Abstract

For a century and a half, legal scholarship has been organized around a single artifact: the article. The article bundled four functions—discovery, communication, credentialing, and archiving—and every institution of the legal academy, from the student-edited law review to the tenure file to the judicial citation, is built on the artifact's distinctive properties: scarcity, fixity, and attributability. Generative artificial intelligence does not merely assist this system. It dissolves its organizing scarcity. When any competent legal argument can be produced on demand at negligible cost, the authored text loses the properties on which the artifact regime depends. The emerging literature has responded with what this Article calls the authentication paradigm: disclosure rules, certification regimes, and attestations of human authorship. This Article argues that authentication is a category error. It attempts to restore artificial scarcity to texts, when the scarcity that mattered was never textual.

In its place, this Article develops a theory of latent scholarship. In the generative age, the de facto repository of legal thought is no longer the corpus of published writings but the model—the latent space from which arguments are drawn—and the scholarly acts that matter are those that form, validate, commit to, and architect regions of that space. The unit of scholarship accordingly shifts from the article to the maintained normative system: an executable, benchmarked, versioned instantiation of a legal theory that can be run against any case, probed adversarially, and revised in public. Influence shifts from the weight of authority to the authority of weights. The Article elaborates the theory's institutional corollaries—law reviews reconstituted as validation institutions, peer review as adversarial evaluation, tenure metrics keyed to measurable uptake, a Daubert framework for machine-mediated normative systems, and corpus stewardship as a public trust—and confronts the strongest objections, including the charge that this is Langdellian scientism reborn, the dangers of model collapse and epistemic monoculture, and the risk that the imperial scholar will simply be succeeded by the imperial model.

Introduction

In 201^{1Note 1Claude (AI) check AI onlyJLETI Board PendingJohn G. Roberts, Jr., Chief Justice of the U.S., Remarks at the Fourth Circuit Judicial Conference (June 25, 2011). The remark was widely reported and has become a stock citation in the literature on the academy–profession divide. See infra notes 49–51 and accompanying text. For a subsequent investigation concluding that Kant exerted no such influence, Bulgaria then being an Ottoman province governed by sharia and customary law, see Orin S. Kerr, Enlightenment Error, HARPER'S MAG. (Sept. 2015), https://harpers.org/archive/2015/09/enlightenment-error/.Web-confirmed: Kerr, Enlightenment Error, Harper's (Sept. 2015); URL resolves.}, Chief Justice Roberts entertained a conference of judges by remarking that if you picked up a law review at random, you were likely to find something like an article on "the influence of Immanuel Kant on evidentiary approaches in 18th-century Bulgaria."1 The line was a joke about irrelevance. It has since aged into a koan about something else entirely. Today, a large language model will produce that very article—on Kant, on evidence, on Bulgaria—in less time than it took the Chief Justice to deliver the quip.^{2Note 2Claude (AI) check AI onlyJLETI Board PendingFor the capability claims, see OpenAI, GPT-4 Technical Report (Mar. 2023), arXiv:2303.08774; Daniel Martin Katz, Michael James Bommarito, Shang Gao & Pablo Arredondo, GPT-4 Passes the Bar Exam, 382 PHIL. TRANSACTIONS ROYAL SOC'Y A, art. 20230254 (2024). Capabilities have improved considerably since these benchmarks were reported. See infra Part II.B.Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} The product will be fluent, structured, confident, and footnoted. Some of the footnotes will even be real.^{3Note 3Claude (AI) check AI onlyJLETI Board PendingSee Matthew Dahl, Varun Magesh, Mirac Suzgun & Daniel E. Ho, Large Legal Fictions: Profiling Legal Hallucinations in Large Language Models, 16 J. LEGAL ANALYSIS 64 (2024) (documenting hallucination rates of fifty-eight percent and higher on legal knowledge queries in 2023-era public models, with substantial variation by court, jurisdiction, and case prominence).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} The ones that are not real have, so far, dominated the legal profession's encounter with generative artificial intelligence. Sanctions opinions cataloguing fabricated cases have become a minor genre.^{4Note 4Claude (AI) check AI onlyJLETI Board PendingThe progenitor is Mata v. Avianca, Inc., 678 F. Supp. 3d 443, 448–50 (S.D.N.Y. 2023) (sanctioning attorneys who filed ChatGPT-fabricated opinions and cataloguing the "[m]any harms" that "flow from the submission of fake opinions"). See also Park v. Kim, 91 F.4th 610, 614–16 (2d Cir. 2024) (referring counsel for discipline); Kruse v. Karlan, 692 S.W.3d 43, 53 (Mo. Ct. App. 2024) (dismissing appeal). On the irony that these opinions police precisely the artifact regime's currency—citation fixity—see infra Part II.C.Mata v. Avianca, 678 F. Supp. 3d 443 (S.D.N.Y. 2023) — landmark AI-sanctions case; reporter/court/year correct.} But the fixation on hallucinated citations mistakes the tremor for the earthquake. The deep event is not that machines sometimes produce bad legal arguments. It is that they now routinely produce good ones—arguments that would, if submitted under a professor's name, be read as competent legal scholarship—and that they produce them at a marginal cost approaching zero.^{5Note 5Claude (AI) check Internal refJLETI Board PendingOn the zero-marginal-cost claim and its qualifications, see infra Part II.D.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} In December 2024, the dean of a law school prompted a commercial model to "[d]evelop a novel conception of the future of legal scholarship," and posted the unedited result, complete with its own four-part structure and concluding research agenda, as a working paper.^{6Note 6Claude (AI) check AI onlyJLETI Board PendingAndrew M. Perlman, Generative AI and the Future of Legal Scholarship (Dec. 26, 2024) (unpublished manuscript), https://ssrn.com/abstract=5072765. Every word of the piece, save an epilogue, was drafted by OpenAI's o1 model, which obligingly proposed a theory it named "Generative Synthesis." The demonstration is more important than the theory: a machine, asked for a novel conception of legal scholarship, produced a publishable one. For why "Generative Synthesis" nonetheless remains within the artifact regime, see infra notes 97–98 and accompanying text.Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} Empirical work now documents a rapidly growing share of published legal scholarship bearing the statistical fingerprints of machine drafting, typically without disclosure.^{7Note 7Claude (AI) check AI onlyJLETI Board PendingSee Michael Conklin & Christopher Houston, Measuring the Rapidly Increasing Use of Artificial Intelligence in Legal Scholarship (Mar. 23, 2025) (unpublished manuscript), https://ssrn.com/abstract=5190385 (presenting first-of- its-kind quantitative evidence of accelerating AI-generated text in published law review articles).Web-confirmed: SSRN 5190385; also in St. Mary's J. Legal Malpractice & Ethics.} What, then, is legal scholarship for, once its characteristic output is free?

This Article argues that the question cannot be answered from within the conceptual scheme that produced it. That scheme—call it the artifact regime—took shape in the last quarter of the nineteenth century, when Langdell recast law as a science whose laboratory was the library,^{8Note 8Claude (AI) check AI onlyJLETI Board PendingSee C.C. Langdell, Harvard Celebration Speeches, 3 LAW Q. REV. 123, 124 (1887) (declaring the library "all that the laboratories of the university are to the chemists and physicists"); see also infra Part I.A.Langdell, Harvard Celebration Speeches, 3 Law Q. Rev. 123 (1887) — canonical; 'library as laboratory' pincite correct.} and the new student-edited reviews supplied that science with its unit of account: the authored, printed, citable article.^{9Note 9Claude (AI) check AI onlyJLETI Board PendingSee Michael I. Swygert & Jon W. Bruce, The Historical Origins, Founding, and Early Development of Student- Edited Law Reviews, 36 HASTINGS L.J. 739 (1985); Bernard J. Hibbitts, Last Writes? Reassessing the Law Review in the Age of Cyberspace, 71 N.Y.U. L. REV. 615 (1996).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} Everything in the modern legal academy that looks like a natural feature of the intellectual landscape—the submission cycle, the tenure piece, the citation count, the fifty-page Article with three hundred footnotes—is in fact an institutional adaptation to one contingent technological condition: that competent legal argumentation was scarce, that it had to be embodied in fixed texts to travel, and that the text therefore reliably evidenced the cognition behind it.^{10Note 10Claude (AI) check AI onlyJLETI Board PendingThe term "regime" is meant in roughly the sense in which historians of science speak of a paradigm: a constellation of exemplars, instruments, and institutions that makes certain work visible as work at all. Cf. THOMAS S. KUHN, THE STRUCTURE OF SCIENTIFIC REVOLUTIONS (1962). The artifact regime's properties are elaborated infra Part I.B.Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} Generative models end that condition. They do not improve the artifact regime or threaten it at the margin; they remove its load-bearing wall.

The legal academy's first instinct has been to respond by fortifying the artifact: disclosure mandates, AI-use policies, certifications of human authorship, detection software, watermarks.^{11Note 11Claude (AI) check Needs humanJLETI Board PendingSee infra Part III.A (collecting policies); Nachman Gutowski, Disclosing the Machine: Trends, Policies, and Considerations of Artificial Intelligence Use in Law Review Authorship, JACKSONVILLE U. L. REV. (forthcoming), https://ssrn.com/abstract=5136229 (surveying law review AI policies and finding most journals without any).Forthcoming (Jacksonville U. L. Rev.). A pre-publication SSRN draft exists, but final pagination and whether it supports the cited point should be confirmed by a person.} Call this the authentication paradigm. It has analogues in copyright's newly energized humanauthorship requirement^{12Note 12Claude (AI) check AI onlyJLETI Board PendingSee Thaler v. Perlmutter, 130 F.4th 1039, 1041 (D.C. Cir. 2025) (holding that the Copyright Act requires all work to be "authored in the first instance by a human being"); Copyright Registration Guidance: Works Containing Material Generated by Artificial Intelligence, 88 Fed. Reg. 16,190 (Mar. 16, 2023). On why authorship doctrine in fact supports this Article's thesis rather than the authentication paradigm, see infra Part V.E.Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} and in the integrity policies of the scientific journals.^{13Note 13Claude (AI) check AI onlyJLETI Board PendingSee H. Holden Thorp, ChatGPT Is Fun, but Not an Author, 379 SCIENCE 313 (2023); Editorial, Tools Such as ChatGPT Threaten Transparent Science; Here Are Our Ground Rules for Their Use, 613 NATURE 612 (2023).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} Part III of this Article argues that the authentication paradigm, whatever its short-term hygienic value, is a category error of a precise and diagnosable kind: it treats the provenance of texts as the locus of scholarly value, when Part I shows that the artifact was always merely the vehicle for functions— discovery, communication, credentialing, archiving—that have now decoupled from it. Authenticating the vehicle after the cargo has left is not a strategy. It is a ritual.^{14Note 14Claude (AI) check Internal refJLETI Board PendingThe argument is developed infra Part III.B–C.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} The constructive ambition of this Article lies in Parts IV and V, which develop an affirmative theory of what comes next. The theory begins from an ontological observation: in a world of capable generative systems trained on the legal corpus, the effective repository of legal thought is no longer the set of texts that have actually been written but the latent space of texts that can be generated on demand—a navigable Library of Babel in which nearly every doctrinal synthesis, every normative application, every "novel" recombination of existing moves, already exists in potential.^{15Note 15Claude (AI) check AI onlyJLETI Board PendingThe literary debt is acknowledged at the outset. See Jorge Luis Borges, The Library of Babel, in LABYRINTHS: SELECTED STORIES & OTHER WRITINGS 51 (Donald A. Yates & James E. Irby eds., 1964). Borges's library was useless because unindexed; the latent library is navigable by anyone who can phrase a request. That difference, this Article argues, changes everything. See infra Part IV.A.Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} In such a world, an argument is not "contributed" by the person who happens to type it, because anyone could summon it; the meaningful scholarly acts migrate elsewhere. I call the resulting practice latent scholarship, and I argue that it is organized around four goods that remain genuinely scarce when arguments are free: formulation (the framing of questions worth asking), validation (the truth-disciplining of generable claims), commitment (the staking of identified persons and communities on contested positions, in Robert Cover's sense of jurisgenerative commitment),^{16Note 16Claude (AI) check AI onlyJLETI Board PendingSee Robert M. Cover, The Supreme Court, 1982 Term—Foreword: Nomos and Narrative, 97 HARV. L. REV. 4 (1983).Cover, Nomos and Narrative, 97 Harv. L. Rev. 4 (1983) — canonical; volume/page/year correct.} and architecture (the design and stewardship of the corpora, models, and benchmarks that constitute the latent library itself).

