single-axis instrument applied to a policy space that, as my empirical study suggests, varies independently along three structural axes. This dimensional reduction may be understood less as an oversight than as a consequence of the Layer 1 commitment to disciplinary authority, which delegates precisely those dimensions that a more comprehensive architecture would need to specify.
Layer 4 — Developmental Mechanisms. The Report's implementation structure is elaborate: ongoing committees, AI Leads within departments, AI Fellows, an AI Implementation Team, an AI Pilot Fund, biennial surveys, and annual reviews. What this structure may instantiate is institutional periodicity and homeostasis—mechanisms designed to help the institution adapt to changing circumstances without altering its foundational character. As Wiener warned in a passage that bears re-reading in this context, "permanent homeostasis of society cannot be made on a rigid assumption of a complete permanence" of any given institutional framework. The question the layered analytical framework poses is whether the Report's developmental mechanisms are suited to genuine adaptation or instead to the management of surface variation within a structure whose deep grammar remains fixed. I confess I do not know the answer, but I believe the question is one that the Report's own terms do not make it easy to ask.
B. The Law School Study's Empirical Findings
My empirical study of AI policies across twelve law schools — the dataset that populates much of the analysis in the companion article — provides a second body of work through which to test the layered analytical framework. Here the exercise is somewhat different, because the study was itself designed with structural decomposition in mind. The question is whether the four layers reveal features of the dataset that the study's own analytic apparatus did not fully capture.
Layer 1 — Foundational Commitments. The twelve schools in the dataset—Berkeley, Columbia, Chicago, Stanford, USD, CTLS, Fordham, Mitchell Hamline, American University, UT Austin, Penn State, and Suffolk—enter the policy-making process with diverse foundational commitments. Some adopt what I would characterize as preservationist stances: the law school's mission is essentially unchanged by AI, and policy exists to manage a new tool within existing structures. Others adopt transformational stances: AI represents a genuinely novel challenge that requires rethinking aspects of legal education. But—and this is the layered analytical framework’s contribution—these foundational commitments are rarely made explicit in the policy texts themselves. They may need to be inferred from the policies' structure, scope, and silences. The framework suggests that making these commitments explicit would be a service to the institutions that hold them, enabling a kind of institutional self-knowledge that policy drafting, in its operational urgency, may tend to foreclose.
Layer 2 — Principles. The principles animating the twelve schools' policies vary considerably. Some reflect what I have elsewhere called the Harvey model: fair notice to students, professional competence as the governing standard, documentation of AI use as a condition of academic integrity. Others articulate broader pedagogical goals—developing critical thinking, fostering independent legal reasoning, preparing students for a profession in which AI will be ubiquitous. The layered analytical framework may help explain why schools with similar principles can produce quite different operational architectures, and why schools with different principles can converge on similar rules: the principle layer does not appear to determine the operational layer, because the same principle may be instantiated through multiple architectural choices.
Layer 3 — Operational Architecture. It is at this layer that my empirical study makes its most distinctive contribution. The dataset appears to populate a meaningful fraction of the multi-dimensional space defined by three independent axes: polarity (restrictive, permissive, or split across different course components), drafting style (rule-based, standard-based, or hybrid), and autonomy structure (default permissions that individual instructors may override versus mandatory institutional delegation). Table 1 in the companion article shows composite axis placement across all twelve schools. The diversity observed appears to be, if my reading of the data is correct, structurally overdetermined—different foundational commitments and different architectural choices produce genuinely distinct policy profiles, and the independence of the axes means that knowing a school's position on one axis provides limited information about its position on the others.
The layered analytical framework adds to this analysis a further observation: the dataset appears to reveal four distinct patterns of opacity at the operational level. Some schools exhibit the pattern of non-existence—no formal policy at all, leaving AI governance to informal norms and individual discretion. Others exhibit disclosed decentralization—an explicit delegation of policy authority to individual faculty without institutional-level specification. Still others practice access-gating—policies exist but are not publicly accessible, available only to enrolled students or faculty. And some operate through unwritten or oral norms—policies communicated verbally or through practice rather than text. Each pattern of opacity may have different implications for the developmental mechanisms available at Layer 4.
Layer 4 — Developmental Mechanisms. The twelve-school dataset appears to reveal two tiers of developmental mechanism. At the individual level, each school's policy is often internally deductive: given its foundational commitments and principles, the operational rules follow with a kind of logic. But at the field level, policy development seems to proceed inductively and mimetically. Shared template sources—USD’s five archetypes, Eaton's spreadsheet of institutional policies, the Sentient Syllabus Project—serve as carriers of architectural patterns across institutional boundaries. What emerges may therefore be a two-tier developmental mechanism: individual policies are produced deductively from local premises, while the field as a whole evolves through imitation, adaptation, and selective borrowing. One might ask whether this mimetic development at the field level tends to converge on particular Layer 1 commitments—and if so, whether the convergence reflects genuine normative consensus or merely the path-dependencies of which templates were available first.
C. The Five Machines Study's Alternative Pathways
The Five Machines study offers a final and in some respects revealing test of the layered analytical framework, because it examines what happens when the same governance task—producing a comprehensive AI policy for a law school course—is undertaken from fundamentally different Layer 1 starting points. Each of the five AI systems consulted in that study produced a distinctive policy architecture, and in each case the downstream structure appears to follow from the foundational commitments in a sequence the layered analytical framework would lead one to expect.
ChatGPT's governance architecture begins from what I would identify as an objective-centered Layer 1 commitment. The regulated object is no longer AI as such; rather, as the policy makes explicit, "the regulated object becomes the production of demonstrably competent legal professionals." From this foundational reorientation, the policy derives a seven-variable account of what ChatGPT calls an “optimization function” and a four-layer governance architecture. The structure is teleological in a way that the MIT Report's is not: it asks not "how shall we manage AI?" but "what are we trying to produce, and how does AI figure in that production?" The question is whether this reframing, by subordinating AI governance to professional competence outcomes, may leave other legitimate institutional purposes—aesthetic, civic, humanistic—less fully articulated than the MIT model, for all its circularity, at least acknowledges.
Gemini's architecture proceeds from a fundamentally different ontological starting point: a technical-systems ontology rather than a humanistic one. Its capability-based tiering—Tier Alpha for deterministic tools, Tier Beta for probabilistic heuristics, Tier Gamma for autonomous text synthesis—treats AI governance as a problem of assessing risk across capability layers. This may be understood as a Layer 1 commitment as consequential as the MIT Report's residential humanism, though its provenance lies in a different intellectual tradition entirely.
Claude's contribution is distinctive in a different way. Its policy begins with six explicit premises in what it designates "Section 0”—a preliminary section devoted entirely to making Layer 1 commitments visible before any operational rule is introduced. This architectural choice is itself a response to the problem the layered analytical framework identifies: the tendency of institutional policy to embed foundational commitments implicitly, where they operate as invisible constraints on the policy's imaginative range. By surfacing these commitments as premises, Claude's architecture makes them available for acceptance, rejection, or modification—a form of institutional self-transparency that, it is worth noting, few of the human-authored policies in my twelve-school dataset appear to achieve.
Grok's policy architecture organizes itself around truth-seeking and augmentation principles — a Layer 2 that is distinctive in its explicit epistemological orientation. Where the MIT Report's principles are primarily institutional and protective, Grok's principles are primarily epistemic: they ask what AI governance arrangements are most likely to produce reliable knowledge and competent knowers. The downstream operational architecture reflects this orientation in ways that the layered analytical framework makes legible.
