Companion publications: Chicago School Accelerated — The Integrated Framework · Integrated Application: AI Hallucinations, AI Copyright · The Dual Nash-Stigler Equilibrium Architecture
Executive Summary
The Seattle Times and Newsday sued OpenAI and Microsoft for copyright infringement on September 4, 2026. The filing landed the same day summary judgment briefing opened in the consolidated New York case brought by The New York Times and more than a dozen other publishers. One question now sits before Judge Sidney Stein: may AI companies copy journalism to build and operate their products without permission or payment?
The training question misleads because it collapses several legally distinct uses. An AI product uses an article several separate times: it acquires the copy, trains on it, and then retrieves or reproduces it in answers. The same article can be fair to learn from, unlawful to acquire, and infringing to reproduce.
The litigation now presents a live pricing problem: briefing is open, and positions taken today price into the settlement that follows. The analysis predicts where that settlement lands and what it prices.
Courts and markets will not settle generative-AI copyright through a universal rule on training; they will divide the pipeline according to coordination feasibility, market substitution, and control. Fair use will protect non-substitutive training on lawfully acquired works where work-by-work licensing remains infeasible.
Liability and licensing will concentrate on unlawful acquisition, protected-content retrieval, and substitutive outputs. Developers can prevent harm and meter access in those layers. Prevention capacity decides allocation.
The result is the fair-use settlement equilibrium of the title: publishers lose universal training control, developers lose unrestricted acquisition and output freedom, and licensing occupies the valuable territory between them.
Litigation and provenance infrastructure will then move the boundary itself. Markets created by the copyright contest will consume part of the market failure that originally supported fair use. Market formation alone cannot eliminate fair use for the historical training core, and the divided rule’s durability depends on propagation across courts and contracts.
MindCast AI reads the contest through the Chicago School of Law and Behavioral Economics and the Dual Nash-Stigler architecture: game theory supplies payoff structure and equilibrium selection, and behavioral economics supplies decision rules and salience effects. Predictive behavior emerges from the combination. MindCast registered this migration’s direction in December 2025, nine months before the first summary judgment motion.
Prediction Highlights
The MindCast AI Proprietary Cognitive Digital Twin Foresight Simulation (MP CDT FS) run of September 6, 2026 released fourteen Simulation Predictions: eight primary and six secondary. Section II carries the full register. Headlines:
The court rules on each stage of the AI pipeline separately rather than deciding training as one question: 81–90%.
Training on lawfully acquired articles survives as fair use: 65–78%. Claims over how articles were obtained and what the products reproduce survive alongside it: 75–85%.
No court orders an AI model destroyed: 87–95%.
The summary judgment decision most likely arrives around May 2027.
A mixed ruling triggers at least three publisher licensing deals or settlements within 18 months: 58–70%.
An appellate ruling or federal statute resets the whole contest by 2029: 25–40%.
Stakeholders
🏛️ Policymakers: An 87–95% likelihood of non-structural remedies puts the design space at provenance, retrieval, and attribution duties. Legislate there rather than at training.
💼 Executives: Segment pipelines now. The 75–85% retrieval-prevalence prediction says the licensing market forms at the grounding layer first.
⚖️ Counsel: Plead and defend by stage. The register prices each claim class, and acquisition and output records carry the surviving leverage.
📊 Investors: Price content exposure to the four-route tree in Section XI. The modal divided rule carries 55% and the defense branch 22%.
I. The December 2025 Scorecard
MindCast’s Chicago School Accelerated — Integrated Application: AI Hallucinations, AI Copyright, and Crypto ATMsregistered three structural forecasts for AI copyright in December 2025. Litigation would shift from transformation metaphysics to measured substitution. Liability would emerge through output duties before courts imposed training rules, and enforcement would arrive first in news and music. The December registration makes the present paper a scored extension rather than a fresh theory.
The September 2026 record permits scoring. The news plaintiffs’ September 4 brief seeks liability at five pipeline stages, with one stage redacted throughout. Microsoft’s motion in the consolidated book cases answers with counting: 24 matching responses across 8.2 million Copilot conversations.
