AI Provenance: The Copyright Fight That Will Define What AI Was Allowed to Learn

The most consequential AI lawsuits may not determine what artificial intelligence can generate. They may determine what artificial intelligence was lawfully allowed to learn.

For the past several years, public debate has focused on AI outputs: hallucinations, bias, fabricated citations, copyright infringement, privacy leaks, and safety failures. Those concerns are real. But the more fundamental legal question arises earlier in the causal chain. Before an AI system produces anything, it must first be trained. The legitimacy of that training process is becoming one of the defining legal questions of the AI era.

That is why the author and news organization lawsuits matter. These cases are not merely about whether ChatGPT or another model can reproduce protected expression. They are about whether copyrighted works were copied, acquired, retained, or used lawfully during model development. In Tremblay v. OpenAI, Inc., 716 F. Supp. 3d 772 (N.D. Cal. 2024), the court addressed early claims by authors alleging that OpenAI used copyrighted books to train its models. The court dismissed most of the claims, including vicarious infringement and the Digital Millennium Copyright Act count, leaving the direct infringement claim to proceed. Those author cases were later consolidated into the multidistrict litigation now proceeding as In re OpenAI, Inc., Copyright Infringement Litigation, No. 25-md-3143 (SHS) (OTW) (S.D.N.Y.), before Judge Sidney H. Stein and Magistrate Judge Ona T. Wang.

Discovery in that MDL shows why this fight is bigger than ordinary copyright litigation. On January 5, 2026, Judge Stein affirmed Magistrate Judge Wang’s orders requiring OpenAI to produce a sample of 20 million de-identified ChatGPT consumer logs. That sample represents roughly 0.5 percent of the tens of billions of logs OpenAI preserves, produced after the news plaintiffs, led by The New York Times Company and the Chicago Tribune, initially sought 120 million. The court concluded that even logs that do not reproduce plaintiffs’ works could be relevant to OpenAI’s fair use defense. In re OpenAI, Inc., Copyright Infringement Litigation, No. 25-md-3143 (SHS) (OTW), 2026 WL 21676 (S.D.N.Y. Jan. 5, 2026). Judge Stein wrote that “no case law requires the court to order the least burdensome discovery possible,” and he distinguished the consumer logs from surreptitious wiretaps because users voluntarily submitted their communications. The reach of the ruling has since grown. On March 9, 2026, the court granted a further motion to compel and ordered OpenAI to produce reservoirs of tens of millions of additional logs on top of the original sample.

Provenance asks a simple but powerful question. Where did the model’s knowledge come from, and was it lawfully acquired? Explainability tells us why a model reached a decision. Provenance tells us whether the knowledge enabling that decision came through a legitimate process. One concerns reasoning. The other concerns legitimacy. Both are becoming essential AI governance capabilities.

Across mature legal systems, legitimacy rarely depends only on the final product. It depends on the process that produced it. Administrative law cares about procedure. Corporate governance cares about controls. Due process cares about fairness before the result. Intellectual property law cares not only about what is distributed, but also about what was copied, transformed, retained, or commercially exploited along the way.

Artificial intelligence should not be treated differently.

Bartz, Kadrey, and the New Line

Recent AI copyright decisions already point toward this line. In Bartz v. Anthropic PBC, 787 F. Supp. 3d 1007 (N.D. Cal. 2025), Judge William Alsup distinguished three uses: training on works, which he found fair; digitizing lawfully purchased print books, which he also found fair; and building a permanent central library from pirated copies, which he held was not excused by fair use. The court found the training use transformative on that record, but drew a sharp line against the pirated library copies.