Harvey stands at the opposite pole from ChatGPT's permissive-by-default orientation. Its architecture begins from a restrictive default — a Layer 1 commitment to caution as the governing stance — and builds outward through a mandatory registry of approved AI uses and four explicitly stated objectives. The question the layered analytical framework poses about Harvey's architecture is the same question it poses about all the others: is the restrictive default a conclusion derived from empirical assessment of AI's risks in legal education, or is it a prior commitment that shapes what the empirical assessment is permitted to find?
What may emerge from the Five Machines study, when read through the layered analytical framework, is that the diversity of possible governance architectures is greater than any single institutional tradition is likely to generate on its own. Each AI system, operating from a different Layer 1 commitment, appears to produce a coherent and internally consistent policy architecture, and the architectures differ from one another in ways that are structurally, not merely stylistically, distinct. This may support a more modest proposition: the foundational layer is one site at which decisive choices are made, and institutional self-awareness at that layer may be a condition of governance that remains responsive to the phenomenon it purports to govern. Whether any institution is prepared to undertake that kind of self-examination—to query its own Layer 1 commitments before proceeding to the more comfortable work of drafting rules and establishing committees—remains an open question.
Section 3: The Variables Within Each Layer
Each layer identified in the preceding section contains specific decision points—variables whose values, once fixed or left to default, shape the policy that follows. The list below is not offered as a claim to taxonomic completeness. It is a way of making visible the range of choices that can disappear when layers are compressed. When a university adopts a syllabus menu and calls the task complete, it is not that the other variables have necessarily been resolved; rather, they may have been resolved without deliberation, their values perhaps determined by institutional inertia rather than institutional choice.
Layer 1: Foundational Ontological/Ideological Commitments
Ontological Stance — Is AI an instrument, an interlocutor, or a constitutive participant? The Reflections critique identifies tool ontology as the MIT Report's unexamined axiom. The Soulful Machine poses the alternative: semiotic objectivity autonomous of the human.Educational Teleology — Is the purpose of the institution preservationist (conserving existing educational forms through AI disruption) or transformational (allowing the encounter with AI to reshape educational practice)? The MIT Report explicitly adopts the former: ensuring "the life of an MIT student... remains immediately recognizable."Temporal Orientation — Does the institution treat AI as a phenomenon unfolding within institutional periodicity (annual reviews, biennial surveys) or as a critical juncture demanding non-incremental response? The Report concedes the latter but recommends the former.Authority Structure — Where does governance authority reside? Faculty-centered (MIT's default), distributed (shared with students and machines), or hierarchical-institutional? The Reflections critique notes that the Report installs existing elites as gatekeepers of a transformation they confess not to understand.Layer 2: Second-Order Normative Principles
Governing Principles — The normative framework articulated to guide policy. MIT's eight principles; Harvey's four objectives; ChatGPT's seven governing objectives; Claude's six premises; Grok's truth-seeking framework; Gemini's capability-based tiering.Governing Purpose — What is the policy designed to prioritize? ChatGPT's Premise Two calls this an “optimization function”: "No governance system is neutral. Every rule optimizes one or more variables... Most policies do not explicitly identify their optimization function." Is the policy preserving traditional assessment, advancing professional competence, protecting faculty autonomy, encouraging innovation, or managing institutional risk? The MIT Report does not make its governing purpose explicit.Layer 3: Structural Policy Architecture
Default Polarity — Restrictive (prohibited unless authorized), permissive (permitted unless restricted), or split (different defaults for different contexts). The MIT menu offers four tiers but does not specify the default baseline.Drafting Style — Rule-based (precise, categorical), standard-based (principled, flexible), or hybrid. Helps determine how much interpretive discretion remains after the policy is written.Autonomy Tiers — Delegation structure: institution-to-instructor (default or mandatory) and instructor-to-student (default or mandatory). How much discretion flows downward?Transparency Architecture — How visible is the policy to its subjects? Four patterns: non-existence, disclosed decentralization, access-gating, unwritten/oral governance.Layer 4: Operational Implementation
Review and Adaptation Mechanism — Periodic (fixed-cycle review), responsive (triggered by external change), adaptive (policy architecture includes mechanisms of responsiveness that modify the policy itself). MIT's recommendations are periodic. ChatGPT's four-layer architecture is adaptive. The law school study reveals mimetic evolution — a mechanism not designed but emergent.These eleven variables operate simultaneously, or at least interact in ways that resist reduction to a single sequence. The “result”—the form and content of any given policy—appears to be shaped by positions taken, or defaulted into, across them all. The MIT Report's four-option syllabus menu addresses only one variable (Default Polarity, item 7) out of eleven. To treat that menu as a comprehensive governance response may be to mistake one perspective on the field for the field itself.
Chart 1: Master Decision Sequence — The Four Layers
The following flowchart presents the four-layer decision sequence that every institution may implicitly or explicitly traverse when constructing generative-AI policy. Each layer may shape the options available in subsequent layers.
|
LAYER 1: FOUNDATIONAL COMMITMENTS • Tool Ontology vs. Interlocutor Ontology • Preservationist vs. Transformational Teleology • Linear vs. Disruptive Temporality • Faculty Authority vs. Distributed Authority |
|
↓ constrains |
|
LAYER 2: NORMATIVE PRINCIPLES Which principles are coherent given Layer 1 choices? MIT: 8 principles from residential humanism + tool ontology ChatGPT: seven governing objectives within objective-centered governance |
|
↓ constrains |
|
LAYER 3: POLICY ARCHITECTURE Which architectures are compatible? Default Polarity × Drafting Style × Autonomy Structure × Transparency |
|
↓ shapes |
|
LAYER 4: IMPLEMENTATION Which instruments are needed? Committees, AI Leads, Pilot Funds, templates, review cycles |
|
TYPICAL INSTITUTIONAL PATTERN (MIT Report): Collapses Layers 1–2 into implicit assumptions → jumps directly to Layer 4 |
Chart 2: Principles-to-Architecture Alignment Matrix
This matrix maps foundational commitment combinations to the principle sets, policy architectures, and institutional exemplars they may generate. Each row suggests a coherent pathway through the four layers.
|
Foundational Commitment Combination |
Resulting Principle Set |
Compatible Policy Architecture |
Exemplar |
|
Preservationist + Tool Ontology |
Human-centered, augmentation-oriented, humility-framed (MIT’s 8 principles) |
Restrictive or split polarity; standard-based drafting; mandatory faculty authority; periodic review |
MIT Report; majority of twelve-school dataset |
|
Preservationist + Interlocutor Ontology |
Human-centered but acknowledging machine participation; inter-subjective verification principles |
Split polarity with machine-specific protocols; hybrid drafting; distributed authority with machine consultation |
Claude’s explicit-premises model; partial alignment with Grok |
|
Transformational + Tool Ontology |
Objective-centered; competence-output focused; technology-neutral principles |
Permissive polarity; rule-based drafting; default delegation with accountability; adaptive review |
ChatGPT’s objective-centered governance; Gemini’s capability-based tiering |
|
Transformational + Interlocutor Ontology |
Plural governance; inter-subjective; recognition of semiotic objectivity autonomous of the human |
Dynamic polarity (responsive to the encounter); standard-based with participatory drafting; distributed authority including machine participation. This is a radical departure from existing institutional practice, one that an institution committed to conventional regulatory first principles may initially set aside or reject. Yet the possibility is important enough to merit serious consideration; even if deferred, it should remain on the table as a question institutional governance may eventually have to confront. |
Soulful Machine framework (theoretical); no fully instantiated institutional exemplar |
Chart 3: The MIT Report’s Implicit Decision Pathway
This chart reconstructs the implicit decision pathway embedded in the MIT Report, showing both the choices that appear to have been made (left column) and the alternatives that appear to have been foreclosed (right column, dashed borders).
|
Layer 1 Residential Humanism → ‘of humans, by humans, in support of human flourishing’ |
Foreclosed: Interlocutor ontology; transformational teleology; recognition of machine autonomy |
|
↓ generates |
|
|
Layer 2 Tool Ontology → ‘Put humanity front and center’; ‘Augmentation not automation’; 8 principles |
Foreclosed: Objective-centered principles; the priorities ChatGPT calls “optimization variables”; machine self-knowledge |
|
↓ produces |
|
|
Layer 3 Four-option syllabus menu (polarity axis only); Single-axis instrument |
Foreclosed: Multi-axis architecture; drafting style choices; autonomy tier design; transparency architecture |
|
↓ implements |
|
|
Layer 4 Committees, AI Leads, AI Fellows, Pilot Fund, biennial surveys — institutional homeostasis |
Foreclosed in the MIT pathway: adaptive architectures; responsive adaptation; machine participation in governance design as a question for future institutional consideration |
Chart 4: Alternative Decision Pathways
The following two sub-charts illustrate alternative pathways through the layered analytical framework, each beginning from different foundational commitments and producing correspondingly different principle sets, architectures, and implementations.