Both sides now litigate substitution measurement. Transformation remains contested, but both sides now operationalize it through substitution and frequency evidence.
Traffic economics leads the plaintiffs’ case. The briefing cites a crawl-to-referral ratio of 1,500 to 1 for OpenAI against 18 to 1 for Google. The Seattle Times complaint alleges industry data showing a 47% referral decline for midsize publishers, describing harm channels rather than adjudicated causation.
The forecast’s sector ordering largely held. News consolidated in the Manhattan Multidistrict Litigation (MDL). Music escalated through a lyrics suit against Anthropic seeking more than $3 billion in alleged damages, and books settled first, consistent with stronger acquisition-taint leverage in those cases.
Grading to date: the substitution turn and the output-first sequencing score as hits. The forecast did not call books settling before music, and the sequence miss grades on the page beside them.
The September record turns the December forecast into a base for extension. The December analysis answered where liability lands. The open questions are when and how the contest terminates.
II. MindCast AI Proprietary Simulation Predictions
The MP CDT FS run of September 6, 2026 adjudicated the contest through adversarial Cognitive Digital Twin (CDT)simulation. Each Cognitive Digital Twin models an actor’s incentives and constraints, its available moves, and its likely responses to other actors. The run released fourteen Simulation Predictions: eight primary and six secondary. P marks a Primary Simulation Prediction and S marks a Secondary Simulation Prediction. Four entries are conditional and resolve as unresolved rather than failed if their qualifying event never occurs.
Two hypotheses framed the run. Hypothesis A (stage-specific allocation): courts treat acquisition, training, and delivery as separate uses and allocate liability by stage. Hypothesis A fails if a controlling opinion analyzes the pipeline as one integrated exploitation. Hypothesis B (categorical rule): one answer governs the whole pipeline, for or against training. Hypothesis B fails if the opinion divides liability by stage. The simulation favors Hypothesis A, and the register follows.
The Ruling
P1. Segmented adjudication (81–90%). Judge Stein’s summary judgment opinion in the OpenAI MDL analyzes each pipeline stage as a distinct use. Fails if the opinion adopts one integrated use for the material stages.
P2. Training-core protection (65–78%). The opinion imposes no liability on training itself where works were lawfully acquired. Fails if lawful-source training draws liability without a substitution showing.
P3. Flank survival (75–85%). Acquisition or output claims survive summary judgment against at least one defendant. Fails if defendants win clean across every stage.
P4. The divided rule jointly (62–74%). P2 and P3 land together: some training protected, material edge exposure preserved. Fails if either side wins categorically.
Remedies
P5. Non-structural remedies (87–95%). No publisher case produces a model-destruction order that takes effect and survives direct appellate review. Fails if one does.
Timing and Opinion Shape
P6. Ruling timing (median May 2027). The decision window runs February 2027 at the 10th percentile to January 2028 at the 90th, scored against the docket date. Fails if the order lands before mid-December 2026 or after June 2028.
P7. Opinion architecture (61–74%). Factors one and four align within each analyzed stage and diverge across stages. Fails if a dispositive stage shows the two factors opposed.
Settlement Cascade
P8. Licensing cascade (58–70%). After a mixed ruling, at least three qualifying publisher settlements or licenses with OpenAI or Microsoft arrive within 18 months; a faster wave of five top-50 publishers within 12 months carries (45–60%). Unresolved if no mixed ruling issues. Fails on fewer than three.
The Licensing Market
S1. Retrieval prevalence (75–85%). A majority of new publisher agreements within 18 months of the ruling carry retrieval or real-time terms distinct from training terms, measured across at least three agreements with disclosed scope. Unresolved below that denominator. Fails if the majority are training-only.
S2. Rights differentiation (70–80%). At least three disclosed post-ruling deals separate current-content access from historical corpora through carve-outs, refresh obligations, or usage-based compensation. Unresolved below three disclosed deals. Fails if disclosed deals run undifferentiated.