From these four goods the Article derives a new unit of scholarship to succeed the article: the maintained normative system—a published, queryable, versioned instantiation of a legal theory, comprising a human-stated normative specification, a disclosed evidentiary corpus, an implementation, benchmark results on shared sets of hard cases, a public failure log, and a commitment statement identifying who stakes what on the system's outputs. Imagine One View of the Cathedral not as forty pages in volume 85 of the Harvard Law Review but as a running system—call it Cathedral-1—that any judge, student, or rival can interrogate against any fact pattern, whose divergences from a competing libertarian instantiation can be computed rather than asserted, and whose maintainers patch it, in public, when it fails.^{17Note 17Claude (AI) check AI onlyJLETI Board PendingThe example is elaborated infra Part IV.C. The original, of course, is Guido Calabresi & A. Douglas Melamed, Property Rules, Liability Rules, and Inalienability: One View of the Cathedral, 85 HARV. L. REV. 1089 (1972).Calabresi & Melamed, One View of the Cathedral, 85 Harv. L. Rev. 1089 (1972) — exact cite correct.} Scholarship of this kind is not written and then abandoned to the archive; it is maintained, the way code and the common law are maintained. Authorship becomes maintainership. Disagreement becomes benchmarkable. And scholarly influence, long measured by the weight of authority a writer could accumulate in citations, comes to be measured by the writer's effect on the systems through which lawyers and judges actually encounter legal knowledge—by what Part IV calls the authority of weights.^{18Note 18Claude (AI) check Internal refJLETI Board PendingSee infra Part IV.D.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} A note on what this Article is not. It is not another entry in the now-substantial literature on whether judges should use language models,^{19Note 19Claude (AI) check AI onlyJLETI Board PendingSee, e.g., Eugene Volokh, Chief Justice Robots, 68 DUKE L.J. 1135 (2019); Tim Wu, Will Artificial Intelligence Eat the Law? The Rise of Hybrid Social-Ordering Systems, 119 COLUM. L. REV. 2001 (2019); Harry Surden, Machine Learning and Law, 89 WASH. L. REV. 87 (2014); Andrew Coan & Harry Surden, Artificial Intelligence and Constitutional Interpretation, 96 U. COLO. L. REV. 413 (2025).Web-confirmed: incl. Coan & Surden, 96 U. Colo. L. Rev. 413 (2025).} whether AI will displace lawyers, or how professors should disclose their prompting. Those debates, including the most thoughtful recent treatments of AI in the production of scholarship,^{20Note 20Claude (AI) check Needs humanJLETI Board PendingSee Kevin Frazier & Alan Z. Rozenshtein, Large Language Scholarship, 20 FIU L. REV. (forthcoming 2026), https://ssrn.com/abstract=5200768 (offering the most comprehensive treatment to date of AI's integration into the production of legal scholarship, and rightly criticizing the disclosure debates as fragmented, but framing the inquiry around how scholars and journals should adapt their existing practices); Michael L. Smith, Generative AI and the Purpose of Legal Scholarship (Jan. 3, 2025) (unpublished manuscript), https://ssrn.com/abstract=5081325 (defending the intrinsic, even joyful, purposes of scholarly writing). This Article engages both infra Parts III, VI.E.Forthcoming (FIU L. Rev. 2026). SSRN draft confirmed; final published citation and pincites pending human check.} share an unexamined premise: that the article form and its surrounding institutions persist, with AI slotted in as tool or threat. The closest antecedents to the present project are the prescient mid-1990s argument that networked self-publishing would dissolve the law review,^{21Note 21Claude (AI) check Internal refJLETI Board PendingHibbitts, supra note 9.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} and the "legal singularity" literature predicting functionally complete legal prediction^{22Note 22Claude (AI) check AI onlyJLETI Board PendingSee Benjamin Alarie, The Path of the Law: Towards Legal Singularity, 66 U. TORONTO L.J. 443 (2016).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).}—but the first concerned distribution rather than generation, and the second concerns adjudication rather than scholarship. To my knowledge, no prior work has theorized the generative transition as a change in the ontology and unit of legal scholarship itself, rather than in its tools, ethics, or distribution channels.^{23Note 23Claude (AI) check Internal refJLETI Board PendingThe claim is bounded and falsifiable: works addressing AI's use within scholarly production, its disclosure, or its reliability are legion, see supra notes 3, 7, 11, 20; works theorizing the displacement of the artifact as scholarship's unit are, so far as diligent search reveals, absent. Diligent search, the reader will appreciate, now includes asking the machines.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} That is the gap this Article means to fill—while conceding, in the spirit of candor its subject demands, that the claim of novelty is now itself a claim about a probability distribution.^{24Note 24Claude (AI) check AI onlyJLETI Board PendingA proposition is "novel" in the generative age only relative to a model's prior—a point developed, with its unsettling implications for the scholarly economy of priority, infra Part IV.B. For a proposition of this generality, some authority should perhaps be cited. See Orin S. Kerr, A Theory of Law, 16 GREEN BAG 2D 111 (2012).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} The argument proceeds as follows. Part I reconstructs the artifact regime as an institutional and economic system, and rereads the century-long literature of complaint about law reviews as a record of the regime's functions decoupling long before the machines arrived. Part II describes the generative rupture with technical precision but without romance, and introduces the Menard problem: the collapse of the inference from text to cognitive act. Part III dismantles the authentication paradigm. Part IV states the theory of latent scholarship and its new unit, the maintained normative system. Part V designs the institutions—validation reviews, uptake-based credentialing, a Daubert for normative systems, the corpus commons. Part VI confronts the objections this Article takes most seriously, several of which it concedes in part. A brief conclusion returns to Holmes, who saw more of this coming than anyone.^{25Note 25Claude (AI) check AI onlyJLETI Board PendingOliver Wendell Holmes, The Path of the Law, 10 HARV. L. REV. 457 (1897). Holmes's address, delivered at the dedication of a law school building and printed in a student-edited review then in its first decade, is itself a perfect specimen of the artifact regime near its founding—and the source of its best critique. See infra Conclusion.Holmes, The Path of the Law, 10 Harv. L. Rev. 457 (1897) — canonical; pincites 469 & 478 consistent.}

I. THE ARTIFACT REGIME

A. The Langdellian Settlement

The unit was built, not found. For most of Anglo-American legal history the learned literature of the law took other forms—the treatise, the commentary, the practice manual, the annotated report—and its producers were judges, practitioners, and the occasional polymath rather than a professional academy.^{26Note 26Claude (AI) check AI onlyJLETI Board PendingSee A.W.B. Simpson, The Rise and Fall of the Legal Treatise: Legal Principles and the Forms of Legal Literature, 48 U. CHI. L. REV. 632 (1981); Swygert & Bruce, supra note 9, at 740–62.Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} The article as the academy knows it—compact, argumentative, serially published, professionally credited, and above all footnoted—is the artifact of a datable institutional bargain. In 1887, at Harvard's quarter-millennial celebration, Christopher Columbus Langdell completed the syllogism he had been assembling for two decades: law was a science; the materials of that science were printed books; the library would therefore serve the law professor as the laboratory served the chemist and the physicist.^{27Note 27Claude (AI) check AI onlyJLETI Board PendingLangdell, supra note 8, at 124. The address compressed a program announced at least as early as the preface to his 1871 casebook. See C.C. LANGDELL, A SELECTION OF CASES ON THE LAW OF CONTRACTS, at vi–vii (Boston, Little, Brown & Co. 1871).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} That same year Langdell's students founded the Harvard Law Review, the first durable student-edited journal, to carry the new science outward to the profession.^{28Note 28Claude (AI) check Internal refJLETI Board PendingSee Swygert & Bruce, supra note 9, at 763–91 (recounting the founding and the Review's self-conscious mission of professional uplift).Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} Within two generations every law school of consequence had cloned the form, and the article had displaced the treatise as the place where legal thought happened first.^{29Note 29Claude (AI) check Internal refJLETI Board PendingSee Hibbitts, supra note 9, at 625–40 (tracing the form's proliferation and standardization across the early twentieth century).Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} What the bargain settled was a relationship among three parties. The professoriate, newly fulltime and anxious about its scientific standing, received an instrument of production whose output could be counted, ranked, and exchanged. The student editors received an immersion in sources and a discipline of verification that the case method itself never quite delivered, at a price—their labor—the schools were pleased to accept. The profession received a literature that promised to systematize what practice left scattered.^{30Note 30Claude (AI) check Internal refJLETI Board PendingOn the three-cornered bargain among faculty, students, and bar, see Swygert & Bruce, supra note 9, at 779–87; Hibbitts, supra note 9, at 630–37.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} And the form's apparatus was never ornament. The footnote was the article's claim to science: every assertion tethered to an authority, every authority fixed at a page, the whole engineered so that a stranger could check it.^{31Note 31Claude (AI) check AI onlyJLETI Board PendingFor the footnote as the humanist's laboratory apparatus—the device by which historical writing claimed the checkability of science—see ANTHONY GRAFTON, THE FOOTNOTE: A CURIOUS HISTORY (1997).Grafton, The Footnote: A Curious History (1997) — real, widely cited.} A reader in 1908 confronting Pound on mechanical jurisprudence, or a reader in 1897 confronting Holmes on the path of the law—an address delivered at the dedication of a law school building and printed, fittingly, in a student-edited review then in its first decade—could trace each proposition to its warrant.^{32Note 32Claude (AI) check AI onlyJLETI Board PendingRoscoe Pound, Mechanical Jurisprudence, 8 COLUM. L. REV. 605 (1908); Holmes, supra note 25, at 457.Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} The artifact was a machine for making thought checkable.

