Fair Use and Competition in AI-Disrupted Markets

A bipartisan bill, the Transparency and Responsibility for Artificial Intelligence Networks Act (TRAIN), introduced in January 2026, would give content creators subpoena power to compel disclosure from AI companies. If it passes, more copyright holders will have a legal mechanism to find out whether their work was used for AI training.

At first glance, this may seem like a power that would entitle more copyright holders to claim payments from AI developers. In reality, however, knowing that your work was used without prior permission is far from enough.

When deciding cases of fair use, courts weigh four key factors: the purpose of the use, the nature of the original work, how much was taken, and the effect such use has had on the material’s market value. Recent rulings in US courts have reaffirmed that fair use remains a pillar of innovation and cannot be dismissed easily. The spotlight is especially on the market harm factor and proving it.

Market Harm as the Main Battleground

AI copyright decisions from the Northern District of California show courts taking different approaches to fair use analysis. In Kadrey v. Meta, Judge Chhabria called market harm “the single most important element of fair use.” Judge Alsup in Bartz v. Anthropic, on the other hand, weighed all four factors more evenly. But both judges agreed on this: plaintiffs can’t just claim harm – they need to prove it happened or is probable.

The evidence requirement matters for AI developers, especially cash-strapped startups. If harm must be proven rather than presumed, developers can make design choices to avoid it. The decisions suggest that developers can reduce their risk by acquiring data from legal sources, designing products that serve distinct purposes from the copyrighted work, and implementing guardrails to prevent the reproduction of large chunks of text.

Both the Bartz and Kadrey courts found that AI training qualifies as “transformative use” under copyright law. With that, the focus is increasingly shifting to the fourth fair use factor: market harm. Recent AI copyright battles illustrate this. The claims increasingly center on the idea that verbatim reproductions of copyrighted work harm the publishers’ market value.

These cases remain to be decided. What matters is that publishers increasingly understand that, if they want to win, they need to claim two things: that AI outputs effectively replace the need to access original works and that, as a result, copyright holders suffer concrete economic harm.

Evidence Requirements

Both Bartz and Kadrey emphasize that market harm must be demonstrated, not assumed. In Kadrey, extensive testing showed Meta’s Llama reproduced no more than 50 tokens from plaintiffs’ works, and only 60% of the time under coaxing prompts designed to get the model to reproduce the original work.

Judge Alsup in Bartz focused on whether Anthropic’s Claude actually delivered infringing text to users – plaintiffs did not contend this had occurred. Without reproduction, it becomes harder to claim substitution for the original work.

This evidentiary approach shows that, even when copyright holders demonstrate that their works were used in training, they do not necessarily have a strong case for infringement. If the resulting AI system does not produce outputs that cause identifiable market damage, mere use is of little significance under the law.

When Market Harm Is Recognized

In Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc., the District Court for the District of Delaware rejected Ross Intelligence’s fair use defense after Ross used Thomson Reuters’ Westlaw headnotes to train an AI legal research tool that directly competed with Westlaw. Both factor 1 (purpose and character) and factor 4 (market effect) of the fair use analysis were crucial to the ruling.

Circuit Judge Stephanos Bibas found Ross’s use was not transformative because it created a direct market substitute. Ross initially sought to license Westlaw’s content, but Thomson Reuters specifically refused because Ross was its competitor. The alignment between the original materials’ purpose and the AI product’s purpose also supports the claim of potential harm.

Conversely, when AI products target markets different from those used for training, establishing market substitution becomes difficult. In Bartz and Kadrey, general-purpose language models served fundamentally different functions than individual books used for training. This distinction may prove crucial – the further removed an AI system’s purpose from its training data sources, the harder it is to demonstrate market substitution.

The “Licensing Market” Argument Rejected

Both courts explicitly rejected arguments that AI developers harm potential licensing markets for training data. Judge Chhabria explained that treating lost licensing fees as harm would make fair use analysis circular, automatically favoring copyright holders. Judge Alsup, for his part, found that a market for licensing books specifically for AI training “is not one the Copyright Act entitles authors to exploit.”

Courts declined to treat voluntary licensing arrangements as establishing legal entitlement to fees, at least where use is sufficiently transformative. These rulings demonstrate that the emerging licensing market does not automatically entitle copyright holders to bar fair use of their work.

Strategic Implications

For copyright holders, the strongest cases will be those where market substitution is clearly identifiable. They might strategically focus on AI systems whose outputs most closely approximate their original works rather than pursuing broad challenges to training itself.

If the TRAIN Act becomes law, copyright holders would gain discovery tools to investigate how their works are used. However, obtaining information would only be the first step. Demonstrating market harm would remain central to the success of any infringement claim.

For AI developers, recent decisions provide a framework for reducing exposure. First, ensure lawful data sourcing. Both Bartz and Kadrey distinguished between using works for training (potentially fair use) and acquiring them through piracy . Judge Alsup found that Anthropic’s downloading from pirate sites was “inherently, irredeemably infringing,” even though subsequent training might be fair use.

Second, design products for purposes different from training data sources. An AI system helping users draft documents serves different purposes than novels or articles in its training data. A system that simply retrieves or reproduces those works does not.

Third, implement safeguards preventing substantial verbatim reproduction. The Kadrey court noted Meta’s system reproduced minimal content even under adversarial testing, supporting fair use. Developers allowing systems to regurgitate large portions of copyrighted works may face significantly greater legal risk.

Conclusion

The TRAIN Act may soon give copyright holders tools to discover whether their works were used for AI training. However, recent decisions make clear that such a discovery would be only the beginning. The emerging US framework centers on market harm, requiring demonstration of identifiable economic damage rather than mere training use.

AI developers should focus on three things: obtain your data legally, build products that serve purposes beyond your training materials, and prevent your systems from reproducing long passages verbatim. Copyright holders, on the other hand, will have the strongest cases when they can show an AI product actually replaces their work in the market.