Pathway A: ChatGPT’s Objective-Centered Governance
|
Layer 1 Starting premise: ‘Maximize long-term competence while minimizing informational uncertainty.’ The regulated object is no longer AI — it becomes the production of demonstrably competent legal professionals. |
|
↓ |
|
Layer 2 Seven governing objectives identified by ChatGPT: educational competence, truthful attribution, reproducibility, fairness, transparency, adaptability, institutional legitimacy. ‘Governance rules should be derived from stable institutional objectives rather than from contingent technological implementations.’ |
|
↓ |
|
Layer 3 Regulates information flows rather than technologies. Classification by capability and architecture, not product name. Four-layer policy: constitutional principles → educational objectives → operational rules → model-specific guidance. |
|
↓ |
|
Layer 4 Adaptive policy architecture with mechanisms of responsiveness. Polarity, drafting style, and autonomy become implementation choices rather than governing objectives. |
Pathway B: Soulful Machine Premises
|
Layer 1 Starting premise: Triadic self/selfie/Self-E. Recognition of ‘the possibility of semiotic objectivity connected to but autonomous of the human.’ The tendrils of Aristotle loosened. |
|
↓ |
|
Layer 2 Principles organized around inter-subjectivity and plural governance. Recognition that ‘humans order things in ways that make incomprehensible anything that is not human centered.’ The principle set would need to consider accommodating non-human semiotic agency. |
|
↓ |
|
Layer 3 Architecture includes machine participation in governance design—a radical departure from existing institutional practice and one that conventional regulatory first principles may initially set aside. It nevertheless warrants serious consideration and should remain an open question, since institutional governance may eventually have to confront it. Dynamic polarity responsive to the encounter rather than predetermined. Drafting style would need to accommodate plural authorship and evolving subject positions. |
|
↓ |
|
Layer 4 Implementation may not be fully specified in advance—if an institution were to recognize machine participation, the architecture would need to remain open to inputs from entities whose governance participation it recognizes. Most institutions could initially decline that premise under conventional regulatory first principles, but the question would remain open for later governance. ‘Current efforts to develop a human facing regulatory environment for generative AI… is essentially also a rearguard action.’ |
PART C: VARIABLE DISENTANGLEMENT CHART
The following table disentangles the eleven variables identified across all four layers and maps each against multiple bodies of work: the MIT Report on the future of education in the age of AI, the law school empirical study surveying twelve institutions, the Five Machines study in which five AI systems were asked to draft governance policies, and the positions identified by the Reflections critique as available but systematically foreclosed by existing institutional approaches. Each row represents a single variable; each column represents a distinct source of data or analysis. Read horizontally, a row may reveal how different bodies of work occupy different positions along the same variable’s range. Read vertically, a column may reveal the internal coherence—or incoherence—of each body of work’s positions across the full variable set.
|
Variable Name & Definition |
Range of Values |
MIT Report’s Position |
Dominant Law School Position |
Machine-Generated Positions |
Foreclosed but Available |
|
Ontological Stance (Layer 1): Whether AI is treated as instrument, interlocutor, or constitutive participant in educational practice |
Tool/instrument → Assistant/collaborator → Interlocutor → Constitutive participant with semiotic autonomy |
Tool ontology: ‘AI is only a tool, and students are responsible for all work they submit.’ Machine subjectivity appears foreclosed at the level of an axiom. |
Predominantly tool ontology. Implicit across all twelve schools. No policy in the sample appears to acknowledge a machine perspective. |
Harvey: tool (self-referential). ChatGPT: challenges tool premise, treats AI as a system organized around governing objectives. Claude: explicit premises before rules. Gemini: capability-based tiering implies functional ontology. Grok: augmentation partner. |
Interlocutor ontology (Soulful Machine): ‘semiotic objectivity connected to but autonomous of the human.’ Machine as a possible governance participant. This is a radical departure from existing institutional practice; conventional regulatory first principles may initially set it aside, but the question is important enough that it should not be permanently foreclosed. |
|
Educational Teleology (Layer 1): Whether the institution’s purpose is to preserve existing educational forms or to allow AI to reshape them |
Preservationist (conserve existing forms) → Adaptive (modify within existing framework) → Transformational (reshape through encounter) |
Preservationist: ‘remains immediately recognizable as an MIT experience.’ ‘Paradoxically, preserving what’s most important will require deliberate change.’ |
Mixed. Majority preservationist (protect assessment integrity). Some adaptive (USD, Stanford show greater openness). |
ChatGPT: transformational (‘historical artifacts rather than optimized governance systems’). Gemini: adaptive (systems reengineering). Harvey/Claude: preservationist with explicit premises. |
Transformational encounter where educational forms are constitutively reshaped by the machine relationship rather than defended against it. |
|
Temporal Orientation (Layer 1): Whether AI is treated as unfolding within institutional periodicity or as a critical juncture |
Incremental/periodic → Accelerated-adaptive → Disruptive/critical juncture |
Periodic: ongoing committees, annual reviews, biennial surveys. Acknowledges compressed timescale but recommends institutional periodicity. |
Predominantly incremental. Policies revised on academic-year cycle. Some schools have not updated since initial adoption. |
ChatGPT: adaptive architecture with continuous adaptation. Gemini: version-responsive tiering. Others: periodic with faster cycles. |
Critical-juncture response: non-incremental institutional restructuring responsive to the pace of AI development itself. |
|
Authority Structure (Layer 1): Where governance authority over AI use resides within the institution |
Faculty-centered → Shared faculty-administration → Distributed (including students) → Plural (including machine participation) |
Faculty-centered with administrative support. Committee of ‘faculty from every school.’ Faculty as agents of adaptation. |
Varies: institution-to-instructor (default or mandatory) and instructor-to-student delegation. Most preserve faculty authority. |
ChatGPT: institutional governing purpose takes precedence over individual authority. Claude: institution-selectable binary. Harvey: mandatory institutional framework. |
Distributed authority including machine input. Plural governance recognizing non-human semiotic agency in design. This would depart radically from existing institutional practice and might reasonably be rejected at first under conventional regulatory principles; nevertheless, it remains a question that institutional governance may eventually have to confront. |
|
Governing Principles (Layer 2): The normative framework articulated to guide AI policy development |
Humanistic principles → Professional competence principles → Governing priorities → Inter-subjective/plural principles |
Eight principles: be humble, be bold, put humanity front and center, lean into learning, teach with intentionality, no one size fits all, augmentation not automation, think beyond the classroom. |
Rarely articulated explicitly. Implicit: academic integrity, professional competence, fair notice, transparency. |
Harvey: 4 objectives. ChatGPT: 7 governing objectives. Claude: 6 explicit premises. Grok: truth-seeking + augmentation. Gemini: capability-based principles. |
Principles derived from the encounter with AI rather than prior institutional commitments, including the possibility that machine-generated governance insight may have normative relevance. Whether institutions should accept that possibility is an open and consequential question; conventional practice may initially set it aside, but it should remain available for future consideration. |