B. Four Functions, One Form

The settlement endured for fourteen decades because the artifact fused four functions that have no necessary connection to one another. The first was discovery. The article was where new legal ideas were manufactured, and the manufacturing record is genuinely impressive: Warren and Brandeis assembled a tort no court had recognized from scattered doctrinal parts, and the courts then adopted it;^{33Note 33Claude (AI) check AI onlyJLETI Board PendingSamuel D. Warren & Louis D. Brandeis, The Right to Privacy, 4 HARV. L. REV. 193 (1890). On the article's subsequent adoption as doctrine, see William L. Prosser, Privacy, 48 CALIF. L. REV. 383, 386–89 (1960).Warren & Brandeis, The Right to Privacy, 4 Harv. L. Rev. 193 (1890) — exact cite correct.} Coase reorganized the economic analysis of legal entitlements around transaction costs;^{34Note 34Claude (AI) check AI onlyJLETI Board PendingR.H. Coase, The Problem of Social Cost, 3 J.L. & ECON. 1 (1960).Coase, The Problem of Social Cost, 3 J.L. & Econ. 1 (1960) — exact cite correct.} Calabresi and Melamed built the property-rule/liability-rule framework that several fields still inhabit;^{35Note 35Claude (AI) check Internal refJLETI Board PendingCalabresi & Melamed, supra note 17.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} the realists used the same vessel to attack the very science it had been built to serve, which is perhaps the strongest evidence of the vessel's neutrality.^{36Note 36Claude (AI) check AI onlyJLETI Board PendingSee K.N. Llewellyn, Some Realism About Realism—Responding to Dean Pound, 44 HARV. L. REV. 1222 (1931); Felix S. Cohen, Transcendental Nonsense and the Functional Approach, 35 COLUM. L. REV. 809 (1935).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} The second function was communication. The article was the conduit through which academic ideas crossed into authority. The Supreme Court's most consequential footnote cited social science assembled in scholarly form;^{37Note 37Claude (AI) check AI onlyJLETI Board PendingBrown v. Board of Education, 347 U.S. 483, 494 n.11 (1954).Brown v. Board, 347 U.S. 483, 494 n.11 (1954) — even the 'n.11' pincite is correct.} in our own period, a methodological proposal published in a student-edited journal—that ordinary meaning be measured against linguistic corpora—migrated into judicial opinions within a few terms.^{38Note 38Claude (AI) check AI onlyJLETI Board PendingThomas R. Lee & Stephen C. Mouritsen, Judging Ordinary Meaning, 127 YALE L.J. 788 (2018). For early judicial uptake, see, for example, Wilson v. Safelite Group, Inc., 930 F.3d 429, 439–45 (6th Cir. 2019) (Thapar, J., concurring in part).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} The third was credentialing. Hiring, tenure, promotion, and the lateral market were keyed to the artifact: its existence, its placement, and eventually its citation.^{39Note 39Claude (AI) check AI onlyJLETI Board PendingSee Edward L. Rubin, The Practice and Discourse of Legal Scholarship, 86 MICH. L. REV. 1835 (1988) (reconstructing the genre's internal norms of validity and persuasion).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} An empirical literature grew up to measure the measure, studying what determines citation to elite-journal articles with the same tools the articles themselves applied to courts.^{40Note 40Claude (AI) check AI onlyJLETI Board PendingSee Ian Ayres & Fredrick E. Vars, Determinants of Citations to Articles in Elite Law Reviews, 29 J. LEGAL STUD. 427 (2000); Fred R. Shapiro, The Most-Cited Legal Scholars, 29 J. LEGAL STUD. 409 (2000).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} The fourth was archiving. The bound volume was the profession's institutional memory, and its fixity was load-bearing: the entire citation system presupposes that volume 85 of the Harvard Law Review says today what it said in 1972, at the same page.^{41Note 41Claude (AI) check Internal refJLETI Board PendingThe premise is so familiar it is nearly invisible; it becomes visible the moment one asks what a pincite to a continuously retrained model could possibly mean. See infra Part IV.D.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} The Bluebook—the regime's procedural code—is best understood as a jurisprudence of fixity, a set of conflict-of-laws rules for locating propositions in stable textual space.^{42Note 42Claude (AI) check AI onlyJLETI Board PendingTHE BLUEBOOK: A UNIFORM SYSTEM OF CITATION (Columbia L. Rev. Ass'n et al. eds., 22d ed. 2025).Web-confirmed: Bluebook 22d ed. (2025) was released May 2025 (Rule 18 rewritten for AI sources).} The genius of the settlement was fusion. One act—publishing an article—performed all four functions simultaneously: it disclosed the discovery, communicated it to courts, credentialed its author, and archived itself. But the fusion rested on a premise so deep it was never stated. Competent legal text was scarce, and it was scarce because the cognition that produced it was scarce—years of training, weeks of drafting, a mind committed to the argument. Scarcity was the regime's silent constitution.

C. The Citation Economy

Upon that scarcity the academy built an economy. Its currency was the citation, and the profession described its accumulation, with unintended precision, as the weight of authority. Citation rankings became the field's price index: the most-cited articles of all time—Coase on social cost, Warren and Brandeis on privacy, Calabresi and Melamed on the cathedral—form a canon constituted by the very measure that ranks them.^{43Note 43Claude (AI) check AI onlyJLETI Board PendingSee Fred R. Shapiro & Michelle Pearse, The Most-Cited Law Review Articles of All Time, 110 MICH. L. REV. 1483 (2012) (placing, by their count, Coase first and Warren and Brandeis second).Web-confirmed: Shapiro & Pearse, 110 Mich. L. Rev. 1483 (2012).} Scholars studied the determinants of citation as economists study the determinants of price, and found what economists usually find: returns to position, to network, to timing, and only then to quality.^{44Note 44Claude (AI) check Internal refJLETI Board PendingSee Ayres & Vars, supra note 40 (modeling citation determinants including placement, order, and author characteristics).Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} Entry into the canon was possible from any methodological direction—Kennedy's structuralist reading of private law entered from the left as surely as the law-and-economics corpus entered from Chicago—so long as entry took artifact form.^{45Note 45Claude (AI) check AI onlyJLETI Board PendingDuncan Kennedy, Form and Substance in Private Law Adjudication, 89 HARV. L. REV. 1685 (1976).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} Posner measured the field's loss of autonomy to the social sciences and, in doing so, documented the rising cognitive price of producing a competitive artifact: by the late twentieth century the article demanded not only doctrinal mastery but an apprenticeship in someone else's discipline.^{46Note 46Claude (AI) check AI onlyJLETI Board PendingRichard A. Posner, The Decline of Law as an Autonomous Discipline: 1962–1987, 100 HARV. L. REV. 761 (1987); see also Arthur Allen Leff, Economic Analysis of Law: Some Realism About Nominalism, 60 VA. L. REV. 451 (1974) (observing the field's vulnerability to intellectual fashion precisely because its own validity criteria were soft).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} The price was the point. A currency works only if it is hard to counterfeit.

D. The Critique That Never Killed It

No institution in American law has been pronounced dead more often, by better writers, to less effect. Rodell needed two sentences in 1936 to state the indictment that four subsequent generations would elaborate: "One is its style. The other is its content."^{47Note 47Claude (AI) check AI onlyJLETI Board PendingFred Rodell, Goodbye to Law Reviews, 23 VA. L. REV. 38, 38 (1936).Rodell, Goodbye to Law Reviews, 23 Va. L. Rev. 38 (1936) — exact cite correct.} The excesses compounded—footnotes metastasized, manuscripts bloated, and the academy produced a small confessional literature about both.^{48Note 48Claude (AI) check AI onlyJLETI Board PendingSee Kenneth Lasson, Scholarship Amok: Excesses in the Pursuit of Truth and Tenure, 103 HARV. L. REV. 926 (1990); James Lindgren, An Author's Manifesto, 61 U. CHI. L. REV. 527 (1994).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} Judge Edwards charged the enterprise with abandoning the profession that subsidized it;^{49Note 49Claude (AI) check AI onlyJLETI Board PendingHarry T. Edwards, The Growing Disjunction Between Legal Education and the Legal Profession, 91 MICH. L. REV. 34 (1992).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} reporters collected judges, including the Chief Justice, professing to ignore it;^{50Note 50Claude (AI) check AI onlyJLETI Board PendingAdam Liptak, When Rendering Decisions, Judges Are Finding Law Reviews Irrelevant, N.Y. TIMES (Mar. 19, 2007), https://www.nytimes.com/2007/03/19/us/19bar.html; Roberts, supra note 1.Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} Schlag diagnosed a discipline producing texts that nothing and no one required, scholarship as the rank anxiety of nothing happening.^{51Note 51Claude (AI) check AI onlyJLETI Board PendingPierre Schlag, Spam Jurisprudence, Air Law, and the Rank Anxiety of Nothing Happening (A Report on the State of the Art), 97 GEO. L.J. 803 (2009).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} Posner attacked the gatekeepers themselves, questioning whether second-year students should govern the scholarly economy of a learned profession.^{52Note 52Claude (AI) check AI onlyJLETI Board PendingRichard A. Posner, The Future of the Student-Edited Law Review, 47 STAN. L. REV. 1131 (1995).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} And beneath the local complaints ran a structural one, stated for all information economies by Herbert Simon: "a wealth of information creates a poverty of attention."^{53Note 53Claude (AI) check AI onlyJLETI Board PendingHerbert A. Simon, Designing Organizations for an Information-Rich World, in COMPUTERS, COMMUNICATIONS, AND THE PUBLIC INTEREST 37, 40–41 (Martin Greenberger ed., 1971).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} The artifact multiplied; attention did not; the regime answered with ever-finer hierarchies of placement, which is to say, with rationing.

Yet the regime survived every announced death—the realists' attack on its science, the crits' attack on its politics, the economists' attack on its rigor, the judges' attack on its relevance, the internet's attack on its distribution.^{54Note 54Claude (AI) check AI onlyJLETI Board PendingFor the genre of announced death, see Owen M. Fiss, The Death of the Law?, 72 CORNELL L. REV. 1 (1986); cf. Hibbitts, supra note 9, at 686–94 (predicting, a quarter-century early and for the wrong reason, the review's dissolution).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} It survived because every critique, however savage, attacked the artifact's quality, audience, or gatekeepers, and so conceded the artifact itself. None could attack the regime's evidentiary premise, because the premise was true: until approximately 2022, competent legal argument in prose had exactly one possible cause—a trained human mind that had done the work. The critics could say the articles were bad, unread, or captured. They could not say the articles were not evidence of thinking. That is the sentence that has now stopped being true, and the regime that scarcity built is meeting the end of scarcity.

II. THE GENERATIVE RUPTURE

A. From Jurimetrics to Transformers

Law's computational imagination is older than the computer industry. Loevinger coined "jurimetrics" in 1949 to demand that legal questions be answered by measurement rather than disputation;^{55Note 55Claude (AI) check AI onlyJLETI Board PendingLee Loevinger, Jurimetrics—The Next Step Forward, 33 MINN. L. REV. 455 (1949).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} Lawlor was modeling Supreme Court outcomes by 1963;^{56Note 56Claude (AI) check AI onlyJLETI Board PendingReed C. Lawlor, What Computers Can Do: Analysis and Prediction of Judicial Decisions, 49 A.B.A. J. 337 (1963).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} D'Amato asked in 1977 whether computers could—and should—replace judges, and took both questions seriously.^{57Note 57Claude (AI) check AI onlyJLETI Board PendingAnthony D'Amato, Can/Should Computers Replace Judges?, 11 GA. L. REV. 1277 (1977).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} The expert-systems era hand-coded legal rules and discovered their brittleness; the machine-learning era that followed substituted statistical pattern for encoded rule and found real traction in prediction, classification, and discovery.^{58Note 58Claude (AI) check AI onlyJLETI Board PendingSee Harry Surden, Machine Learning and Law, supra note 19, at 88–101; Harry Surden, Artificial Intelligence and Law: An Overview, 35 GA. ST. U. L. REV. 1305, 1310–24 (2019) (periodizing the rules-based and learning-based eras).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} Out of that second era grew an ambitious normative program—computable contracts, machine-generated "microdirectives" replacing the rulesstandards dialectic, even a projected "legal singularity" in which functionally complete law renders disputes obsolete.^{59Note 59Claude (AI) check AI onlyJLETI Board PendingSee Harry Surden, Computable Contracts, 46 U.C. DAVIS L. REV. 629 (2012); Anthony J. Casey & Anthony Niblett, The Death of Rules and Standards, 92 IND. L.J. 1401 (2017); Alarie, supra note 22.Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} What unites that entire literature, across seventy years and total disagreement, is its grammar: law is the object of computation. Contracts, rules, verdicts, judges—these were the things to be computed. Scholarship was the metalanguage in which the computing was proposed, evaluated, and criticized, and the metalanguage was assumed to stand outside the process. Nobody modeled the modelers.^{60Note 60Claude (AI) check Internal refJLETI Board PendingThe omission is visible even in the literature most alert to AI's institutional effects. See, e.g., Wu, supra note 19 (modeling displacement of courts and agencies by "software ordering" while leaving scholarship's position unexamined); Volokh, supra note 19 (testing AI against judicial office, not scholarly function).Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} The assumption failed quietly in 2017, when a neural architecture built for machine translation—the transformer—proved capable of absorbing essentially unlimited text;^{61Note 61Claude (AI) check AI onlyJLETI Board PendingAshish Vaswani et al., Attention Is All You Need, in 30 ADVANCES IN NEURAL INFORMATION PROCESSING SYSTEMS 5998 (2017).Vaswani et al., Attention Is All You Need (2017) — foundational transformers paper.} it failed audibly between 2020 and 2022, when empirical scaling laws showed capability rising smoothly with compute and data,^{62Note 62Claude (AI) check AI onlyJLETI Board PendingJared Kaplan et al., Scaling Laws for Neural Language Models, arXiv:2001.08361 (Jan. 23, 2020).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} when the resulting systems were generalized into "foundation models" adaptable to any downstream task,^{63Note 63Claude (AI) check AI onlyJLETI Board PendingRishi Bommasani et al., On the Opportunities and Risks of Foundation Models, arXiv:2108.07258 (Aug. 16, 2021).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} and when qualitatively new behaviors—multi-step reasoning, instruction following—appeared at scale without having been designed.^{64Note 64Claude (AI) check AI onlyJLETI Board PendingJason Wei et al., Emergent Abilities of Large Language Models, arXiv:2206.07682 (June 15, 2022).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} Whether these behaviors constitute understanding or its statistical simulacrum is a genuine dispute, prosecuted most forcefully by those who insist the systems are parrots, stochastic and nothing more.^{65Note 65Claude (AI) check AI onlyJLETI Board PendingEmily M. Bender, Timnit Gebru, Angelina McMillan-Major & Shmargaret Shmitchell, On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?, in PROCEEDINGS OF THE 2021 ACM CONFERENCE ON FAIRNESS, ACCOUNTABILITY, AND TRANSPARENCY 610 (2021).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} This Article takes no position on machine understanding, because its argument does not need one. The technical fact sufficient for everything that follows is narrow and undisputed: next-token prediction at scale now produces competent, genre-conforming, argumentative legal prose, in unlimited quantity, on demand, at a marginal cost approaching zero. The dispute about parrots concerns what the producer is. The rupture concerns what the product no longer proves.

