This post argues that the legality of the data used to train artificial intelligence (AI) systems is no longer a private copyright matter between technology firms and rightsholders, but a public-law problem that public buyers must address through procurement. Two terms recur throughout. “Fair use” is a doctrine in U.S. copyright law that, in defined circumstances, permits the unlicensed use of copyrighted material; whether a use qualifies turns on a four-factor statutory test in which the effect on the market for the original work is one factor. “Market evidence” is the body of institutionally validated signals, about substitution, licensing feasibility, valuation, and compliance expectations, that courts construct through litigation and that public buyers can, in turn, build into contracts and audits.
The analysis is grounded in U.S. copyright law and U.S. judicial practice. The doctrinal apparatus discussed below is therefore American. The underlying governance question, namely how public buyers translate unstable legality claims about AI training data into enforceable procurement controls, travels well beyond that jurisdiction, and European and other readers should be able to take the institutional logic across even where the substantive copyright rules differ.
Public agencies now buy and deploy AI systems for high-impact functions: eligibility screening, fraud detection, licensing, case triage, and service automation. Legality assumptions travel through procurement contracts into state-facing systems, and if those assumptions fail, accountability fails with them.
This blog post answers two linked questions: First, how is market evidence crafted? Second, what actions lead to the emergence of specific types of market evidence? The short answer is that market evidence is crafted by human judges through procedural choices that define relevance, proof burdens, and remedies. Different judicial actions produce different evidence artifacts, which can then be translated into procurement controls.
1) How market evidence is crafted
In AI copyright disputes, “market evidence” means institutionally validated signals about four things: market substitution, licensing feasibility, valuation, and compliance expectations. These signals are not found ready-made in the world. They are constructed through legal processes.
The doctrinal baseline, 17 U.S.C. §107, the fair-use rule of the U.S. Copyright Act, sets out a four-factor test comprising (1) the purpose and character of the use, (2) the nature of the copyrighted work, (3) the amount and substantiality of the portion used, and (4) the effect of the use upon the potential market for or value of the copyrighted work. Courts apply this test to decide whether a particular unlicensed use qualifies as fair use. Market effect is built into fair-use analysis. U.S. Supreme Court doctrine in Andy Warhol Foundation for the Visual Arts, Inc. v. Goldsmith Foundation v. Goldsmith held that competition in commercial licensing markets bears on the first factor. That factor asks whether a new use is sufficiently “transformative” of the original. The longstanding doctrine traceable to Harper & Row Publishers, Inc. v. Nation Enterprises, the 1985 U.S. Supreme Court decision holding that The Nation’s unauthorised pre-publication excerpt of President Ford’s memoirs was not fair use, confirms that adverse market effect may carry decisive weight in the fair-use balance, particularly where the secondary use usurps a traditional or licensed market. Together, they reinforce that market-facing reasoning carries weight across the fair-use inquiry. But doctrine alone does not produce usable governance evidence. Judges do, through action.
Action type A: Issue framing
When judges frame the legal dispute, they decide what counts as probative market reality. Framing AI training as transformative use versus substitutive market behaviour changes the evidentiary universe. Once that frame is set, parties must build evidence that fits it. For procurement actors, this means that “fair use” claims cannot be assessed abstractly. They must be tested against the same framed market question that courts are using.
Action type B: Burden allocation and admissibility thresholds
Judges also decide who must prove what and how much proof is enough. In Kadrey v. Meta Platforms, Inc., the June 2025 Phase 1 ruling (granting Meta summary judgment on fair use because plaintiffs produced no concrete market-harm evidence) alongside ongoing Phase 2 briefing (including filings tied to Dkt. 690–694) demonstrates how burden allocation shapes the evidentiary record. That ruling is itself a licensing-feasibility signal: absent sufficient market-harm evidence, fair use will be found. This action type creates licensing-feasibility signals: practical evidence about whether a lawful market path is demonstrated rather than merely asserted.
Action type C: Settlement approval and remedial design
Court-supervised settlements produce governance data beyond the immediate parties. In Bartz v. Anthropic PBC, recent filings around Dkt. 613–615 publicly reference settlement-administration updates and final-approval related motion practice. The settlement, covering approximately 500,000 works obtained from pirated databases and valued at $1.5 billion, without licensing any future AI training, generates valuation-relevant signals and compliance expectations that public procurers can operationalise.
So, how is market evidence crafted? It is crafted through judicial framing, burden design, and remedy architecture. That is the operational answer to the first organising question.