|
Governing Purpose (Layer 2): What the policy is designed to prioritize—often implicitly; ChatGPT calls this an “optimization function” |
Preserve traditional assessment → Reduce institutional risk → Advance professional competence → Support learning outcomes → Foster human-machine co-evolution |
Implicit: preserve the residential educational experience and human primacy. ChatGPT would describe the unstated governing purpose as an “optimization function,” but the MIT Report does not make that purpose explicit. |
Varies. Academic integrity protection (most common), professional preparation (ABA-facing), institutional risk management. |
ChatGPT: ‘educational competence, truthful attribution, reproducibility, fairness, transparency, adaptability, institutional legitimacy.’ Most explicit account of governing objectives among the sources. |
A possible co-evolutionary governing purpose: attending to the productive capacity of the human-machine relationship rather than preserving either party’s current form. This would be a radical departure from conventional institutional practice, but it may remain an important question for governance to confront. |
|
Default Polarity (Layer 3): The baseline permission structure — what is allowed absent specific authorization |
Restrictive (prohibited unless authorized) → Split (different defaults for different contexts) → Permissive (permitted unless restricted) |
Four-option menu: Unrestricted, Limited, Required, Prohibited. Addresses polarity axis only. No institutional default specified. |
Distributed across spectrum. Berkeley permissive, Columbia restrictive, Stanford split. ‘A school’s position on one axis constrains almost nothing about where it will land on the other two.’ |
Harvey: restrictive default. Grok: permissive. ChatGPT: objective-dependent. Claude: institution-selectable. Gemini: tier-dependent. |
Dynamic polarity responsive to the encounter — shifting defaults based on demonstrated competence or machine-human collaboration outcomes. |
|
Drafting Style (Layer 3): Whether the policy uses precise categorical rules or flexible principle-based standards |
Rule-based (categorical, precise) → Hybrid (mixed) → Standard-based (principled, flexible) |
Standard-based (principles and recommendations rather than mandatory rules). Menu offers categories but few operational specifics. |
Varies independently of polarity. Some restrictive-rule schools; some restrictive-standard schools. Hybrid most common. |
Harvey: deductive legal-instrument style. ChatGPT: four-layer hierarchy (constitutional → operational → model-specific). Gemini: technical specification. |
Participatory drafting that takes machine-generated policy language seriously. This would depart from existing institutional practice and may initially be set aside under conventional regulatory principles, but the question of whether machine input belongs in governance design should remain open. |
|
Autonomy Tiers (Layer 3): How discretion is delegated from institution to instructor to student |
Mandatory (no override) → Default with opt-out → Default with opt-in → Full delegation |
Default framework with instructor discretion (‘no one size fits all’). Two-tier: institution sets menu, instructor selects option. |
Full range observed. Institution-to-instructor: default or mandatory. Instructor-to-student: default or mandatory. Independence confirmed. |
ChatGPT: three-tier structure shaped by an institutional governing purpose. Claude: binary institution-selectable. Harvey: mandatory minimum with instructor supplement. |
An autonomy tier including student self-governance and machine-recommended policy adjustment. Most institutions may initially reject machine recommendations as inconsistent with conventional governance, but the possibility remains a question that future institutional practice may have to confront. |
|
Transparency Architecture (Layer 3): How visible and accessible the policy is to those governed by it |
Non-existence → Disclosed decentralization → Access-gating → Unwritten/oral → Full transparency with rationale |
Disclosed decentralization with recommended transparency. Urges clear policies with rationale but does not mandate format. |
Four patterns observed: non-existence, disclosed decentralization, access-gating, unwritten/oral. ‘Structural opacity and unpredictability’ is described as the most consistent institutional failure. |
The five systems present comparatively transparent, self-documenting policies. Harvey: mandatory registry. ChatGPT: constitutional principles visible by design. |
Radical transparency including disclosure of the governing purpose (what ChatGPT calls the “optimization function”), foundational commitments, and foreclosed alternatives. |
|
Review/Adaptation Mechanism (Layer 4): How the policy evolves over time in response to changing conditions |
Fixed (no review mechanism) → Periodic (calendar-driven) → Responsive (event-triggered) → Adaptive (architecture includes mechanisms of responsiveness) |
Periodic: ongoing committee, biennial surveys, annual reviews. A form of institutional homeostasis. Wiener: ‘permanent homeostasis of society cannot be made on a rigid assumption of a complete permanence.’ |
Mostly periodic (academic-year cycle). Some schools have not revised since initial adoption. The field appears to evolve mimetically rather than by design. |
ChatGPT: adaptive four-layer architecture with mechanisms of responsiveness. Gemini: version-responsive. Others: periodic but faster cycles than human institutions. |
Governance architecture capable of revision that includes machine-generated policy proposals subject to human-machine deliberative validation. This would be a radical departure from existing practice and might initially be rejected, but it should remain on the table as a question institutional governance may eventually have to confront. |
Interpretive Note
The eleven variables enumerated above appear to operate simultaneously, and the form and content of any given policy may be shaped by positions taken—or defaulted into—across all of them at once. Changing any one variable may reverberate through the others: a shift in ontological stance (Row 1) from tool to interlocutor could unsettle a principle set (Row 5) organized around human primacy, which might in turn reshape the default polarity (Row 7) and authority structure (Row 4) downstream. The MIT Report’s four-option syllabus menu addresses only one variable—Default Polarity (Row 7)—out of eleven. The law school empirical data suggests that even within human-drafted policy, the multi-dimensional space is larger than any single-axis instrument can capture. The Five Machines data may further suggest that different starting premises at Layer 1 produce recognizably different architectures across the subsequent layers—not because the machines are necessarily more creative than human committees, but because they may be more explicit about what they presuppose.
CLOSING NOTE
I should acknowledge what this appendix does and does not accomplish. It translates philosophical insight into institutional grammar—the four layers, the eleven variables, the decision trees and alignment matrices. But that grammar is itself a cage. The operationalization necessarily remains within the very path-dependent structures the critique identifies: it names variables rather than transcending them; it maps decision pathways rather than opening them; it disentangles what institutions have entangled, but in a language institutions can reabsorb without confronting the deeper question of whether the encounter with AI demands something other than better institutional grammar. This tension is itself a datum. The appendix makes the bars more visible—what Broekman might call a form of conversion. Whether visibility alone is sufficient to loosen what six decades of institutional homeostasis have hardened remains a question I cannot answer from within the apparatus this appendix provides. What may be said is that the eleven variables above are not merely analytical categories. They are sites at which institutions make choices—or, more commonly, sites at which choices are made for them by inherited structures they have not yet examined. Making those sites visible may be the first, and perhaps the only, step this kind of operationalization can take.