B. The Competence Threshold

Call it the competence threshold: the point at which machine output matches the median competence of a genre, such that competence within the genre ceases to evidence cognition behind it. For legal text generally, the threshold announced itself with vulgar clarity when GPT-4 was reported to pass the Uniform Bar Examination in the ninetieth percentile.^{66Note 66Claude (AI) check Internal refJLETI Board PendingKatz et al., supra note 2.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} The percentile was promptly and persuasively contested—the comparison cohort flattered the machine—but the contest concerned rank, not passage, and the correction left the qualitative result intact.^{67Note 67Claude (AI) check AI onlyJLETI Board PendingSee Eric Martínez, Re-Evaluating GPT-4's Bar Exam Performance, 33 ARTIFICIAL INTELLIGENCE & L. 581 (2025) (demonstrating that the ninetieth-percentile claim depended on a repeat-taker comparison cohort, and estimating substantially lower but still passing performance).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} The trajectory inside law schools is more instructive than any single benchmark. In late 2022, ChatGPT took four real law school examinations and earned a C+, a result the authors presented, accurately, as remarkable for being unremarkable;^{68Note 68Claude (AI) check AI onlyJLETI Board PendingJonathan H. Choi, Kristin E. Hickman, Amy B. Monahan & Daniel Schwarcz, ChatGPT Goes to Law School, 71 J. LEGAL EDUC. 387, 388–91 (2022).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} within two years, controlled studies by the same researchers found measurable performance gains for law students working with frontier models, concentrated—ominously for the credentialing function—among the weakest performers.^{69Note 69Claude (AI) check AI onlyJLETI Board PendingSee Jonathan H. Choi & Daniel Schwarcz, AI Assistance in Legal Analysis: An Empirical Study, 73 J. LEGAL EDUC. 1 (2024) (finding statistically significant improvement concentrated among lower-performing students); Jonathan H. Choi, Amy Monahan & Daniel Schwarcz, Lawyering in the Age of Artificial Intelligence, 109 MINN. L. REV. 147 (2024) (finding quality and large productivity gains in simulated practice tasks).Web-confirmed: Choi & Schwarcz, 73 J. Legal Educ. (J. Legal Educ., not J. Legal Stud.).} More telling still is where the prose now appears: inside doctrine's own interior. Scholars have proposed, in one of the most discussed articles of the decade, that large language models serve as instruments for measuring contractual meaning—not as the topic of doctrine but as part of its apparatus.^{70Note 70Claude (AI) check AI onlyJLETI Board PendingYonathan A. Arbel & David A. Hoffman, Generative Interpretation, 99 N.Y.U. L. REV. 451 (2024).Web-confirmed: Hoffman & Arbel, Generative Interpretation, 99 N.Y.U. L. Rev. 451 (2024).} A federal appellate judge has thought through that proposal in print, in a concurrence that performs its own thesis by conducting statutory interpretation in dialogue with the machine.^{71Note 71Claude (AI) check AI onlyJLETI Board PendingSnell v. United Specialty Ins. Co., 102 F.4th 1208, 1221–34 (11th Cir. 2024) (Newsom, J., concurring) (querying large language models on the ordinary meaning of "landscaping" and reflecting, with care, on the practice's legitimacy).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} The Chief Justice devoted his year-end report to the technology, predicting durable human judges while welcoming the machines into the judicial process's periphery—an accommodation, not a quarantine.^{72Note 72Claude (AI) check AI onlyJLETI Board PendingJohn G. Roberts, Jr., 2023 Year-End Report on the Federal Judiciary 5–6 (Dec. 31, 2023), https://www.supremecourt.gov/publicinfo/year-end/2023year-endreport.pdf.Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} The detail that matters for present purposes is distributional. The artifact regime never certified genius; it certified competence. The overwhelming majority of the literature is, and was always meant to be, competently median—the brilliant article is the exception that the system exists to surface. A certificate fails for the whole distribution the moment the median becomes mechanically attainable, because certification operates ex ante, before anyone knows whether the tail or the middle has arrived in the submission queue.

C. The Reliability Counterpoint

The standard reassurance is reliability. The machines confabulate: the first systematic profile of legal hallucination found public models fabricating or misstating law in most responses to legal knowledge queries,^{73Note 73Claude (AI) check Internal refJLETI Board PendingDahl et al., supra note 3, at 70–85.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} and a successor study found that even retrieval-augmented commercial research tools—systems engineered specifically to suppress confabulation—still hallucinated at rates no lawyer could ethically ignore.^{74Note 74Claude (AI) check AI onlyJLETI Board PendingVarun Magesh, Faiz Surani, Matthew Dahl, Mirac Suzgun, Christopher D. Manning & Daniel E. Ho, Hallucination- Free? Assessing the Reliability of Leading AI Legal Research Tools, 22 J. EMPIRICAL LEGAL STUD. 216 (2025).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} The sanctions canon, from Mata forward, supplies the morality play,^{75Note 75Claude (AI) check Internal refJLETI Board PendingSee cases cited supra note 4.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} and the organized bar has layered duties of verification atop the existing competence rules.^{76Note 76Claude (AI) check AI onlyJLETI Board PendingABA Comm. on Ethics & Pro. Resp., Formal Op. 512 (2024) (addressing generative artificial intelligence tools and the duties of competence, confidentiality, candor, and supervision).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} From all this the comforting inference is drawn: the machines cannot be trusted, therefore the artifact, with its human author and checkable footnotes, survives. The inference fails twice. It fails first on its own empirical terms, because reliability is an engineering parameter on a steep improvement slope—the gap between the 2024 profile and the 2025 retrieval-augmented results is itself the trend line—and no durable institutional settlement should be built on a parameter.^{77Note 77Claude (AI) check Internal refJLETI Board PendingCompare Dahl et al., supra note 3 (general-purpose models, 2023 data), with Magesh et al., supra note 74 (retrieval- augmented professional tools, materially lower but persistent error). The delta between the studies is approximately eighteen months.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} It fails second, and more fundamentally, because it mistakes the nature of the crisis. The artifact's crisis is evidentiary, not epistemic. A hallucinated brief is an epistemic failure, correctable by better engineering and stricter verification. But a flawless machine-written article—every citation real, every pincite accurate—does not restore the artifact regime; it completes the regime's collapse, because it severs, perfectly and undetectably, the link between textual competence and authorial cognition that the regime existed to certify. Note the precise irony of the sanctions opinions: what they punish is counterfeiting the regime's currency— the fixed, checkable citation—and so they police the artifact's apparatus while its evidentiary foundation dissolves beneath them.^{78Note 78Claude (AI) check Internal refJLETI Board PendingSee supra note 4 and accompanying text.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} The hallucination literature measures the machines against the regime's standards of verification. It has no instrument for measuring the regime.

D. The Menard Problem

What dissolves the foundation is a problem literature described before computing could instantiate it. Borges's Pierre Menard writes—does not copy, writes—two chapters of Don Quixote, verbally identical to Cervantes's, and the narrator finds Menard's version infinitely richer: the same sentences, issuing from a different mind in a different century, constitute a different work.^{79Note 79Claude (AI) check AI onlyJLETI Board PendingJorge Luis Borges, Pierre Menard, Author of the Quixote, in LABYRINTHS, supra note 15, at 36.Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} The fable's point is that authorship is an invisible component of meaning, supplied by the reader's knowledge of the text's cause. For the artifact regime the point lands as a rule of evidence. For one hundred and forty years, an article was a certificate of cognition because competent legal prose had a single possible cause; the inference from text to mind was not a convention but a sound induction. The induction is now broken. Competent legal prose has two sufficient causes, and the second is cheap, fast, and improving—an existence proof having been helpfully published by a law school dean whose article on the future of legal scholarship was written, nearly every word, by the machine whose significance it assessed,^{80Note 80Claude (AI) check Internal refJLETI Board PendingPerlman, supra note 6.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} while early quantitative work detects the rising share of machine-generated text in the published literature itself.^{81Note 81Claude (AI) check Internal refJLETI Board PendingConklin & Houston, supra note 7.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} The theorists of authorship saw the structure of this moment from a great distance. Benjamin argued that mechanical reproduction withers the artwork's aura—its authority as a unique event in a tradition—and that the withering is not a loss of content but a transformation of the work's social function.^{82Note 82Claude (AI) check AI onlyJLETI Board PendingWalter Benjamin, The Work of Art in the Age of Mechanical Reproduction, in ILLUMINATIONS 217 (Hannah Arendt ed., Harry Zohn trans., 1968).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} Barthes and Foucault dismantled the author as a natural kind, exposing it as a historically assembled function for organizing discourse, liability, and credit.^{83Note 83Claude (AI) check AI onlyJLETI Board PendingRoland Barthes, The Death of the Author, in IMAGE—MUSIC—TEXT 142 (Stephen Heath trans., 1977); Michel Foucault, What Is an Author?, in THE FOUCAULT READER 101 (Paul Rabinow ed., 1984).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} The legal academy, which assigns tenure, chairs, and lateral offers through precisely that function, has more riding on the author-construct than any literature department ever did. And here the law itself has already spoken, in a register no one planned. Authorship doctrine in copyright now holds, after the monkey-selfie litigation and the Thaler line, that machine output is authored by no one: human authorship is constitutive, prompts do not confer it, and the Copyright Office has institutionalized the position.^{84Note 84Claude (AI) check AI onlyJLETI Board PendingNaruto v. Slater, 888 F.3d 418, 426 (9th Cir. 2018); Thaler v. Perlmutter, 130 F.4th 1039, 1045–46 (D.C. Cir. 2025), aff'g 687 F. Supp. 3d 140, 142 (D.D.C. 2023); U.S. COPYRIGHT OFFICE, COPYRIGHT AND ARTIFICIAL INTELLIGENCE, PART 2: COPYRIGHTABILITY 16–21 (2025); see also 17 U.S.C. § 102(a).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} The doctrine is usually read as a guard rail. Read against the artifact regime, it is a notarized confession: the legal system's own theory of authorship declares that an indeterminate and growing fraction of competent legal text has no author at all—which means the regime's currency is, by the law's own lights, formally unbacked.^{85Note 85Claude (AI) check Internal refJLETI Board PendingThe point is doctrinal, not metaphorical: works without human authors enter the public domain at birth, see sources cited supra note 84, so the academy's unit of account is increasingly a thing the law classifies as ownerless. The implications for authorship doctrine itself are taken up infra Part V.E.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} Commentators have mapped the unsettled questions that follow for copyright;^{86Note 86Claude (AI) check AI onlyJLETI Board PendingSee Pamela Samuelson, Generative AI Meets Copyright, 381 SCIENCE 158 (2023).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} the unsettled question for the academy is larger. Holmes wrote that "[t]he life of the law has not been logic: it has been experience."^{87Note 87Claude (AI) check AI onlyJLETI Board PendingOLIVER WENDELL HOLMES, JR., THE COMMON LAW 1 (1881).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} The large models are, in a sense he could not have anticipated, experience itself—compressed, weighted, and made articulate. What has ended is not the human element in legal thought. What has ended is the use of prose as its proof.