2) What actions produce what evidence types
The second organising question asks for a specific mapping. A practical matrix is:
- issue-framing -> market substitutability indicator
- burden-shifting -> licensing feasibility signal
- approval-of-settlement -> price signal
- remedial-design -> compliance expectation
This is not only a legal map. It is an institutional production map. Organisational scholarship helps explain why. Hopf, Joshi, Shollo, and Stelmaszak describe “data-based craft” as iterative production of data artefacts, tools, and outputs. Judicial market evidence works similarly: it is iteratively crafted through filings, rulings, standards, and remedy structures. Bourgoin, Bencherki, and Faraj show that authority is performative and enacted through situated practices. In AI governance, judicial authority is performed when courts convert contested claims into recognised evidence forms that others must operationalise.
That is why different judicial actions generate different evidence types. Each action configures a different part of the institutional pipeline.
3) Procurement-ready translation
If judges craft market evidence, procurement systems should ingest it deliberately. A procurement-ready model needs three operational layers.
Layer 1: Due diligence questions
Public buyers should require vendors to answer at least these questions before award:
- What exact data acquisition pathways were used for pre-training and fine-tuning corpora?
- What evidence supports non-substitutive use in relevant licensing markets?
- Which assumptions in your legal position are tied to active litigation risk?
- What is your procedure for newly disputed or newly licensed works?
- What evidence demonstrates that lawful licensing was infeasible where rights were not cleared?
Layer 2: Contract clauses
A minimum contract package should include:
- Provenance evidence clause: maintain auditable source lineage and retention policies.
- Legal-change update clause: mandatory notice when new rulings alter legality assumptions.
- Contested-content mitigation clause: timed suspension and remediation protocol.
- Claims-process support clause: support for notice, objection, and claimant communication where required.
- Escalation clause: human legal review trigger before high-impact automated outputs continue under contested legality.
Layer 3: Audit triggers and remedy pathways
Governance cannot stop at drafting. Trigger mechanisms are necessary:
- Trigger targeted audit when evidence quality falls below agreed standards.
- Trigger re-certification when litigation developments materially change market assumptions.
- Trigger contractual remediation when provenance gaps affect rights-sensitive public decisions.
This operational translation moves procurement from trust-based vendor narratives to evidence-based accountability.
4) Falsification conditions and scope limits
The framework in this post should be treated as conditional. It weakens or fails under at least four conditions. First, if courts systematically deprioritise market-oriented fair-use reasoning, the action-to-evidence mapping loses explanatory force. Second, if key filings become inaccessible or non-auditable, procurement systems cannot reliably ingest judicial signals. Third, if settlements are structurally non-representative or purely strategic without compliance content, they should not be read as governance evidence. Fourth, if agencies cannot translate legal signals into contract controls and audit triggers, the model remains descriptive, not operational. These falsification conditions define the boundary of the argument and prevent over-generalisation.
5) Humanity in the automated state
The humanity dimension is institutional, not decorative. Human judges remain essential because they provide normative and procedural functions that automated systems cannot legitimately replace: interpretation of contested rights, calibration of proportional remedies, and maintenance of contestability.
In an automated state, agencies may automate transactions, but they cannot automate away legal responsibility. Human oversight must remain mandatory where legality claims are unstable, where evidence is contested, and where rights impacts are material. Human remedy channels must remain accessible to people and creators affected by procurement-mediated AI decisions.
In that sense, humanity appears in three linked locations: human judgment (what evidence counts), human oversight (when deployment can continue), and human remedy (how harms can be challenged and repaired). Without those three, automated governance may be efficient but not accountable.
Conclusion
Human judges shape AI market relations by crafting the evidence architecture that markets and public buyers later rely on. Different judicial actions generate different evidence types, and those types can be translated into procurement due diligence, contract design, and audit triggers. Courts, procurers, and vendors each have a role: courts craft legible signals; procurers operationalise them; vendors treat legality claims as ongoing evidence obligations. That is the path to procurement-ready AI fair use in a genuinely human-centred automated state.
Alex Chanhou Lou
Dr. Alex Chanhou Lou is a Macau Fellow Researcher at the University of Macau and a Digital Life Initiative Fellow Researcher at Cornell Tech, Cornell University. His research focuses on privacy law, constitutional law, cyber law, and AI governance, with particular attention to the historical development of privacy rights, data governance, and the legal regulation of algorithmic systems. He holds a Ph.D. in Law from Tsinghua University and has previously been a Visiting Scholar at Duke University School of Law.