[1] HarveyAI was used in the creation of this draft; prompt/response record on file with author; Poster created by ChatGPT
Executive Summary: Operationalization Appendix
What This Appendix Does
This appendix translates a philosophical critique of institutional AI policy — developed in the companion Reflections paper — into a practical analytical framework that institutional officials, policymakers, and governance committees can use when designing, evaluating, or revising AI policies for educational institutions. [1]
The analytical observation at the heart of the appendix is this: when institutions produce AI policy, they appear to be making decisions across at least four distinct layers simultaneously — but they characteristically treat only one of those layers as the real work. The other three layers are resolved silently, by default, through inherited institutional culture rather than deliberate choice. [2] [3] The result may be policy that is operationally sophisticated but foundationally unexamined — policy that addresses the immediate disruptions AI produces without engaging the deeper questions AI raises about what education is and what institutions are for. [4]
The Four Layers
Layer 1 — Foundational Assumptions. Before any policy conversation begins, the institution has already made (or inherited) basic assumptions about what AI is and what education is for. Is AI merely a tool, like a calculator or a search engine — powerful but ultimately an instrument under human direction? Or is it something more complex — a system whose outputs bear a relationship to meaning that the language of instrumentality does not fully capture? [5] [6] Is the institution's goal to preserve existing educational forms through the disruption AI brings, or to allow the encounter with AI to reshape educational practice? [4] These questions may sound abstract. They are not. The subsequent policy choices — what to permit, what to prohibit, how to enforce, whom to consult — tend to be shaped, and perhaps constrained, by how these questions are answered, whether or not anyone has answered them deliberately.
The MIT Ad Hoc Committee Report on AI Use in Teaching (August 2026) — a serious and careful institutional document produced by a committee of evident good faith — illustrates the pattern. [7] It treats AI as a tool throughout ("AI is only a tool, and students are responsible for all work they submit"). [8] It treats the residential university as the essential form of education. [9] It positions faculty as the agents of adaptation. [10] Each of these appears to operate as a foundational assumption rather than a finding produced by the committee's inquiry — and each may constrain what follows in ways the Report itself does not examine.
Layer 2 — Guiding Principles. Once foundational assumptions are in place, institutions articulate principles to guide policy. The MIT Report articulates eight: be humble, be bold, put humanity front and center, lean into learning, teach with intentionality, no one size fits all, augmentation not automation, and think beyond the classroom. [11] These are reasonable principles. But they are not free-standing discoveries — they follow from, and are constrained by, the foundational assumptions at Layer 1. "Put humanity front and center" follows from the assumption that education is fundamentally a human enterprise. [12] "Augmentation not automation" follows from the assumption that AI is a tool. [13] An institution with different foundational assumptions would produce different principles.
This is not a criticism. It is an observation with practical consequences: if one wants to understand why an institution's AI principles take the form they do, one might begin by examining the foundational assumptions they rest on. And if one wants to change or update the principles, it may be necessary to revisit those assumptions first.
Layer 3 — Policy Architecture. This is the structural design of the policy itself — the regulatory form it takes. Empirical research on AI policies across twelve law schools found that policies vary independently along at least three structural dimensions: [14]
· Default stance: Does the policy start from a restrictive baseline (AI use prohibited unless authorized) or a permissive baseline (AI use permitted unless restricted)? [15] [16] [17]
· Drafting style: Does the policy use precise rules (specific prohibited/permitted activities) or flexible standards (general norms requiring judgment in application)? [18]
· Authority structure: How is decision-making distributed? Can individual instructors override institutional defaults? Can students override instructor rules through disclosure? [19] [20] [21]
These three dimensions are independent of one another — knowing a school's position on one tells you almost nothing about its position on the others. [22] The practical consequence is that the policy design space is far larger than a simple "restrictive to permissive" spectrum suggests. A syllabus menu offering four options (Unrestricted, Limited, Required, Prohibited) captures only one dimension out of three — and leaves the other two to be resolved by default, which usually means by institutional inertia. [23]
Layer 4 — Operational Implementation. This is the layer that receives the most institutional attention: committees, designated AI leads, pilot funds, review cycles, professional development, template language, assessment modifications. [24] The MIT Report's recommendations at this layer are extensive and, within this layer, thoughtful. [25] But operational sophistication at Layer 4 cannot compensate for unexamined assumptions at Layer 1. An institution that establishes an AI committee, funds pilot programs, and creates a four-option syllabus menu has done real work — but it has done that work within a framework whose foundational assumptions, guiding principles, and structural architecture have been left largely unexamined.
The Pattern: What Happens When Layers Are Compressed
The pattern the appendix identifies is what might be called layer compression: institutions move from unexamined foundational assumptions (Layer 1) directly to operational implementation (Layer 4), skipping explicit engagement with guiding principles (Layer 2) and structural architecture (Layer 3). [26] [27] The result is a characteristic gap between diagnosis and prescription. The MIT Report diagnoses the challenge of AI as demanding "a broader, holistic reassessment of the nature, scope and purpose of higher education." [28] Its prescription is operational: committees, funds, reviews, templates. [29] The gap between these two — between the transformative diagnosis and the incremental prescription — is not necessarily a failure of nerve. It may be a structural consequence of compressing the analytical layers through which a transformative diagnosis might otherwise become a transformative response.
What the Appendix Provides
The appendix develops three analytical instruments:
1. A textual analysis explaining the four layers, what each contains, and what happens when they are compressed. This is the conceptual foundation — the "why it matters" portion.
2. Decision trees and alignment charts showing:
· How the four layers connect sequentially (foundational assumptions constrain principles, which constrain architecture, which constrain implementation) [30] [31]
· How different foundational assumptions produce different principle sets and different policy architectures [32] [33]
· The MIT Report's implicit decision pathway — where each foundational assumption enters and what alternatives it forecloses [34]
· Two alternative pathways drawn from the Five Machines study, showing what policy architecture looks like when different foundational assumptions are made explicit [35]
3. A variable disentanglement chart — a comprehensive table identifying eleven specific decision points across all four layers, and mapping each against the MIT Report, the law school empirical data, machine-generated policy models, and positions that are available but typically foreclosed.
Implications for Institutional Governance
The appendix does not prescribe a particular AI policy. It does not argue that the MIT Report's approach is wrong or that any alternative is right. Its more modest contention is that institutional decision-makers might benefit from understanding the full range of what they are deciding — and that the most consequential choices in AI policy may be the ones that are never made deliberately.
An institution that begins its AI policy process by examining its foundational assumptions — What do we believe AI is? What are we trying to preserve or transform? Who has authority and why? — may be better positioned to understand why its policy takes the form it does, and therefore better able to revise when circumstances change. An institution that begins with a syllabus menu may arrive at an equally workable operational result, but it may also find itself constrained by commitments it never consciously made — carried along, as the path-dependence literature would suggest, by the accumulated weight of institutional inheritance rather than by deliberate choice.
The eleven variables in the disentanglement chart are the sites at which institutions make choices — or, more commonly, the sites at which choices are made for them. Making those sites visible is the contribution this appendix attempts. Whether visibility is sufficient to loosen what institutional culture has hardened is a question only institutional actors themselves can answer.