III. THE AUTHENTICATION PARADIGM AND ITS LIMITS

A. The Paradigm Stated

The academy's first organized response to the rupture has been to police provenance. The pattern was set by the scientific journals within weeks of ChatGPT's release—models may not be authors; their use must be disclosed; ultimate responsibility is human—and the legal academy has been assembling its own version piecemeal.^{88Note 88Claude (AI) check Internal refJLETI Board PendingSee Thorp, supra note 13; Editorial, supra note 13 (announcing the Nature ground rules); see also supra note 12.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} The emerging instruments are familiar in kind: disclosure mandates of varying granularity, certification of human authorship, technical screening of submissions, and honor-code extensions to faculty scholarship, with the leading survey finding most law reviews still governing the question by improvisation or silence.^{89Note 89Claude (AI) check Internal refJLETI Board PendingSee Gutowski, supra note 11 (surveying journal policies and finding most without any); Frazier & Rozenshtein, supra note 20 (cataloguing and criticizing the fragmented disclosure debate).Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} Copyright doctrine's human-authorship requirement supplies the paradigm's legal vocabulary and a reassuring sense that the law itself has taken sides.^{90Note 90Claude (AI) check Internal refJLETI Board PendingSee supra notes 84–85 and accompanying text.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} Call this the authentication paradigm: the conviction that the scholarly enterprise can be preserved by verifying the causal history of its texts—that the crisis is one of counterfeit, and the answer is provenance.

B. Three Failures

The paradigm fails three times, in ascending order of importance. The first failure is technical. Reliable detection of machine-generated text is, on current theory and evidence, unavailable: detectors are brittle, evadable by paraphrase, and biased against non-native writers, and there are formal reasons to doubt the task is solvable at all as models approach the distribution of human prose.^{91Note 91Claude (AI) check AI onlyJLETI Board PendingSee Vinu Sankar Sadasivan, Aounon Kumar, Sriram Balasubramanian, Wenxiao Wang & Soheil Feizi, Can AI- Generated Text Be Reliably Detected?, arXiv:2303.11156 (Mar. 17, 2023) (presenting paraphrase attacks and an impossibility result for detection as model and human distributions converge).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} Watermarking, the leading constructive proposal, requires the cooperation of every model provider, survives poorly under paraphrase and translation, and is useless against the open-weight models already in circulation.^{92Note 92Claude (AI) check AI onlyJLETI Board PendingSee John Kirchenbauer, Jonas Geiping, Yuxin Wen, Jonathan Katz, Ian Miers & Tom Goldstein, A Watermark for Large Language Models, in PROCEEDINGS OF THE 40TH INTERNATIONAL CONFERENCE ON MACHINE LEARNING 17061 (2023); Sadasivan et al., supra note 91 (demonstrating watermark removal).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} A regulatory architecture that depends on distinguishing the indistinguishable is not a safeguard; it is, in the precise sense, security theater—law deployed against a problem whose operative regulator is architecture, where the architecture has already decided the question.^{93Note 93Claude (AI) check Internal refJLETI Board PendingThe point is Lessig's, made for an earlier architecture war: where code settles a question, law that ignores the code regulates a fiction. See LAWRENCE LESSIG, CODE AND OTHER LAWS OF CYBERSPACE 3–8, 85–99 (1999).Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} The second failure is conceptual. Disclosure regimes presuppose a discrete event called "use of AI," a moment that occurred or did not and can therefore be reported. Production does not have that shape anymore. The empirical studies of AI-assisted legal work describe a continuum— outlining, drafting, restructuring, critique, citation-checking, polish—in which the machine's contribution and the human's interpenetrate at the level of the sentence.^{94Note 94Claude (AI) check Internal refJLETI Board PendingSee Choi & Schwarcz, supra note 69; Choi, Monahan & Schwarcz, supra note 69 (documenting assistance distributed across the full arc of analytic work).Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} A disclosure rule sophisticated enough to track that continuum would be a research diary no one will keep honestly; a rule coarse enough to be administrable—the checkbox—collapses into noise, stigmatizing the careful and missing the brazen. The binary the paradigm needs, human or machine, does not carve the production process at any joint it actually has.

The third failure is the one that matters, and it is a category error. Suppose authentication worked perfectly: every artifact arrives stamped with a complete, tamper-proof causal history. What exactly has been preserved? Provenance was never what the artifact regime sold. Provenance was a proxy for cognition—the text certified the thinking because only thinking could produce the text—and the rupture consists precisely in the severing of proxy from target.^{95Note 95Claude (AI) check Internal refJLETI Board PendingSee supra Part II.D.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} Once severed, certifying the proxy preserves a ritual, not a function. A certified-human article that a frontier model could have written in four minutes attests to abstinence, not to thought; the certificate has the structure of a relic, an object venerated for its causal contact with a body rather than for anything it does.^{96Note 96Claude (AI) check Internal refJLETI Board PendingBenjamin distinguished the artwork's cult value—grounded in unique causal history—from its exhibition value, and observed that reproduction forces the transition. Benjamin, supra note 82, at 223–26. The authentication paradigm is an attempt to re-found cult value by administrative rule.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} Menard, again, states the problem exactly: if the words are the same, the difference authentication preserves is theological. An academy that adopts the paradigm will find itself, within a decade, solemnly certifying the humanity of texts no one believes are evidence of anything—a cult of authorship maintained after the cult's epistemic god has died.

C. Two Conservatisms

It may seem that the paradigm's opposite—enthusiastic adoption—escapes the criticism. It does not, and the most instructive demonstration was performed by a sitting dean. Perlman's experiment, an article on the future of legal scholarship drafted nearly in its entirety by a frontier model, had the machine propose its own theory, "Generative Synthesis": scholarship as the rapid, machine-enabled synthesis of vast literatures into new arguments and artifacts.^{97Note 97Claude (AI) check Internal refJLETI Board PendingPerlman, supra note 6.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} Read carefully, the machine's theory is the artifact regime's theory, with the assembly line accelerated. Its unit is still the synthesized text; its measure is still the persuasive article; its picture of progress is still more arguments, faster. The accelerationists and the authenticators are thus two wings of a single conservatism: both hold the artifact fixed and dispute only its permissible causes, the one guarding the old production function, the other turbocharging it.^{98Note 98Claude (AI) check Internal refJLETI Board PendingThe most sophisticated work in the adaptationist register is Frazier & Rozenshtein, supra note 20, which sees clearly that disclosure debates are unstable but retains the article, and the scholar's relation to it, as the unit to be adapted. The disagreement here is foundational rather than operational: their question is how the artifact's institutions should metabolize the machines; this Article's claim is that the metabolizing institution is itself the variable.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} The pattern has a precedent. When the internet arrived, the academy's most prescient critic predicted the law review's death by disintermediation—scholars would self-publish, and the journals' distributive monopoly would dissolve.^{99Note 99Claude (AI) check Internal refJLETI Board PendingHibbitts, supra note 9, at 686–94.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} The prediction failed because distribution was never the scarce factor; prestige and certification were, and the journals kept both. Today's conservatisms invert the error: they treat production as the thing to be governed, when production is precisely what has ceased to be scarce. What the regime actually rationed—and what now requires an entirely different institutional form—is the thing the artifact only ever proxied: validated normative cognition. Specifying that form is the task of the next Part.^{100Note 100Claude (AI) check Internal refJLETI Board PendingSee infra Part IV.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.}

IV. LATENT SCHOLARSHIP

A. From Corpus to Latent Space

An objection of genre should be met before the theory is stated. Easterbrook famously derided cyberlaw as "the law of the horse"—the study of a technology's legal cross-section teaches nothing general.^{101Note 101Claude (AI) check AI onlyJLETI Board PendingFrank H. Easterbrook, Cyberspace and the Law of the Horse, 1996 U. CHI. LEGAL F. 207.Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} Lessig's reply was that some cross-sections are different: they expose the architecture of regulation itself, the way a constitutional crisis exposes a constitution.^{102Note 102Claude (AI) check AI onlyJLETI Board PendingLawrence Lessig, The Law of the Horse: What Cyberlaw Might Teach, 113 HARV. L. REV. 501 (1999).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} So here. Asking where scholarship lives after the rupture forces an answer to a question the artifact regime never had to pose: what, exactly, was the medium of legal scholarship? The reflexive answer—text—was true only superficially. The medium was the corpus: the public, accumulating, citable mass of artifacts, within which influence was a position and authority an address.

The generative systems interpose a new medium between the corpus and its use. A large model does not consult volume 85 of the Harvard Law Review; it consults a learned representation—a high-dimensional geometry induced from the corpus during training, in which concepts are directions, doctrines are regions, and the relations among them are distances.^{103Note 103Claude (AI) check Internal refJLETI Board PendingSee Vaswani et al., supra note 61; Bommasani et al., supra note 63, at 5–7, 21–26 (describing learned representations and their reuse across tasks).Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} In that space, The Cathedral is not a location but a pressure: a set of directions in which questions about entitlements, once asked, get bent. The technical literature calls this representational interior latent space, and the term deserves adoption by lawyers, because it names the place where the profession's questions are now increasingly answered. Borges supplies the inversion that makes the migration vivid. His Library of Babel contains every possible text and is therefore useless—a total corpus with zero retrieval, melancholy shelving without a mind.^{104Note 104Claude (AI) check Internal refJLETI Board PendingBorges, The Library of Babel, supra note 15, at 51.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} The model is the anti-Library: it contains almost no text verbatim, and for that very reason it can be consulted about everything. The corpus was scholarship's habitat for fourteen decades. It is becoming scholarship's feedstock. And the point of use has already moved—into chambers, where a circuit judge interrogates the model about ordinary meaning; into doctrine, where interpretation-by-model is a live methodological proposal; into the profession's daily workflow, with the Chief Justice's measured blessing.^{105Note 105Claude (AI) check Internal refJLETI Board PendingSee Snell v. United Specialty Ins. Co., 102 F.4th 1208, 1221–34 (11th Cir. 2024) (Newsom, J., concurring); Coan & Surden, supra note 19; Arbel & Hoffman, supra note 70; Roberts, supra note 72.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} Scholarship goes where the law's questions are answered. The discipline whose questions are answered in latent space, but whose institutions all face the corpus, is a discipline pointed at its own past. The reorientation—of units, of measures, of institutions—is what this Article calls latent scholarship.

B. The Four Scarce Goods

Economies reorganize around their scarcities. When competent argumentative text becomes free, four goods remain scarce, and together they define the new discipline.

The first is formulation. The model answers; it does not ask. What remains expensive is the upstream act: identifying the question, the framing, the hypothetical that will organize a field's attention. Simon's law—information wealth produces attention poverty—does not relax when answers become free; it tightens, and the binding constraint migrates from the production of arguments to the agenda that selects among infinitely many producible arguments.^{106Note 106Claude (AI) check Internal refJLETI Board PendingSimon, supra note 53, at 40–41.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} Kuhn saw that a paradigm's deepest function is to generate puzzles worth solving;^{107Note 107Claude (AI) check Internal refJLETI Board PendingKUHN, supra note 10, at 35–42.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} in the latent regime, the scholar's first product is the puzzle itself, posed with enough precision that machine labor can be usefully spent on it. Formulation cannot be delegated, not because machines lack ingenuity, but because which problems matter is an indexical fact about a community with stakes, and the machine is not a member of any community at stake.