Enforcement and Doctrine
S3. Instrument migration (60–75%). Within 18 months of the ruling at least three new publisher actions materially foreground Copyright-Management Information (CMI), contract, or state-law theories. Fails on fewer than three.
S4. Executive differential (50–65%). If liability reaches a remedy ruling, the remedy reasoning tracks the government’s innovation concerns more closely than the liability reasoning tracks its fair-use position. Unresolved if no remedy ruling issues. Fails if the influence pattern runs even or inverted.
S5. Doctrinal propagation (55–70%). At least two additional federal courts adopt stage-segmented fair-use analysis within 18 months of the ruling. Fails on fewer than two adoptions.
S6. Replacement tail (25–40%). A controlling appellate rule or federal statute replaces the district-level game by September 2029. Fails if the horizon passes without one.
The fourteen predictions form one structure. The ruling predictions establish the divided rule, the settlement and market predictions price it, and the enforcement predictions trace where pressure flows when federal remedies compress. Every entry grades from public sources.
III. The False Training Binary: Courts Choose the Use Before the Factors
Public debate frames the litigation as a binary: training is either fair use or mass infringement. The frame is false, and recent rulings have begun to abandon it. The operative question comes before the four factors: what is the use?
Modern AI systems fragment copyright into distinct acts. Developers acquire corpora and train models. Products then ground answers in retrieval and emit outputs that may reproduce protected expression.
Each pipeline act presents different facts to each factor. Training on ten million articles resembles Authors Guild v. Google: massive copying for a non-substituting capability. A grounded answer that replaces a paywalled story resembles Hachette v. Internet Archive: substitute delivery.
Unit-of-use selection is therefore the hidden zeroth factor. Courts choose between one integrated exploitation and four separable uses before weighing anything. The choice largely determines how the enumerated factors align.
Warhol v. Goldsmith pushed doctrine toward the specific challenged use and its commercial character. Bartz v. Anthropicoperationalized the split: transformative training beside independently unlawful retention of pirated copies. One defendant produced two uses and two answers.
The summary judgment contest before Judge Stein is a fight over the zeroth factor. Counsel should brief it as one. A segmented ruling supplies the legal foundation for everything that follows.
Working With MindCast
MindCast AI runs commissioned foresight simulations built on the same Cognitive Digital Twin method that produced this register. A commissioned run models the client’s specific contest: its actors, its filings, and its market. Outputs arrive as banded predictions with falsifiers and dated checkpoints, in the format this paper demonstrates.
For AI developers: pipeline segmentation audits that price exposure stage by stage against the route tree, provenance and output-control design against P2 and P3, and license sequencing against S1 and S2.
For publishers and content companies: claim-architecture review against P3 and S3, rights packaging and license design against S1 and S2, and settlement-timing analysis against P6 and P8.
For counsel: stage-specific pleading and evidence strategy keyed to P1 and P7, and remedy positioning against P5 and S4.
For investors and insurers: content-cost exposure models across the four routes, deal-scope diligence against S1 and S2, and replacement-hazard stress tests against S6.
For policymakers: intervention design in the space P5 leaves open, and state-instrument analysis against S3. The open space means provenance, retrieval, and attribution duties rather than training bans.
Engagements update at the register’s dated checkpoints, and every commissioned prediction carries its own falsifier and settlement source. Contact mcai@mindcast-ai.com.
IV. The Empirical Record: Factors One and Four Decide Fair Use Outcomes
Barton Beebe’s “An Empirical Study of U.S. Copyright Fair Use Opinions, 1978–2005” (2008) anchors the empirical record. Factors one and four aligned in 72.1% of 297 dispositive opinions, and the outcome followed those factors in all but one aligned case. Factor four matched the overall outcome in 83.8% of opinions and factor one in 81.5%.
Beebe’s “An Empirical Study of U.S. Copyright Fair Use Opinions Updated, 1978–2019” (2020) confirmed factor four’s continued dominance. Market effects remain central to fair-use outcomes in practice.