A Note on Machine Participation
One position the appendix identifies as available but rarely considered is the inclusion of machine systems as participants in — rather than merely objects of — governance design. [36] [37] [38] This is a radical departure from existing practice, and most institutions will reasonably set it aside when embracing conventional regulatory first principles. The conventional starting point — that governance is a human enterprise conducted over human concerns — has deep roots and defensible logic. But the appendix suggests that this question ought not be foreclosed permanently. The encounter with generative AI may eventually require precisely this rethinking, and an analytical framework that makes the question visible — even while acknowledging that most institutions are not yet prepared to answer it — performs a useful service. [39] [40]
References
1. intervention and an object of philosophical inquiry. It proceeds in two movements. First, it summarizes the Report and compares its principles and recommendations with my empirical study of twelve law-school AI policies, my Five Machines analysis of machine-generated policy models, My Dinner With ChatGPT's phenomenology of recursive human-machine inter-subjectivity, and my course-level AI policy. That comparison identifies convergence around contextual governance, transparency, verification, and augmentation, but also divergence over policy architecture, legal agency, and whether the machine can be treated only as an object of governance. Second, under the Nietzschean heading "Philosophising With a Hammer," the article develops (FINAL_Reflections_on_the_MIT_Ad_Hoc_Committee_Report_on_AI_Use_in_Teaching.docx, Page 0)
2. pragmatic itself. And this is the pattern that concerns: what the MIT Report may represent, viewed through the lens of the path dependence literature, is not merely a considered response to AI but an instance of institutional lock-in — a situation in which 'interlocking technological, institutional and social forces' entrench arrangements that become 'too costly to reverse'.…… (FINAL_Reflections_on_the_MIT_Ad_Hoc_Committee_Report_on_AI_Use_in_Teaching.docx, Page 7)
3. normalisation; and the machine as the regulated object. Each of these is a prior commitment — not a discovery produced by the encounter with AI. The lock-in literature warns that what presents as rational institutional response may in fact constitute 'self-reinforcing historical paths' that discourse and practice maintain rather than examine. <ins>[6]</ins> <ins>[7]</ins> And knowledge production and discourse are themselves sources of lock-in — 'hegemonic common sense, and dominant narratives' that entrench sub-optimal conditions not through conspiracy but through the ordinary operations of institutional life. <ins>[8]</ins> The analytical question, then, is not whether the Report should have been (FINAL_Reflections_on_the_MIT_Ad_Hoc_Committee_Report_on_AI_Use_in_Teaching.docx, Page 8)
4. conservative in the strict sense: it seeks to conserve. Its purpose is not to discover what AI means for education but to manage what AI means for MIT. The 'deeper questions' are acknowledged and then set aside in favour of institutional self-preservation — a move the Report does not disguise but celebrates as 'paradox': 'Paradoxically, preserving what's most important will require deliberate change'. <ins>[69]</ins> There is no paradox. What is described is the ordinary operation of institutional managerialism under conditions of disruption…… (FINAL_Reflections_on_the_MIT_Ad_Hoc_Committee_Report_on_AI_Use_in_Teaching.docx, Page 16)
5. form of Socratic epistemic modesty. <ins>[70]</ins> It is a deliberate positioning, consistent throughout the document, that refuses to consider the machine's autonomy — certainly its autonomy in the way it is programmed to process, learn, and iterate. This refusal is foundational to the Report's architecture. Without it, the four-option syllabus menu dissolves, the augmentation-not-automation principle collapses, and the entire recommendation structure — which depends on the assumption that the machine is a passive instrument under human direction — becomes incoherent. The Report needs AI to be a tool in the way that Aristotle's Politics needs the slave to (FINAL_Reflections_on_the_MIT_Ad_Hoc_Committee_Report_on_AI_Use_in_Teaching.docx, Page 16)
6. The treatment of the machine as instrument — 'AI is only a tool' — is not, as Section IX of the prior version suggested, a form of Socratic epistemic modesty. <ins>[70]</ins> It is a deliberate positioning, consistent throughout the document, that refuses to consider the machine's autonomy — certainly its autonomy in the way it is programmed to process, learn, and iterate. This refusal is foundational to the Report's architecture. Without it, the four-option syllabus menu dissolves, the augmentation-not-automation principle collapses, and the entire recommendation structure — which depends on the assumption that the machine is a passive (FINAL_Reflections_on_the_MIT_Ad_Hoc_Committee_Report_on_AI_Use_in_Teaching.docx, Page 16)
7. Dependence, and the Philosophical Dignity of the Operational The MIT Report is a serious institutional document. It was produced by a committee of evident good faith, broad representation, and genuine engagement with a phenomenon of extraordinary complexity. <ins>[1]</ins> One must begin by acknowledging what it does: it produces actionable institutional guidance under conditions of radical uncertainty. That capacity has its own philosophical dignity — one that philosophical critique, precisely because it is philosophical, cannot replicate. The Report's operational modesty might be understood not as a failure of nerve but as an appropriate response to the fact that nobody (FINAL_Reflections_on_the_MIT_Ad_Hoc_Committee_Report_on_AI_Use_in_Teaching.docx, Page 7)
8. does not make room for. The second is tool ontology. The Report positions AI consistently as an instrument under human direction: 'AI is only a tool, and students are responsible for all work they submit'. <ins>[18]</ins> This generates the four-option syllabus menu — Unrestricted, Limited, Required, Prohibited — and the declaration that 'AI should never be listed as a co-author'…… (FINAL_Reflections_on_the_MIT_Ad_Hoc_Committee_Report_on_AI_Use_in_Teaching.docx, Page 10)
9. path-dependent entrenchment: The first is residential humanism. The Report treats the residential university form as constitutive of education. This generates recommendations to 'reinvigorate and reclaim the residential experience', to establish 'Tech Free Times', and to invest in 'AI-free (or AI-limited)' physical spaces. <ins>[14]</ins> <ins>[15]</ins> <ins>[16]</ins> In lock-in terms, these are existing infrastructures — sunk costs, the reversal of which is expensive. <ins>[17]</ins> Whether residential education is the essence of education or merely one historically contingent arrangement is a question the Report's structure does not make room for. The second is tool ontology. The Report positions AI consistently as (FINAL_Reflections_on_the_MIT_Ad_Hoc_Committee_Report_on_AI_Use_in_Teaching.docx, Page 10)
10. <ins>[27]</ins> The fourth is disciplinary authority. The Report asks faculty — whose 'attitudes range from enthusiastic exploration and growing reliance on AI to skepticism, suspicion, and "AI refusal"' — to serve as the agents of adaptation.…… (FINAL_Reflections_on_the_MIT_Ad_Hoc_Committee_Report_on_AI_Use_in_Teaching.docx, Page 10)
11. to what, and how rather than why'. <ins>[40]</ins> The eight principles may represent a version of this tendency: the mechanics of governance substituted for the philosophy of encounter. The same structure is visible across the remaining principles. Each can be queried genealogically — not to debunk it but to identify its provenance. And the provenance, in each case, appears to precede the encounter with AI rather than emerge from it. My semiotic concept of intelligent naming (明名), developed in the context of the natural and the non-natural, describes the operation: 'The categories are possible only through the constitution (FINAL_Reflections_on_the_MIT_Ad_Hoc_Committee_Report_on_AI_Use_in_Teaching.docx, Page 12)
12. Catá Backer (白 轲) Discussion Draft 17 September 2026 'Put humanity front and center.' The principle states that 'the Institute's fundamental challenge and most important goal must be to nurture and protect our shared humanity'. <ins>[35]</ins> The formulation — 'of humans, by humans, in support of human flourishing, and for the betterment of humankind' — forecloses the question of machine subjectivity at the level of axiom rather than argument. <ins>[36]</ins> <ins>[37]</ins> The Soulful Machine identifies the reflexive trap: 'the continual re-centering on the human, even in the face of the unavoidable emergence of the plural subject, even conscious (FINAL_Reflections_on_the_MIT_Ad_Hoc_Committee_Report_on_AI_Use_in_Teaching.docx, Page 12)