The second is validation. Free text is unverified text in unbounded supply, and the scarce good becomes warranted confidence—knowing which propositions, frameworks, and systems hold. Here is a fact hiding in plain sight: the first genuinely post-rupture legal scholarship has already been published, and the discipline has not recognized its genre. The hallucination studies are not articles about AI; they are evaluations—constructed test distributions, assembled ground truth, taxonomies of failure—performed on legal-question-answering systems.^{108Note 108Claude (AI) check Internal refJLETI Board PendingSee Dahl et al., supra note 3; Magesh et al., supra note 74.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} LegalBench, built collaboratively by dozens of scholars, decomposes legal reasoning into testable components and subjects systems to them at scale; that is what a literature looks like when its object is a mindshaped system rather than a thesis.^{109Note 109Claude (AI) check AI onlyJLETI Board PendingNeel Guha et al., LegalBench: A Collaboratively Built Benchmark for Measuring Legal Reasoning in Large Language Models, in 36 ADVANCES IN NEURAL INFORMATION PROCESSING SYSTEMS (2023).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} The footnote verified propositions one at a time. The benchmark verifies an entire reasoning system at once. Validation scholarship—benchmark construction, adversarial probing, audit design—is the successor to doctrinal verification, and its early practitioners are the new realists, whatever they take themselves to be doing. The third is commitment. A model can state every position and occupies none. Cover's insight was that legal meaning is generated—jurisgenesis—by communities committed to their normative visions, and that "[w]e inhabit a nomos—a normative universe" sustained not by texts alone but by the willingness of persons to hold them, live under them, and pay for them.^{110Note 110Claude (AI) check Internal refJLETI Board PendingCover, supra note 16, at 4; see id. at 11–19 (developing jurisgenesis).Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} Legal interpretation, he insisted elsewhere, differs from literary interpretation because it ends in force: to propose law is to accept responsibility for what is done in its name.^{111Note 111Claude (AI) check AI onlyJLETI Board PendingRobert M. Cover, Violence and the Word, 95 YALE L.J. 1601, 1601–10 (1986).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} The machine has nothing at stake—no career, no discipline, no body—and the law's own doctrine, having denied it authorship, a fortiori denies it the capacity to answer for a position.^{112Note 112Claude (AI) check Internal refJLETI Board PendingSee supra notes 84–85 and accompanying text; cf. ABA Comm. on Ethics & Pro. Resp., Formal Op. 512, supra note 76 (locating responsibility exclusively in the supervising human).Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} Commitment therefore becomes both a scarce good and the human warrant that replaces authorship. What the scholar of the latent regime signs is not the unverifiable and beside-the-point claim "I produced this text," but the consequential one: "I stake my professional standing on this normative position, and I will answer for it." Authentication asks who typed. Jurisgenesis asks who stands behind. Only the second question was ever the point.

The fourth is architecture: the design of the spaces in which the other three operate—which corpora are assembled and preserved, which benchmarks become canonical, which systems are maintained and by whom, who may query what on what terms. Code regulates;^{113Note 113Claude (AI) check Internal refJLETI Board PendingLESSIG, supra note 93, at 3–8.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} the scholar-asarchitect, and the institutions that must grow around her, are the subject of Part V.

C. The Executable Article

The four scarcities converge on a new unit of scholarship. Call it the maintained normative system: a framework for some region of law, stated with operational precision and kept alive over time. Its canonical components are six. A specification: the normative framework in prose, written to be implemented rather than admired. A corpus: the curated authorities the system treats as its tradition. An implementation: a model configuration—fine-tuned, scaffolded, retrieval-grounded, the engineering is fungible—that answers questions under the framework. A benchmark suite: canonical and adversarial hypotheticals with target outputs and rationales. A failure log: the public record of counterexamples, indeterminacies, and revisions. And a commitment statement: the maintainers' signed normative warrant, the document that makes the system an act of jurisgenesis rather than a product.

A worked example will fix ideas. Begin with the most-cited framework in private law.^{114Note 114Claude (AI) check Internal refJLETI Board PendingCalabresi & Melamed, supra note 17; see Shapiro & Pearse, supra note 43 (documenting its citation rank).Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} The scholar of 2030 who has something to say about entitlements does not write The Cathedral at Sixty. She builds Cathedral-1. Its specification states the entitlement-assignment and protection-rule logic, including the famous fourth rule. Its corpus is the nuisance, takings, and remedies canon. Its benchmark suite includes Boomer, where the framework must select the liability rule, and Spur, where it must generate the compensated injunction that doctrine produced only once—together with adversarial instruments the canon never faced: drone overflight, data pollution, algorithmic nuisance.^{115Note 115Claude (AI) check AI onlyJLETI Board PendingBoomer v. Atlantic Cement Co., 257 N.E.2d 870 (N.Y. 1970); Spur Indus., Inc. v. Del E. Webb Dev. Co., 494 P.2d 700 (Ariz. 1972).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} Its failure log records, honestly, where the framework goes indeterminate— inalienability was always the unfinished room of the cathedral. Its commitment statement says when and why liability rules should dominate, and is signed. Rivals respond in kind: a competing team maintains a Coasean system that resolves the same benchmarks by bargaining analysis, and the divergence between the two systems on shared instruments is the scholarly debate—stated, for the first time in the discipline's history, in executable and therefore precisely contestable form. A clerk in 2031 confronting a novel externality does not pull volume 85; she runs the facts against Cathedral-1.7 and its rival and reads the disagreement. The systems are versioned because they are alive: attacked, patched, conceded, forked.

The unit has a lineage, which is reassurance that it is law's own. Fuller's Speluncean Explorers was a benchmark before the word existed—one fixed fact pattern, five committed jurisprudential systems executed against it—and the genre's long afterlife of sequels shows the profession has always known that frameworks are tested on shared hypotheticals.^{116Note 116Claude (AI) check AI onlyJLETI Board PendingLon L. Fuller, The Case of the Speluncean Explorers, 62 HARV. L. REV. 616 (1949).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} Dworkin's chain novel imagined law itself as a collaboratively maintained system under interpretive constraint; the maintained normative system is the chain novel with a commit history.^{117Note 117Claude (AI) check AI onlyJLETI Board PendingRONALD DWORKIN, LAW'S EMPIRE 228–38 (1986).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} And the distinctions matter as much as the lineage. This is not computable contracts or the microdirective program, which compute law for transactions—law as product, optimized and delivered;^{118Note 118Claude (AI) check Internal refJLETI Board PendingSee Surden, Computable Contracts, supra note 59; Casey & Niblett, supra note 59.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} the maintained system argues to law, and its mode is contestation. Nor is it the legal singularity, which imagines convergence on functionally complete law—one optimum, disputes dissolved;^{119Note 119Claude (AI) check Internal refJLETI Board PendingAlarie, supra note 22.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} latent scholarship is constitutively pluralist, a marketplace of maintained nomoi whose disagreements are precise, queryable, and never final. Nor, finally, is it Generative Synthesis, which accelerates the old unit rather than replacing it.^{120Note 120Claude (AI) check Internal refJLETI Board PendingSee supra notes 97–98 and accompanying text.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} The article itself survives, in the way the sonnet survived the novel and the photograph: practiced for its discipline, capable of perfection, beloved—and no longer the unit of account.

D. The Authority of Weights

If the citation no longer measures influence, what does? The latent regime's answer is that influence becomes, for the first time, an empirical property rather than an honorific one. Define the uptake delta: the change in the behavior of the systems the profession actually consults that is attributable to a scholarly contribution. The quantity is measurable in principle and increasingly in practice—by influence functions, which trace a model's outputs to the training data responsible for them;^{121Note 121Claude (AI) check AI onlyJLETI Board PendingPang Wei Koh & Percy Liang, Understanding Black-Box Predictions via Influence Functions, in PROCEEDINGS OF THE 34TH INTERNATIONAL CONFERENCE ON MACHINE LEARNING 1885 (2017).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} by the same techniques scaled to frontier models;^{122Note 122Claude (AI) check AI onlyJLETI Board PendingRoger Grosse et al., Studying Large Language Model Generalization with Influence Functions, arXiv:2308.03296 (Aug. 7, 2023).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} by ablation, comparing systems trained or retrieved with and without the contribution; by behavioral probing, asking whether the profession's models now generate the compensated injunction where once they did not. The old economy's idiom can now be heard saying what it always almost said. The profession spoke of the weight of authority. The successor concept is the authority of weights: the degree to which a contribution has bent the geometry through which the law's questions pass.

Three consequences follow. First, the pincite dies, and with it the archive function as the regime knew it: one cannot pincite a distribution, and provenance becomes statistical—attribution rather than address.^{123Note 123Claude (AI) check Internal refJLETI Board PendingSee supra notes 41–42 and accompanying text.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} Second, influence measurement becomes honest. The citation count was always a proxy for the unmeasurable fact of having shaped the field's mind; the uptake delta measures the shaping directly, on the artificial minds through which the field increasingly thinks. Third, the pathologies arrive on schedule, and they should be named before they are discovered. There will be model optimization—the search-engine optimization of scholarship, writing not to persuade readers but to be ingested by pipelines, the attention merchants' logic migrating from eyeballs to training runs.^{124Note 124Claude (AI) check AI onlyJLETI Board PendingCf. TIM WU, THE ATTENTION MERCHANTS: THE EPIC SCRAMBLE TO GET INSIDE OUR HEADS (2016).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} There will be corpus attack: data poisoning is demonstrated, cheap, and practical at web scale, which converts the integrity of the training commons from a security topic into a core problem of scholarly governance.^{125Note 125Claude (AI) check AI onlyJLETI Board PendingSee Nicholas Carlini et al., Poisoning Web-Scale Training Datasets Is Practical, arXiv:2302.10149 (Feb. 20, 2023).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} And over every metric falls Goodhart's shadow: a measure of uptake, once targeted, will be gamed. These are not objections to the theory; they are its agenda, and the institutional answers occupy Part V.^{126Note 126Claude (AI) check Internal refJLETI Board PendingSee infra Part V.B (on Goodhart-resistant credentialing).Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.}

E. What Remains Human

The theory will be misread as automation triumphant; it is nearly the opposite, and the point deserves a section rather than a sentence. A predictive system, however complete, yields the distribution of what would be said—the is of legal speech, mastered to the last register. The office of scholarship has always been the ought, and the jurisprudential tradition's most carefully defended border runs exactly there, between reporting a practice and prescribing one.^{127Note 127Claude (AI) check AI onlyJLETI Board PendingSee H.L.A. Hart, Positivism and the Separation of Law and Morals, 71 HARV. L. REV. 593, 607–15 (1958).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} A distribution cannot answer a normative question; it can only inform the one who must. Cardozo, candid about judicial legislation within limits, assigned to the method of sociology the task of steering it and to the jurist the task of supplying its conscience;^{128Note 128Claude (AI) check AI onlyJLETI Board PendingBENJAMIN N. CARDOZO, THE NATURE OF THE JUDICIAL PROCESS 98–141 (1921).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} the assignment stands, with better instruments. Indeed, the realists turn out to have written the latent regime's program reviews sixty years early. They demanded that legal concepts be judged by what they do—that transcendental nonsense give way to functional analysis^{129Note 129Claude (AI) check Internal refJLETI Board PendingCohen, supra note 36, at 821–34; see Llewellyn, supra note 36.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.}—and the demand outran the instruments of every generation that made it. A maintained normative system is the functional approach made finally functional: a framework whose consequences can be executed, measured, and compared against a rival's, hypothetical by hypothetical, rather than asserted in prose and adjudicated by fashion.

The human remainder is therefore not a residue but a constitution: choosing the problems, warranting the confidence, standing behind the positions, designing the commons—and tending the nomos, for Cover's communities of meaning do not dissolve into the machines; they acquire, in them, a new organ.^{130Note 130Claude (AI) check Internal refJLETI Board PendingCover, supra note 16, at 4–11, 40–44.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} Latent scholarship does not remove the scholar from the law. It relocates her—from writing the answers to governing the space in which answers form.