Beebe’s data establish alignment rather than mechanism. Beebe disclaimed statistical proof that judges decide first and conform the factors afterward. MindCast reads the alignment through allocation-first judging: courts identify which actor can prevent the harm without destroying the productive system, and the factors then express the allocation.
The allocation-first reading is an interpretation supported by coherence-based models of judicial reasoning. The interpretation generates a testable expectation. A segmented AI ruling should align factors one and four within each stage and split them across stages.
The predicted opinion favors developers at training and publishers at acquisition and substitutive output. Factor-by-factor doctrine alone does not predict the cross-stage pattern as directly. The four factors are the surface, and liability geometry decides.
V. The Coordination Gradient: Why Fair Use Protects Training but Not Retrieval
Wendy Gordon’s “Fair Use as Market Failure” (1982) explains fair use as a response to failed licensing markets. The AI litigation demands two modernizations. Both supply the paper’s mechanism.
Gordon’s first needed modernization separates bilateral transaction costs from system-wide coordination costs. Bilateral contracting between a developer and a major publisher is inexpensive. Multilateral coordination across tens of millions of works fails on fragmented ownership and incompatible price expectations.
Chicago School Accelerated — The Integrated Framework establishes coordination costs as analytically distinct from bargaining friction. The distinction matters here because low bilateral costs coexist with prohibitive system-level coordination costs. The AI-content market presents exactly that configuration.
Fair use tracks the resulting coordination-cost gradient. Doctrine protects the layer above the coordination threshold because prohibition would impose systemic costs no market can absorb. Premium retrieval and authenticated current content sit below the threshold, where markets form and liability prices access.
A dynamic market-failure model supplies the second modernization. Litigation builds the coordination infrastructure that moves particular uses out of fair use’s market-failure zone. The Anthropic settlement established a reference point near $3,000 per work for pirated-source acquisition.
Discovery verified provenance at industrial scale. Repeat contracting among OpenAI, News Corp, and the Associated Press created reference terms. A settlement is not a judicial price, but a focal point needs no adjudication to coordinate expectations.
Settlement also creates categories before it creates prices. The durable market signal is rights separation: historical training, current retrieval, and grounded delivery priced as distinct categories rather than one per-work number.
Kadrey v. Meta supplies the doctrinal limit. Owners cannot create a cognizable market for a transformative use merely by demanding payment. Licensing infrastructure narrows fair use only where it converts undifferentiated training into a separately identifiable and substitutive service.
Real-time retrieval, authenticated grounding, and protected-content delivery present the strongest claims under Kadrey’s limit. The refined expectation follows: coordination architecture expands licensing from the edges inward and may stop before non-substitutive training. Courts fighting the static circularity question are adjudicating a moving boundary.
Four conditions govern the equilibrium, where s denotes any pipeline stage from acquisition through output.
Fair use holds at stage s when C(s) > V(s). C(s) is the coordination cost of forming a licensing market at that stage, and V(s) is the substitution value a license would price. Training clears the inequality because C is prohibitive across tens of millions of works; retrieval fails it because C is low and V is observable.
Liability assigns upstream when B(s) < P(s) × L(s) and downstream avoidance capacity is near zero. B(s) is the developer’s burden of preventing harm at that stage. P(s) is the probability of the harm and L(s) is its magnitude, so their product is the expected harm. Section VI develops the behavioral extension that collapses the cost comparison to capacity.
The boundary moves: dC/dt < 0 at coordinated layers. The derivative dC/dt is the change in coordination cost over time, and a negative value means the cost falls. Posted prices, verified provenance, and repeated contracts each lower C. The set of stages satisfying the fair-use condition contracts from the edges inward while the training core holds.
The rule locks when ΔPayoff < ε for every actor. ΔPayoff is the gain any actor could capture by deviating alone, and ε is a threshold near zero. No player improves by breaking from the divided rule, and Section VII prices that condition across the route tree.
Collective rights organizations and standardized licenses extend the same mechanism. Every publisher plaintiff therefore faces a sequencing choice. Arguing licensing is impossible supports market-failure fair use and weakens damages.