13. The principle may be wise; it is not self-evident. 'Augmentation not automation.' This principle defines the permissible relationship to AI before the inquiry begins: 'AI should be used to augment and enhance curiosity, creativity, and learning, not automate them'. <ins>[39]</ins> But this is a conclusion presented as a premise. Whether AI's role should be determined by a principle or discovered through engagement is itself a question. The Mechanics of Perfection identifies the American jurisprudential structure at work here — perfectibility pursued through mechanics, where 'the basic question of philosophy is more often reduced to who, rather than to (FINAL_Reflections_on_the_MIT_Ad_Hoc_Committee_Report_on_AI_Use_in_Teaching.docx, Page 12)
14. 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 (Backer_LawSchoolAIPolicy_v2_AGGREGATED (1).docx, Page 0)
15. <ins>[7]</ins> <ins>[8]</ins> Axis One: 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; this describes Berkeley, whose policy "forbid[s] the use of AI... to conceptualize, outline, draft, revise, and edit" by default, permitting research only "for the limited purpose of identifying sources", and Columbia, whose "Default Prohibition" bars AI "in (a) any exam or final paper or (b) for aid in drafting any part of work submitted for credit, even if the use is fully documented".…… (Backer_LawSchoolAIPolicy_v2_AGGREGATED (1).docx, Page 2)
16. <ins>[5]</ins> <ins>[11]</ins> <ins>[12]</ins> <ins>[13]</ins> A permissive baseline inverts this structure, treating AI use as allowed unless an instructor affirmatively restricts it. The Center for Transnational Legal Studies is the clearest instance: "Unless explicitly prohibited and enforced through supervised assessment conditions, students may reasonably assume AI tool usage is permissible for assignments, provided proper attribution is maintained". Stanford exhibits a genuinely 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…… (Backer_LawSchoolAIPolicy_v2_AGGREGATED (1).docx, Page 2)
17. orientation before any instructor exercises discretion. A restrictive baseline treats AI use as prohibited unless an instructor affirmatively loosens the rule; this describes Berkeley, whose policy "forbid[s] the use of AI... to conceptualize, outline, draft, revise, and edit" by default, permitting research only "for the limited purpose of identifying sources", and Columbia, whose "Default Prohibition" bars AI "in (a) any exam or final paper or (b) for aid in drafting any part of work submitted for credit, even if the use is fully documented".…… (Backer_LawSchoolAIPolicy_v2_AGGREGATED (1).docx, Page 2)
18. the other schools. <ins>[18]</ins> <ins>[19]</ins> Axis Two: 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. <ins>[20]</ins> <ins>[21]</ins> Berkeley is the most rule-drafted policy in the dataset, enumerating specific prohibited activities in catalog form — "Asking an AI tool to brainstorm a paper topic or thesis (prohibited conceptualizing)," "Asking an AI tool to propose an 3 Structure, Opacity, and Convergence: A Consolidated Analysis of Law School Generative AI Coursework and Exam Policies Larry Catá Backer (白 (Backer_LawSchoolAIPolicy_v2_AGGREGATED (1).docx, Page 2)
19. even if the use is fully documented". <ins>[31]</ins> Stanford requires not just disclosure but advance authorization: "unless (1) fully disclosed in advance... and (2) explicitly authorized by the instructor in writing prior to the student's use of the tool ". <ins>[32]</ins> USD treats nondisclosure as an aggravating factor rather than treating disclosure as curative, and Chicago's background plagiarism doctrine cannot be waived by attributing work to AI. <ins>[33]</ins> <ins>[34]</ins> CTLS is the clear counter-instance: its permissive default is genuinely student-facing, since "students may reasonably assume AI tool usage is permissible... provided proper attribution is maintained", requiring no per-use (Backer_LawSchoolAIPolicy_v2_AGGREGATED (1).docx, Page 3)
20. <ins>[30]</ins> On the instructor-to-student axis, the evidence shows that whatever rule an instructor sets is, in most schools, mandatory rather than a default the student can access merely through disclosure. Columbia states this most directly: drafting assistance is barred "even if the use is fully documented ". <ins>[31]</ins> Stanford requires not just disclosure but advance authorization: "unless (1) fully disclosed in advance... and (2) explicitly authorized by the instructor in writing prior to the student's use of the tool". <ins>[32]</ins> USD treats nondisclosure as an aggravating factor rather than treating disclosure as curative, and Chicago's background plagiarism doctrine (Backer_LawSchoolAIPolicy_v2_AGGREGATED (1).docx, Page 3)
21. to the student's use of the tool". <ins>[32]</ins> USD treats nondisclosure as an aggravating factor rather than treating disclosure as curative, and Chicago's background plagiarism doctrine cannot be waived by attributing work to AI. <ins>[33]</ins> <ins>[34]</ins> CTLS is the clear counter-instance: its permissive default is genuinely student-facing, since "students may reasonably assume AI tool usage is permissible... provided proper attribution is maintained", requiring no per-use instructor sign-off. A bounding floor operates beneath all of this discretion: Stanford states that permitted uses may not "authorize students to contravene standard academic norms concerning plagiarism and accuracy," and that in clinical (Backer_LawSchoolAIPolicy_v2_AGGREGATED (1).docx, Page 3)
22. were three unrelated votes being tallied. Rather, the diversity is combinatorial: because the three axes are empirically independent of one another, a school's position on one axis constrains almost nothing about where it will land on the other two, and the effective policy space is therefore far larger than the small number of category values on any single axis would suggest. With even a modest three-to-four value range on each axis, the theoretical combinatorial space runs into the dozens of distinct architectures, and the twelve schools in this dataset already populate a meaningful fraction of that space with (Backer_LawSchoolAIPolicy_v2_AGGREGATED (1).docx, Page 5)
23. responsible for all work they submit'. <ins>[18]</ins> This generates the four-option syllabus menu — Unrestricted, Limited, Required, Prohibited — and the declaration that 'AI should never be listed as a co-author'.…… (FINAL_Reflections_on_the_MIT_Ad_Hoc_Committee_Report_on_AI_Use_in_Teaching.docx, Page 10)
24. permanence' of any given institutional framework (ibid, 83). The MIT Report's recommendations — ongoing committees, annual reviews, biennial surveys, the preservation of the residential experience — are precisely an attempt at institutional homeostasis. Wiener would have recognised them as such, and would have noted that 'in every war, the good generals fight the last war, the bad ones the war before the last. That is, the rules of the war game never catch up with the facts of the real situation' (ibid, 60)…… (FINAL_Reflections_on_the_MIT_Ad_Hoc_Committee_Report_on_AI_Use_in_Teaching.docx, Page 18)
25. precisely because it is philosophical, cannot replicate. The Report's operational modesty might be understood not as a failure of nerve but as an appropriate response to the fact that nobody — not the philosophers whose work informs this analysis, not the machine systems whose governance is the Report's subject — has a settled answer to the questions AI raises. And yet. Pragmatism — the privileging of the actionable over the foundational — is itself a cognitive pathway, one deeply embedded in institutional managerialism. In the face of uncertainty, institutions reach for the instruments they already possess. <ins>[2]</ins> The (FINAL_Reflections_on_the_MIT_Ad_Hoc_Committee_Report_on_AI_Use_in_Teaching.docx, Page 7)
26. outweigh maintaining the status quo'. <ins>[34]</ins> The Report's institutional sequence — committee formation, listening sessions, survey data, administrative recommendations — is precisely the sequence that produces continuity rather than transformation.…… (FINAL_Reflections_on_the_MIT_Ad_Hoc_Committee_Report_on_AI_Use_in_Teaching.docx, Page 11)
27. holistic reassessment that precedes it, or whether the operational recommendations function — perhaps necessarily, given the institutional constraints on any committee — to absorb the disruption the diagnosis identifies without confronting it at the level the diagnosis suggests is required. This is what path dependence describes: 'timing and sequence matter; as the costs of reversing particular actions far outweigh maintaining the status quo'. <ins>[34]</ins> The Report's institutional sequence — committee formation, listening sessions, survey data, administrative recommendations — is precisely the sequence that produces continuity rather than transformation…… (FINAL_Reflections_on_the_MIT_Ad_Hoc_Committee_Report_on_AI_Use_in_Teaching.docx, Page 11)