V. INSTITUTIONAL DESIGN

If Part IV is right that the unit of scholarship is migrating from the fixed artifact to the maintained normative system, then every institution calibrated to the artifact—the journal that selects it, the committee that counts it, the court that cites it, the database that shelves it—must be rebuilt around the new unit or watch its function drain quietly away. This Part sketches the rebuilding. It is offered as institutional imagination rather than blueprint: the designs below will be wrong in detail, and are meant to be right in kind.

A. Law Reviews as Validation Institutions

Begin with the institution through which this Article must, by professional custom, itself pass. The student-edited law review has been mocked for a century on one ground above all: that secondand third-year students select and edit the work of the professoriate.^{131Note 131Claude (AI) check Internal refJLETI Board PendingSee Posner, supra note 52 (cataloguing the indictment while predicting the institution's persistence); Lindgren, supra note 48.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} The mockery always mistook the institution's actual competence. Article selection was never what law reviews did distinctively well; verification was. The cite-check—that grinding audit in which a staff of twenty confirms, proposition by proposition, that every claim rests on the authority beneath it—is the most thoroughgoing verification apparatus in the academy, and the Bluebook is its procedural code.^{132Note 132Claude (AI) check Internal refJLETI Board PendingTHE BLUEBOOK, supra note 42. The point is institutional rather than stylistic: the citation manual is the codified procedure of a standing verification bureaucracy staffed, at any given moment, by thousands.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} The artifact regime buried that competence in the back office of publication. The latent regime promotes it to the front, because validation is precisely the scarce good the new regime cannot generate for itself.

Reconceived as a validation institution, the law review curates benchmark suites of hard cases, administers adversarial review of maintained systems before certification, publishes performance results alongside failure logs, and retests on a schedule rather than once at acceptance. Each element has a working prototype. LegalBench was assembled the way a symposium issue is—by soliciting contributions, tasks rather than essays, from across the interpretive community.^{133Note 133Claude (AI) check Internal refJLETI Board PendingGuha et al., supra note 109.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} The hallucination audits of the commercial research tools are, in form, exactly what the article review becomes: structured adversarial evaluation of deployed systems, published with error rates.^{134Note 134Claude (AI) check Internal refJLETI Board PendingMagesh et al., supra note 74.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} The volume becomes a registry of certified systems; the issue becomes a release cycle; the case note becomes a benchmark task. And the institution's famous pathology becomes its qualification. Adversarial evaluation does not require seniority; it requires ingenuity in the construction of counterexamples—a skill legal education has cultivated in students since the hypothetical was invented, and that Fuller raised to literature.^{135Note 135Claude (AI) check Internal refJLETI Board PendingCf. Fuller, supra note 116 (one invented case, five irreconcilable systems—the hypothetical as a complete adversarial instrument).Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.}

B. Credentialing After Citation

Appointments and promotion under the artifact regime run on counts: placements weighted by journal prestige, citations accumulated and ranked, the whole apparatus of measurement that the citation studies at once documented and reinforced.^{136Note 136Claude (AI) check Internal refJLETI Board PendingShapiro & Pearse, supra note 43; Ayres & Vars, supra note 40.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} Part IV proposed the successor measure: uptake delta, the demonstrated change in the behavior of maintained systems attributable to a scholar's contribution.^{137Note 137Claude (AI) check Internal refJLETI Board PendingSee supra notes 121–122 and accompanying text.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} A tenure file in the latent regime is a portfolio rather than a bibliography: formulations contributed, with their measured propagation across systems; validation work— benchmarks authored, audits conducted, failures documented; commitments of record, positions taken and defended through successive versions; and architecture, the corpora curated and the systems maintained in the candidate's care.

The objection writes itself, and Part IV promised it an answer here. When a measure becomes a target it ceases to be a good measure;^{138Note 138Claude (AI) check AI onlyJLETI Board PendingMarilyn Strathern, 'Improving Ratings': Audit in the British University System, 5 EUR. REV. 305 (1997) (stating the canonical generalization of Goodhart's law).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} citation counts were Goodharted within a generation of becoming currency, and uptake metrics will summon flooding and optimization in the forms already confronted. The answer is not a Goodhart-proof metric, which does not exist, but a Goodhart-resistant institution. Evaluation sets rotate and are partially sealed, so that teaching to the test is teaching to a moving and partly invisible test. Metrics are plural, so that optimizing against one degrades visibly against another. And the metrics are themselves objects of scholarship—adversarially audited, revised in public—so that manipulation, when documented, is documented in the regime's own organs and priced into reputation. None of this is exotic. It is the cite-check, pointed at a new object.

There is a quiet payoff for the regime's oldest complaint. The lament that scholarship had abandoned the bench and bar was, under the artifact regime, an anecdote war—uncountable claims about who read what.^{139Note 139Claude (AI) check Internal refJLETI Board PendingSee Edwards, supra note 49; supra notes 49–51 and accompanying text.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} Uptake is countable. Whether a body of work changed the systems through which practice actually flows stops being a matter of conference grievance and becomes a matter of measurement, a development the complaint's authors should welcome: their hypothesis was always that the number would be embarrassing.

C. Courts: A Daubert for Normative Systems

Courts have consumed scholarship since before the footnote settled into its modern form, and they have consumed it untested: the treatise carried authority without an error rate, and the article entered the United States Reports on persuasion alone.^{140Note 140Claude (AI) check Internal refJLETI Board PendingSee supra notes 37–38 and accompanying text.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} The latent regime's quiet irony is that it makes scholarship, for the first time, the kind of input courts already know how to screen. The Daubert factors—testability, known or potential error rate, peer review and publication, general acceptance—map onto the maintained normative system with almost embarrassing precision: the benchmark is testability; the audit supplies an error rate; adversarial review is peer review with consequences; uptake is general acceptance, measured rather than assumed.^{141Note 141Claude (AI) check AI onlyJLETI Board PendingDaubert v. Merrell Dow Pharms., Inc., 509 U.S. 579, 593–94 (1993); see FED. R. EVID. 702.Daubert v. Merrell Dow, 509 U.S. 579 (1993) — exact cite correct.} A prose article could never be Dauberted. A system can. When a brief leans on Cathedral-1.7's resolution of an entitlement problem—and the leaning has already begun at the level of interpretive method^{142Note 142Claude (AI) check Internal refJLETI Board PendingSee supra notes 71–72 and accompanying text.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.}— the opposing party can demand the failure log, rerun the contested class of cases, and impeach the system the way one impeaches an expert: on its record.

The algorithmic-accountability literature supplies the cautions, and they should be imported wholesale. Loomis let a sentencing court consult a proprietary risk instrument while conceding in the same breath that due process constrains reliance on what cannot be examined;^{143Note 143Claude (AI) check AI onlyJLETI Board PendingState v. Loomis, 881 N.W.2d 749 (Wis. 2016).State v. Loomis, 881 N.W.2d 749 (Wis. 2016) — exact cite correct.} Citron's technological due process^{144Note 144Claude (AI) check AI onlyJLETI Board PendingDanielle Keats Citron, Technological Due Process, 85 WASH. U. L. REV. 1249 (2008).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} and the black-box critique^{145Note 145Claude (AI) check AI onlyJLETI Board PendingFRANK PASQUALE, THE BLACK BOX SOCIETY: THE SECRET ALGORITHMS THAT CONTROL MONEY AND INFORMATION (2015).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} built the case that opacity in systems exercising public power is itself a process failure. The contexts differ in a way that matters—those systems judged persons, these advise judges—but the borrowing runs in the right direction. The commitment statement and the public failure log are transparency obligations carried over from that literature into scholarship's new form. A normative system whose maintainers will not disclose its evaluation record should carry no more weight in a brief than an expert who declines to state a methodology, and for the same reason.

D. The Corpus as Commons

Everything above rests on a substrate: the corpus on which systems are trained and against which they are evaluated. Here law is peculiarly fortunate and peculiarly endangered. Fortunate, because its raw material is public by construction—no one may own the edicts of government, a principle the Supreme Court reaffirmed on the eve of the generative era.^{146Note 146Claude (AI) check AI onlyJLETI Board PendingGeorgia v. Public.Resource.Org, Inc., 590 U.S. 255 (2020) (government-edicts doctrine).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} Endangered, because access to that public material has been privately intermediated for a century and a half, and the intermediation now extends to the training rights on which the latent regime runs. The liberation projects—the digitization of the printed case law, the open court-records infrastructure^{147Note 147Claude (AI) check AI onlyJLETI Board PendingCASELAW ACCESS PROJECT, https://case.law (last visited June 9, 2026); FREE LAW PROJECT, https://free.law (last visited June 9, 2026).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).}—and the machine-learning commons assembled atop them, the Pile of Law most prominently,^{148Note 148Claude (AI) check AI onlyJLETI Board PendingPeter Henderson et al., Pile of Law: Learning Responsible Data Filtering from the Law and a 256GB Open-Source Legal Dataset, in 35 ADVANCES IN NEURAL INFORMATION PROCESSING SYSTEMS (2022).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} are accordingly not library conveniences; they are the regime's constitutional infrastructure. Infrastructure theory explains why: nonrival inputs that generate spillovers across an entire economy are exactly the resources that repay commons management,^{149Note 149Claude (AI) check AI onlyJLETI Board PendingBRETT M. FRISCHMANN, INFRASTRUCTURE: THE SOCIAL VALUE OF SHARED RESOURCES (2012).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} and no input in the legal economy is more generative than the corpus itself.

The concentration danger is the regime's hardest political-economy problem. If frontier systems are few, proprietary, and corporately aligned, latent scholarship collapses into company scholarship: the nomos acquires a landlord. The response this Article urges is an epistemic antitrust—structural attention to who owns the substrate of legal knowledge—in the tradition recently revived for the platform economy^{150Note 150Claude (AI) check AI onlyJLETI Board PendingSee Lina M. Khan, Amazon's Antitrust Paradox, 126 YALE L.J. 710 (2017).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} and documented for the extraction stack beneath the models.^{151Note 151Claude (AI) check AI onlyJLETI Board PendingKATE CRAWFORD, ATLAS OF AI: POWER, POLITICS, AND THE PLANETARY COSTS OF ARTIFICIAL INTELLIGENCE (2021).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} And the integrity threat converts a clerical task into a scholarly office. If a public corpus can be poisoned cheaply,^{152Note 152Claude (AI) check Internal refJLETI Board PendingSee Carlini et al., supra note 125.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} then curation—provenance, authentication, the keeping of the record—becomes to the latent regime what the great reporters were to the print one: the unglamorous work on which everything downstream silently depends.

E. The Quiet Testimony of Authorship Doctrine

The law has already been asked, directly, whether machine output can inherit the artifact regime's property logic, and it has answered: no one owns it. The Copyright Act's author is human; the monkey lacked standing and the autonomous machine lacks authorship; registration requires determinative human contribution, and prompts alone do not supply it.^{153Note 153Claude (AI) check Internal refJLETI Board PendingSee supra note 84 and accompanying text.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} The consequence is the one flagged at the rupture's edge: text generated without a human author enters the world unowned, so that an artifact regime extended across the rupture would be a regime of ownerless currency— credit attaching to no one, exclusivity available to no one. The doctrine is usually read as a gatekeeping rule. Read affirmatively, it is a map of where protectable contribution now lives: in selection, arrangement, direction, and creative control—which is to say, in formulation and architecture, two of the four scarce goods.^{154Note 154Claude (AI) check Internal refJLETI Board PendingSee supra note 84 (the Copyright Office's insistence on determinative human contribution); Samuelson, supra note 86.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} Copyright's oldest exclusion completes the chart. Section 102(b) withholds protection from any system or method of operation as such,^{155Note 155Claude (AI) check AI onlyJLETI Board Pending17 U.S.C. § 102(b).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} so the maintained normative system resists ownership at both ends—its generated expression authorless, its operative method unprotectable—and gravitates, by legal design rather than utopian hope, toward the commons that Part V.D argued it requires. Authorship doctrine, asked to police the old regime's border, has quietly chartered the new one.