Building the licensing market strengthens factor-four harm and narrows the doctrine litigated over. Sophisticated plaintiffs deploy the two positions in order rather than holding either absolutely.
The coordination-gradient model scales beyond journalism when four conditions recur. The content field carries a large historical corpus and concentrated owners of current material. Training and retrieval remain separable, and substitution is measurable at delivery.
The model weakens where those layers cannot be separated or ownership stays too fragmented to support contracting. Books, music, and images may therefore produce different prices and remedies without requiring a different theory. Code and other authenticated-data markets follow the same rule.
Fair use in the AI era is a moving boundary. The movement is predictable because the driving architecture is observable: prices posted, provenance verified, and contracts repeated. Executives and counsel should track the boundary as a market variable.
VI. Chicago School Accelerated Across the Pipeline
The integrated Coase-Becker-Posner framework assembles the mechanism into one causal account. Each layer maps onto a stage of the record. Mechanism precedes outcome throughout.
Coase locates the origin. The AI-content market carried relatively manageable bilateral transaction costs yet failed to coordinate at the system level. The market lacked shared prices, standardized rights, and trusted provenance.
Becker explains persistence. Under degraded coordination developers maximized ingestion because immediate returns exceeded expected liability discounted by adjudication lag. Paywall scraping and shadow-library acquisition were predictable responses to the payoff structure rather than isolated anomalies.
The Department of Justice (DOJ) intervention lowered developers’ expected liability and structural-remedy costs. Filed September 1 under 28 U.S.C. § 517, the Statement of Interest argues training is fair use and calls market-dilution theory deeply flawed.
Posner locates the correction. Liability migrates to the lowest-cost capable avoider, and behavioral incapacity makes the migration one-directional. Readers cannot inspect provenance or audit training pipelines.
Developers control the only scalable prevention surfaces at a small fraction of expected harm. S4 tests a narrower institutional prediction: DOJ influence should appear more strongly in remedy design than in substantive liability analysis.
Coordination failure created the conditions and incentive exploitation filled them. MDL consolidation now forces the parties to litigate against a shared evidentiary record. Allocation to the controllable layers is the predicted institutional correction.
VII. How the Contest Ends: The Divided-Rule Settlement and Its Falsifier
The Dual Nash-Stigler Equilibrium Architecture supplies the termination condition and converts settlement from narrative into a testable stopping rule: a contest ends when no actor improves through unilateral deviation. Game theory sets the payoff structure, and behavioral economics sets the decision rules that determine which equilibrium actors select.
The emerging cases create the conditions for a divided rule. The maximal publisher position faces mass-licensing infeasibility, adverse precedent, and the executive posture. The maximal developer position faces the $1.5 billion Anthropic litigation-exposure reference point and output claims that survived dismissal.
The predicted settlement basin is the fair-use settlement equilibrium the title names. Inside it publishers monetize current authenticated content at premium terms. Developers preserve historical training while purchasing retrieval access and adding output controls. Remedies resolve as damages plus licensing plus architectural safeguards rather than model destruction.
The simulation's route tree assigns the divided rule 55% as the trunk outcome, neither side improves its position by defecting from the divided rule. A defense-heavy disposition, in which frequency evidence governs and output claims collapse, carries 22%; a publisher-heavy result carries 15%; early replacement carries 8%.
The divided-rule claim is falsifiable. The prediction fails if controlling authority imposes liability on lawful-source non-substitutive training. The rule equally fails if controlling authority excuses independently unlawful acquisition because later training is transformative.
A categorical holding in either direction would cut against the gradient and favor a categorical ownership rule. The endgame is an equilibrium event rather than a verdict. Timing is the tractable question, and P6 bands it.
VIII. Whether the Divided Rule Spreads: The Propagation Test
A divided rule announced in one courtroom is not yet a regime. Randal Picker’s generative account of norm adoptionposes the governing question: does a seed configuration become self-sustaining or decay? Modern computational foresight operationalizes the question for doctrine.