What happened next is the point businesses cannot miss. Rather than take the piracy claims to trial, Anthropic settled. The exposure was staggering. As the Kluwer Copyright Blog explained, the roughly 482,460 in-class books multiplied by the $150,000 statutory maximum for willful infringement produced theoretical liability exceeding $70 billion. The settlement is valued at $1.5 billion, the largest copyright recovery on record. Class counsel at Susman Godfrey described it as “approximately $3,000 for each work,” anticipating “approximately 500,000 works in the class,” and it requires Anthropic to destroy the pirated datasets. Judge Alsup granted preliminary approval on September 25, 2025. After his retirement, the case was reassigned to Judge Araceli Martínez-Olguín, who held a fairness hearing on May 14, 2026 and took final approval under submission. As of this writing in July 2026, final approval remains pending, with disputes over attorneys’ fees, including a reduced request of $187.5 million, and a small number of late opt-outs still open. The lesson is already clear. Anthropic largely prevailed on training as a concept, yet still faced existential exposure. The driver of that exposure was acquisition, not training. Provenance, not output.

In Kadrey v. Meta Platforms, Inc., 788 F. Supp. 3d 1026 (N.D. Cal. 2025), Judge Vince Chhabria granted Meta summary judgment on fair use, but he was emphatic that the ruling did not establish that training on copyrighted works is lawful. He wrote that the plaintiffs simply made the wrong arguments and failed to develop a record on market harm. He went further, suggesting that in many cases unauthorized training will likely be illegal because of market dilution, the flooding of a market with competing works. The win was narrow. It bound only the thirteen named plaintiffs, it was not a class ruling, and claims tied to Meta’s torrenting and distribution of pirated files remain live. A stronger evidentiary showing, the court signaled, could change the result.

The appellate courts are now engaging. In Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc., the Third Circuit heard oral argument on June 11, 2026, the first federal appeals argument on whether copying copyrighted material to train an AI system can be fair use. The appeal follows Judge Stephanos Bibas’s February 2025 ruling that Ross’s copying of 2,243 Westlaw headnotes to build a competing legal research tool was not fair use. A decision will bind courts in the Third Circuit and will influence every training data case.

From Litigation Risk to Disclosure Obligation

Regulators are reinforcing the same signal. California’s Generative Artificial Intelligence Training Data Transparency Act, AB 2013, took effect on January 1, 2026 and requires developers to publish a high-level summary of the datasets used to train generative AI systems made available to Californians since 2022. In the European Union, the AI Act’s obligations for general purpose AI model providers began applying on August 2, 2025, requiring a published summary of training content and a copyright compliance policy. Provenance is no longer only a litigation risk. It is a disclosure obligation.

That is the legal signal businesses should not miss. The next generation of AI risk will not be limited to whether the model gives a bad answer. It will include whether the organization can prove that the model, dataset, workflow, and vendor ecosystem were built on lawful, auditable foundations.

For law firms, corporate legal departments, municipalities, and regulated businesses, this means AI governance must go beyond reviewing the output. Output review is necessary, but it is no longer enough.

The AI Provenance File

The practical governance tool is an AI provenance file. Every serious AI deployment should have a record that documents the following:

That may sound administrative. It is actually strategic. This inventory maps naturally to the COUNSEL framework for AI governance, whose pillars, Confidentiality, Oversight, Understanding, Notice, Scrutiny, Equity, and Learning, operationalize the duties described in ABA Formal Opinion 512 on generative AI tools.

The companies that win the next stage of AI adoption will not simply be the ones with the largest models or the most computing power. They will be the ones with governable intelligence: intelligence whose origin, acquisition, development, and deployment can withstand legal scrutiny, regulatory oversight, client expectations, and public trust.

Trust in AI will not be built by marketing claims. It will be built by evidence. Can the organization show where the system’s knowledge came from and that the data was lawfully obtained? Can it separate licensed data from scraped data, public data from confidential data, and user data from training data? Can it prove that its AI workflow respects privilege, privacy, intellectual property, and professional obligations? An organization that cannot answer those questions may have an AI capability. It does not yet have an AI governance program.

Ultimately, the defining question for AI will not be whether it is intelligent enough to transform society. It will be whether the process by which that intelligence was created can itself withstand the rule of law.