VI. OBJECTIONS

Five objections have shadowed the argument, and this Part states each at the full strength its proponents would give it. Two are conceded in part. None, I will argue, is fatal—though the most serious arrives as a refutation and should be kept as a design constraint.

A. The Normativity Objection

The first is jurisprudential bedrock. Scholarship is normative argument, and a language model is, on the standard critique, syntax without semantics—Searle's room, scaled^{156Note 156Claude (AI) check AI onlyJLETI Board PendingJohn R. Searle, Minds, Brains, and Programs, 3 BEHAV. & BRAIN SCI. 417 (1980).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).}—a stochastic parrot whose fluency is correlation rather than conviction.^{157Note 157Claude (AI) check Internal refJLETI Board PendingBender et al., supra note 65.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} Law, on a powerful recent statement, is a rule of persons, not machines;^{158Note 158Claude (AI) check AI onlyJLETI Board PendingFrank Pasquale, A Rule of Persons, Not Machines: The Limits of Legal Automation, 87 GEO. WASH. L. REV. 1 (2019).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} and interpretation, on Dworkin's account, is constitutively evaluative—an effort to make the practice the best it can be, which nothing that holds no view of the good can undertake.^{159Note 159Claude (AI) check Internal refJLETI Board PendingDWORKIN, supra note 117.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} If all of that is right, the maintained normative system cannot be scholarship, because nothing inside it means anything.

Concede the semantics arguendo; the objection refutes a thesis this Article does not hold. The claim was never that the machine is a scholar—machine-as-scholar is the artifact regime's category error running in reverse, authorship anxiously reassigned rather than dissolved. The claim is that scholarship relocates to acts only humans perform: formulation, which decides what the space shall be asked; commitment, which stakes a person and a community on an answer.^{160Note 160Claude (AI) check Internal refJLETI Board PendingSee supra notes 110–111 and accompanying text.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} The system is the medium of normative argument, not its author, and a medium has never needed beliefs. The printed page believed nothing either; the treatise held no views, though Williston did. Searle's room is an argument about where understanding is not. Latent scholarship is a theory about where commitment goes.

B. Langdell 2.0

The second objection hears history rhyming as farce. A century after Langdell announced law a science and was answered by generations of jurists who showed the science was politics in a laboratory coat,^{161Note 161Claude (AI) check Internal refJLETI Board PendingSee Pound, supra note 32; supra note 36 and accompanying text.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} this Article proposes—what, exactly? Law as benchmark performance: Langdell with graphics processors. The critical tradition's warning applies with renewed force: a benchmark launders contestable value choices as neutral measurement,^{162Note 162Claude (AI) check Internal refJLETI Board PendingSee Kennedy, supra note 45.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} and the deepest things at stake in legal argument resist commensuration on any common metric.^{163Note 163Claude (AI) check AI onlyJLETI Board PendingCass R. Sunstein, Incommensurability and Valuation in Law, 92 MICH. L. REV. 779 (1994).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} The reply is a single word with institutional consequences: reflexivity. Langdell's science suppressed contestation—the case method presumed right answers, and the laboratory presumed one chemistry. The latent regime institutionalizes contestation as its basic form. Rival maintained systems are what disagreement looks like, and the variance among them is not noise to be engineered away but the regime's most important measurement: where Cathedral-1 and Coase-2 diverge across ten thousand hypotheticals, the indeterminacy thesis—asserted in prose for ninety years—becomes observable, the realist claim rendered empirical at last.^{164Note 164Claude (AI) check Internal refJLETI Board PendingCohen, supra note 36.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} As for commensuration, the design has already conceded it. Metrics are plural by constitution, benchmarks are contestable objects of scholarship, and contesting one is architecture, the fourth scarce good. Langdell offered a science instead of a politics. The latent regime offers instruments to the politics—the opposite program, run on similar hardware.

C. Collapse and Monoculture

The third objection is material, and it is the strongest. The regime may eat its inputs twice over. Models trained recursively on machine output degrade—collapse is now measured, not conjectured;^{165Note 165Claude (AI) check AI onlyJLETI Board PendingIlia Shumailov et al., AI Models Collapse When Trained on Recursively Generated Data, 631 NATURE 755 (2024).Shumailov et al., AI Models Collapse, 631 Nature 755 (2024) — real Nature paper.} a handful of shared foundation systems correlates error across an entire ecosystem, purchasing efficiency with systemic fragility;^{166Note 166Claude (AI) check AI onlyJLETI Board PendingJon Kleinberg & Manish Raghavan, Algorithmic Monoculture and Social Welfare, 118 PROC. NAT'L ACAD. SCIS. e2018340118 (2021).Recognized source; citation form assessed as correct (model knowledge; existence + Bluebook form, not proposition support).} and the humans may degrade in parallel, unaided competence atrophying inside the loop. The first randomized evidence on that last fear is, instructively, mixed: early machine assistance in a legal-reasoning task did not erode later unaided performance—the assisted group did better even after the tool was withdrawn—but machine-aided revision dragged the strongest performers toward the mean.^{167Note 167Claude (AI) check Needs humanJLETI Board PendingNicholas Bednar, David R. Cleveland, Allan Erbsen & Daniel Schwarcz, Artificial Intelligence and Human Legal Reasoning (Apr. 5, 2026) (unpublished manuscript), https://ssrn.com/abstract=6525800.Working paper (SSRN, Apr. 2026). The empirical findings should be confirmed against the manuscript by a person.} Read together, the findings sharpen the worry rather than dissolving it, because regression toward the mean is monoculture operating at the level of the individual mind.

The answer is to treat the objection as the regime's first design constraint. Collapse identifies what must never be automated: the scarce goods are the anti-collapse inputs, since human formulation and human commitment are precisely the fresh, off-distribution data on which the systems' continued usefulness depends. The regime has, as it were, a thermodynamic interest in the humanity of its humans. Provenance-tracked commons answer the corpus problem; constitutive rivalry among maintained systems answers monoculture, since a nomos with one system is not a nomos but a database; and the pedagogical answer is sequencing rather than abstention—the conservatory did not abolish scales when recordings arrived, and the early evidence suggests that timing, not prohibition, is the operative variable.^{168Note 168Claude (AI) check Internal refJLETI Board PendingSee supra Part V.D (provenance); Choi & Schwarcz, supra note 69 (documenting the largest gains among the weakest performers); Bednar et al., supra note 167.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.}

D. The Imperial Model

Fourth, the political objection. Delgado asked who got to write the scholarship and found an imperial circle citing itself;^{169Note 169Claude (AI) check AI onlyJLETI Board PendingRichard Delgado, The Imperial Scholar: Reflections on a Review of Civil Rights Literature, 132 U. PA. L. REV. 561 (1984).Delgado, The Imperial Scholar, 132 U. Pa. L. Rev. 561 (1984) — exact cite correct.} the latent regime threatens an imperial model—capital-intensive, trained on the well-documented and the English-speaking, owned where the compute is— exclusion by infrastructure rather than by network, and infrastructure is harder to crash.^{170Note 170Claude (AI) check Internal refJLETI Board PendingSee supra notes 150–151 and accompanying text.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} The concession comes first: the danger is real, its economic version is the regime's deepest vulnerability, and the commons of Part V.D is necessary rather than sufficient. But two features of the design answer the objection's older form. Open evaluation inverts the artifact regime's asymmetry of standing: under the old regime, an outsider's critique needed placement, patronage, letterhead; under open benchmarks, a documented failure is a failure whoever files it, and the registry cannot see the provenance of a counterexample. And jurisgenesis was never centralized. The communities Cover described generate law out of commitment, not capital,^{171Note 171Claude (AI) check Internal refJLETI Board PendingCover, supra note 16, at 11–19.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} and a community-maintained system—a tribal code's, a movement's, a small republic of meaning's—is within reach of modest means in a way a frontier model is not. The imperial center is a danger to the latent regime. It was a realized fact of the artifact one.

E. Aura, Joy, and the Seminar Room

The last objection is the most human, and it should be stated without embarrassment, because everyone in the discipline feels it. The great article had aura—the singular, authored thing, encountered as an event—and mechanical reproduction is exactly the solvent the aura's theorist said it was;^{172Note 172Claude (AI) check Internal refJLETI Board PendingBenjamin, supra note 82, at 223–26.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} a registry of versioned systems has the charisma of a utilities closet. There is a companion fear for the working life: if the writing is outsourced, the love dies with it, and if a machine wrote it, why should anyone read it.^{173Note 173Claude (AI) check Internal refJLETI Board PendingSmith, supra note 20.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} But the essay carries the answer it is cited against. Reproduction destroyed cult value and created exhibition value, and the arts did not end—they moved, and film became an art no altar could have housed. So here. Prose argument will persist as a practiced art, the way the lecture and the letterpress persist, valued as performance precisely because nothing depends on mistaking it for the scarce thing. The seminar room is promoted, not demolished: formulation and commitment—the goods no system supplies—are what seminars manufacture, and what scholars actually love was never the typing. It was the idea arriving; the argument surviving contact with colleagues who want it dead. That survives by construction, because it is the part the regime cannot generate. The calmest voices in the recent literature are right that integration is coming either way;^{174Note 174Claude (AI) check Internal refJLETI Board PendingFrazier & Rozenshtein, supra note 20.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} the question this Article adds is whether the discipline will design the institutions of its new form, or wake inside institutions designed by others.

Conclusion

Holmes stood before a roomful of artifact-regime scholars in 1897 and told them their successor had already been born: the black-letter man's future belonged to "the man of statistics and the master of economics."^{175Note 175Claude (AI) check Internal refJLETI Board PendingHolmes, supra note 25, at 469.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} The profession spent a century reading the line as a prophecy about economists, and the citation counts eventually agreed. But the essay's deeper instruction came at its end, where Holmes urged the lawyer toward the remoter horizon of the subject—the vantage from which the local quarrels of doctrine resolve into structure.^{176Note 176Claude (AI) check Internal refJLETI Board PendingHolmes, supra note 25, at 478.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} This Article has tried to follow that instruction about scholarship itself. Viewed locally, the present moment is a quarrel about disclosure forms. Viewed from the horizon, it is a change of state in the medium of legal knowledge.

Langdell built a regime on the claim that the library was the lawyer's laboratory, and for a hundred and forty years the claim held: the materials of the science sat on shelves, and scholarship was the writing that arranged them.^{177Note 177Claude (AI) check Internal refJLETI Board PendingSee Langdell, supra note 8, at 124; supra Part I.A.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} The materials no longer sit. They circulate, learned into systems that answer, and the laboratory has changed state—from a room one walks into to a space one queries. The form of work that counted as science in the first laboratory cannot simply be carried into the second. What carries is the office. Someone must still decide which questions the space is made to answer; validate what comes back; stake a community's meaning on a position and hold it through failure; and govern the substrate on which all of it runs. Those are the four offices this Article has called the scarce goods, and the institutions worth building are the ones that protect them.

The nomos, Cover taught, is held in place by commitment, and it has survived its media before—memory to scroll, scroll to codex, codex to print, print to database.^{178Note 178Claude (AI) check Internal refJLETI Board PendingCover, supra note 16, at 4; see supra Part IV.E.Internal cross-reference (“supra/infra”) to a source cited elsewhere; no new external citation.} It will survive this passage too, on one condition the earlier transitions never imposed: that the scholars notice the passage while it is happening. The artifact was never the point. It was the best container the technology of attention could offer for formulation, validation, commitment, and architecture, and it served honorably for a century and a half. Now the container is dissolving into the thing it contained. The black-letter man yielded to the man of statistics, as predicted, and the prediction owed its accuracy to its method: follow the scarce thing. Follow it now, and the conclusion states itself. The scholar of the future is the master of the model.