The current seed comprises Thomson Reuters on competitive substitution, Bartz on acquisition-training separation, and Kadrey’s open market-dilution door for news. The coming MDL ruling adds the first pipeline-segmented adjudication at scale. A settlement layer propagates the divided rule through contracts faster than appellate review propagates it through doctrine.
Decay pressures stand against the seed. A genuine circuit split could invite Supreme Court review that replaces the game. Congressional licensing legislation would moot the judicial equilibrium entirely.
Apparent momentum is not the same as durable adoption. S5 tests whether stage segmentation propagates across federal courts; P8, S1, and S2 separately test propagation through contracts. The distinction determines whether the paper describes a settlement or a regime.
IX. Enforcement Competition and Instrument Migration
When federal remedies compress, enforcement migrates rather than ends. The migration paths are already visible in claim architecture. The Seattle Times complaint pairs copyright counts with CMI claims that the DOJ’s fair-use position does not itself resolve.
Contract theories attach to paywall circumvention independently of § 107. State consumer-protection statutes and attorney general coalitions offer forums where national-innovation framing carries no privileged weight.
Federal Inaction Has Elevated State Authority on Consumer Protection, Antitrust, and Market Integrity documents the pattern across antitrust and AI safety. The same propagation principle now reaches copyright: federal compression changes the instrument and price of enforcement without ending the contest.
A favorable federal fair-use ruling therefore purchases less finality than its advocates expect. The durable resolution remains the licensed equilibrium of Section VII.
The equilibrium narrows the federal training contest; it does not end publisher enforcement. Mapping post-compression enforcement flows bridges to MindCast’s forthcoming analysis of AI distillation and intellectual-property enforcement. Policymakers should expect pressure deflected federally to resurface in state instruments.
X. Operational Consequences: What the Equilibrium Instructs Each Side to Do Now
The divided rule produces different instructions for developers and publishers. Developers should stop treating training data as one undifferentiated legal object, and publishers should stop making universal compensation for historical training the entire case.
The lowest-cost durable developer defense is demonstrable control over the stages where substitution occurs. Establish lawful acquisition and auditable provenance before training. Segregate disputed corpora so acquisition taint cannot contaminate the training record.
Separate training copies from retrieval indexes and production grounding systems. License current, paywalled, and frequently retrieved content first. Measure output overlap and referral effects before plaintiffs measure them in discovery.
Preserve CMI through retrieval and attribution systems. Price historical corpus access differently from real-time retrieval rights, because the divided rule prices both.
Publishers should concentrate evidence where the divided rule preserves liability: unauthorized acquisition, protected-content retrieval, and substitutive outputs. Preserve access logs and documentation of paywall restrictions. Test ordinary-user substitution rather than adversarial regurgitation, which courts increasingly discount.
Package current content, authentication, and provenance as distinct products. Negotiate separate prices for training, retrieval, and real-time access. Use the Anthropic settlement as acquisition-risk evidence rather than a transferable content price. Keep the CMI, contract, and state-law options that survive federal remedy compression.
Publisher sequencing turns on a conflict between market-failure advocacy and market-building. The market-failure argument and the market-building strategy cannot run at full strength simultaneously. Sophisticated plaintiffs deploy them in order.
The equilibrium is an architecture to build toward rather than a forecast to await. Parties who internalize the divided rule before it is announced will set the terms on which the rest transact. Executives own the pipeline segmentation, and counsel own the evidentiary posture.
XI. Risk Mitigation
The initial route tree assigns 45% to non-trunk outcomes. A separate 25–40% replacement hazard through 2029 can attach to any initial route and should not be added to that 45%. Two alternative dispositions and one cross-cutting hazard dominate the risk surface, and each carries a mitigation that costs little if the trunk holds.
The defense-heavy branch carries 22%: frequency evidence governs and output claims collapse. Publishers mitigate by building ordinary-user substitution records before the ruling rather than after. Developers mitigate by not overpaying for peace the branch would deliver on its own.