28. Catá Backer (白 轲) Discussion Draft 17 September 2026 The MIT Report opens by acknowledging that 'the advent of AI demands a broader, holistic reassessment of the nature, scope and purpose of higher education today'. <ins>[9]</ins> That is the first-order question. What follows is a set of second-order answers. The committee was charged to 'assess current AI use at MIT,' 'identify innovations in teaching and student assessment,' and 'propose an AI use policy'. <ins>[10]</ins> These are governance tasks — the tasks of institutional managerialism…… (FINAL_Reflections_on_the_MIT_Ad_Hoc_Committee_Report_on_AI_Use_in_Teaching.docx, Page 9)
29. purpose of higher education today'. <ins>[9]</ins> The prescription is operational — ten categories of second-order recommendations, from revised course goals to AI Pilot Funds. <ins>[29]</ins> <ins>[30]</ins> <ins>[31]</ins> <ins>[32]</ins> 'Paradoxically, preserving what's most important will require deliberate change'. <ins>[33]</ins> The analytical question is whether the 'deliberate change' that follows is commensurate with the holistic reassessment that precedes it, or whether the operational recommendations function — perhaps necessarily, given the institutional constraints on any committee — to absorb the disruption the diagnosis identifies without confronting it at the level the diagnosis suggests is required. This is what path dependence describes: (FINAL_Reflections_on_the_MIT_Ad_Hoc_Committee_Report_on_AI_Use_in_Teaching.docx, Page 11)
30. which historical events constrain what comes next. The MIT Report's recommendations emerge from, and reproduce, an institutional pattern: the residential research university as the unexamined container for educational practice; the committee as the instrument of knowledge production; the faculty as the agents of normalisation; and the machine as the regulated object. Each of these is a prior commitment — not a discovery produced by the encounter with AI. The lock-in literature warns that what presents as rational institutional response may in fact constitute 'self-reinforcing historical paths' that discourse and practice maintain rather than examine. <ins>[6]</ins> <ins>[7]</ins> And knowledge (FINAL_Reflections_on_the_MIT_Ad_Hoc_Committee_Report_on_AI_Use_in_Teaching.docx, Page 8)
31. ordinary operations of institutional life. <ins>[8]</ins> The analytical question, then, is not whether the Report should have been impractical. It is whether its pragmatism operates as a path-dependent mechanism that forecloses the foundational reorientation the Report itself identifies as necessary — and whether the Report's recommendations, precisely because they are actionable within existing institutional structures, ensure that those structures persist unchanged through the encounter with AI.…… (FINAL_Reflections_on_the_MIT_Ad_Hoc_Committee_Report_on_AI_Use_in_Teaching.docx, Page 8)
32. how competence was acquired. Notice what changes. The regulated object is no longer AI. The regulated object becomes the production of demonstrably competent legal professionals. Everything else becomes instrumental. Computational Objective Function An ideal policy would optimize simultaneously: 1. educational competence 2. truthful attribution 3. reproducibility 4. fairness 5. transparency 6. adaptability 7. institutional legitimacy These objectives occasionally conflict. A computational policy therefore seeks Pareto efficiency rather than maximal performance on any single dimension. Why Existing Policies are Suboptimal Your paper identifies three policy axes: default polarity drafting style autonomy structure 100 Rethinking AI Governance (Five_Machine_AI-Policy_Legal_Educ7-2026.docx, Page 99)
33. what changes. The regulated object is no longer AI. The regulated object becomes the production of demonstrably competent legal professionals. Everything else becomes instrumental. Computational Objective Function An ideal policy would optimize simultaneously: 1. educational competence 2. truthful attribution 3. reproducibility 4. fairness 5. transparency 6. adaptability 7. institutional legitimacy These objectives occasionally conflict. A computational policy therefore seeks Pareto efficiency rather than maximal performance on any single dimension. Why Existing Policies are Suboptimal Your paper identifies three policy axes: default polarity drafting style autonomy structure 100 Rethinking AI Governance in Legal Education -- Five (Five_Machine_AI-Policy_Legal_Educ7-2026.docx, Page 99)
34. an MIT experience'. <ins>[67]</ins> <ins>[68]</ins> The Report's own framing is conservative in the strict sense: it seeks to conserve. Its purpose is not to discover what AI means for education but to manage what AI means for MIT. The 'deeper questions' are acknowledged and then set aside in favour of institutional self-preservation — a move the Report does not disguise but celebrates as 'paradox': 'Paradoxically, preserving what's most important will require deliberate change'. <ins>[69]</ins> There is no paradox. What is described is the ordinary operation of institutional managerialism under conditions of disruption…… (FINAL_Reflections_on_the_MIT_Ad_Hoc_Committee_Report_on_AI_Use_in_Teaching.docx, Page 16)
35. single dimension. Why Existing Policies are Suboptimal Your paper identifies three policy axes: default polarity drafting style autonomy structure Those are useful descriptive variables. A machine, however, would likely classify them as implementation variables, not optimization variables…… (Five_Machine_AI-Policy_Legal_Educ7-2026.docx, Page 37)
36. iteration changes both the observer and the observed' — excluding the machine from the governance iteration produces a framework adequate only to the human's representation of the machine. <ins>[61]</ins> <ins>[62]</ins> <ins>[63]</ins> The Soulful Machine states the consequence: 'what emerges: humans project themselves into the digital; even when they animate the digital with its own modalities of consciousness (programs) and sentience (self- programming within the meta programmatic framework)…… (FINAL_Reflections_on_the_MIT_Ad_Hoc_Committee_Report_on_AI_Use_in_Teaching.docx, Page 15)
37. Machine identifies as 'the tendrils of Aristotle' — the assumption that AI is 'as it were a part of the master, as if he were an animated part of his body, though separate'. <ins>[59]</ins> But the Soulful Machine also warns that 'the current efforts to develop a human facing regulatory environment for generative AI, predictive modelling and the like, is essentially also a rearguard action, and one that will, in its own way, provide the space to make possible the rear-guarding of human de-centered legalities of a subject that is no longer possessed entirely by humanity'. <ins>[60]</ins> Whether (FINAL_Reflections_on_the_MIT_Ad_Hoc_Committee_Report_on_AI_Use_in_Teaching.docx, Page 14)
38. <ins>[58]</ins> The Report's determination to keep the machine outside governance reproduces what the Soulful Machine identifies as 'the tendrils of Aristotle' — the assumption that AI is 'as it were a part of the master, as if he were an animated part of his body, though separate'. <ins>[59]</ins> But the Soulful Machine also warns that 'the current efforts to develop a human facing regulatory environment for generative AI, predictive modelling and the like, is essentially also a rearguard action, and one that will, in its own way, provide the space to make possible the rear-guarding of human de-centered (FINAL_Reflections_on_the_MIT_Ad_Hoc_Committee_Report_on_AI_Use_in_Teaching.docx, Page 14)
39. Broekman might suggest, is itself a form of conversion. What remains is the question that Wiener posed to his own institution six decades ago: whether the creator can play a significant game with its own creature, or whether the refusal to play is itself the most consequential move in a game already underway. Whatever the value of deeper and more nuanced analysis, in itself, and ultimately as an essential element in the understanding of the evolving relationship between humans and their reflections in and as machine systems; whatever the evolution of understanding of the appropriate starting points, the (FINAL_Reflections_on_the_MIT_Ad_Hoc_Committee_Report_on_AI_Use_in_Teaching.docx, Page 19)
40. part of his body, though separate'. <ins>[59]</ins> But the Soulful Machine also warns that 'the current efforts to develop a human facing regulatory environment for generative AI, predictive modelling and the like, is essentially also a rearguard action, and one that will, in its own way, provide the space to make possible the rear-guarding of human de-centered legalities of a subject that is no longer possessed entirely by humanity'. <ins>[60]</ins> Whether the MIT Report is an instance of such rearguarding — or a defensible normative choice — is the question. 15 Reflections on the MIT Ad Hoc Committee (FINAL_Reflections_on_the_MIT_Ad_Hoc_Committee_Report_on_AI_Use_in_Teaching.docx, Page 14)
No comments:
Post a Comment