The publisher-heavy branch carries 15%: a cognizable training market emerges and training-stage exposure opens. Developers mitigate by pre-negotiating historical-corpus options and segregating disputed corpora. Publishers mitigate by preparing damages models that survive circularity scrutiny.
The replacement hazard carries 25–40% through 2029: an appellate rule or federal statute resets the game. Both sides mitigate by drafting licenses with change-of-law adjustment terms. Policymakers should treat the window before appellate resolution as the design opportunity.
Measurement risk rounds out the surface. Confidential deal terms and the partially sealed record limit observation. The register therefore grades on public proxies: rights architecture, claim structure, and citations. Risk in this contest is positional rather than existential. The equilibrium’s shape holds across branches while the branches move prices and timing.
XII. What to Watch
Judge Stein’s first consequential choice is whether to analyze one integrated use or several pipeline stages, and every dated observable below feeds it. September 11 and 18, 2026 bring the stay filings that help determine which plaintiff record drives the opinion. September 17 brings public re-filing of the summary judgment record, including whatever survives redaction of the plaintiffs’ fifth claimed stage.
November 6 closes reply briefing. An argument date follows, and the decision window opens in February 2027. After the ruling, the first three disclosed publisher deals test S1 and S2, and the claim mix of new complaints tests S3.
Conclusion
The litigation will not end with a yes or a no on training. It will end with an allocation: which pipeline layers can bear liability without destroying the productive system, and which actor controls prevention at each layer. Prevention capacity answers both questions.
Protect computational learning. Price controlled access. Constrain substitutive delivery.
The litigation builds the coordination infrastructure that converts the allocation from doctrine into market structure, and the excusable market failure erodes from the coordinated edges inward while the training core holds. The Section II register carries the released Simulation Predictions, and public re-filing of the summary judgment record arrives September 17, 2026. Each checkpoint that follows grades an entry, and the register tells readers exactly where to look.
Sources
MindCast AI
Chicago School Accelerated — The Integrated, Modernized Framework of Chicago Law and Behavioral Economics(2025). Establishes coordination costs as analytically distinct from transaction costs and integrates Coase, Becker, and Posner into the allocation framework this paper applies.
Chicago School Accelerated — Integrated Application: AI Hallucinations, AI Copyright, and Crypto ATMs (2025). Registered the substitution turn and the output-first liability migration this paper scores in Section I.
The Dual Nash-Stigler Equilibrium Architecture (2026). Supplies the settlement termination logic and inquiry-sufficiency discipline governing the released simulation.
Federal Inaction Has Elevated State Authority on Consumer Protection, Antitrust, and Market Integrity (2026). Documents the enforcement-migration pattern Section IX extends to copyright.
Randy Picker, Visionary (2026). Grounds the propagation question in Picker’s generative account of norm adoption.
External
In re OpenAI, Inc. Copyright Infringement Litigation, No. 1:25-md-03143 (S.D.N.Y.).
The Seattle Times Co. v. OpenAI, Inc., No. 1:26-cv-07644 (S.D.N.Y.).
Barton Beebe, “An Empirical Study of U.S. Copyright Fair Use Opinions, 1978–2005,” 156 University of Pennsylvania Law Review 549 (2008); “An Empirical Study of U.S. Copyright Fair Use Opinions Updated, 1978–2019,” 10 NYU Journal of Intellectual Property and Entertainment Law 1 (2020).
Wendy J. Gordon, “Fair Use as Market Failure: A Structural and Economic Analysis of the Betamax Case and Its Predecessors,” 82 Columbia Law Review 1600 (1982).
Bartz v. Anthropic, order on fair use (N.D. Cal. 2025).
Kadrey v. Meta Platforms, order on summary judgment (N.D. Cal. 2025).
Authors Guild v. Google, 804 F.3d 202 (2d Cir. 2015).
Hachette Book Group v. Internet Archive (2d Cir. 2024).
Andy Warhol Foundation for the Visual Arts v. Goldsmith, 598 U.S. 508 (2023).