The Uncertain Future of AI

The debate over AI copyright infringement has reached a pivotal moment. In late June 2025, two landmark summary judgment rulings from the Northern District of California, Bartz v. Anthropic and Kadrey v. Meta Platforms, Inc. became the first federal decisions to directly address whether generative AI platforms’ use of copyrighted works for training constitutes copyright infringement or fair use. These rulings, alongside the earlier Thomson Reuters v. Ross Intelligence decision (involving non-generative AI), now provide the most authoritative guidance for AI litigation attorneys, content creators, and AI developers navigating this rapidly evolving terrain.

The Legal Framework: Copyright Infringement and Fair Use

At the heart of every AI copyright infringement dispute is the Copyright Act’s fair use doctrine, codified in 17 U.S.C. § 107. Courts weigh four non-exclusive factors:

1. Purpose and character of the use, including whether it is commercial or nonprofit educational
2. Nature of the copyrighted work
3. Amount and substantiality of the portion used
4. Effect of the use upon the potential market for or value of the copyrighted work

No single factor is dispositive; courts engage in a holistic, fact-intensive inquiry, balancing the interests of copyright holders against the societal benefits of new technology.

The Bartz and Kadrey Decisions: A Tale of Two Fair Use Analyses

In Bartz v. Anthropic, the Court found that using copyrighted books to train large language models (LLMs) was “quintessentially transformative,” analogizing it to a human reading books to gain knowledge and then creating new works. The court held that:

• Training LLMs with copyrighted works is fair use, regardless of whether the books were lawfully purchased or pirated (though the library use of pirated books was not fair use and remains for trial).
• The court placed significant weight on the transformative purpose of AI training, reasoning that the LLMs did not output infringing copies and thus did not harm the market for the original works.
• The amount and substantiality factor was analyzed in line with Google Books: the focus was not on the amount ingested for training, but on whether the outputs made substantial portions accessible to the public. Since plaintiffs did not show that AI outputs copied their books, this factor favored fair use.
• Market effect was minimized because there was no evidence of AI-generated knockoffs or direct substitution.

Kadrey v. Meta Platforms: Market Harm Takes Center Stage

Judge Chhabria, in Kadrey v. Meta Platforms, took a more skeptical approach:

• While acknowledging that LLM training is highly transformative, the court emphasized that market harm is the most important factor in the fair use analysis.
• The court warned that generative AI has the potential to “flood the market with endless amounts of images, songs, articles, books, and more,” thereby undercutting the incentive for human creativity.
• However, the plaintiffs failed to present evidence that Meta’s Llama models could regurgitate substantial portions of their works or that the outputs diluted the market for their books. As a result, the court granted summary judgment for Meta, but explicitly limited its ruling to these facts and plaintiffs, noting that future cases with a stronger market harm showing could reach a different result.

The Test in Practice: What Are Courts Actually Doing?

The fair use test especially as applied to AI copyright infringement has become intensely fact-specific:

• Transformative Use: Both courts recognize that training AI is transformative, but disagree on how much this should weigh against potential market harm.
• Market Harm: Kadrey elevates market effect as the “single most important element,” signaling that future plaintiffs who can show AI outputs dilute or substitute for their works may prevail.
• Evidence Required: Plaintiffs must present concrete evidence of market harm, not just speculation or theoretical risks.

Thomson Reuters v. Ross Intelligence Case: Non-Generative AI and the Limits of Fair Use

The Thomson Reuters v. Ross Intelligence case, while not involving generative AI, is instructive for its detailed analysis of direct copyright infringement and fair use in the context of AI-powered legal research tools. The court:

• Found Ross directly infringed thousands of Westlaw headnotes by copying and using them to train its AI search engine.
• Rejected the fair use defense, holding that Ross’s use was commercial and not transformative, as it created a competing legal research product rather than a new or different purpose.
• Emphasized that copying for the purpose of developing a direct market substitute weighs heavily against fair use.

This case underscores that AI systems designed to compete directly with the original works or markets are less likely to be protected by fair use.

The Likelihood of Finding AI Copyright Infringement

Based on recent Court rulings, the trend would suggest:

• AI platforms are more at risk of copyright infringement liability if:

• • They train on copyrighted works without permission;
  • The outputs of the AI system are shown to substitute for, dilute, or compete with the original works in the marketplace; and
  • Plaintiffs can demonstrate actual or likely market harm with empirical evidence.

• AI platforms are more likely to prevail on fair use if:

• • Their use is highly transformative and does not result in outputs that substitute for or reproduce the original works; and
  • Plaintiffs cannot demonstrate market harm beyond speculation.

AI Copyright Infringement, Fair Use, and Legal Risks

For those with questions about whether a particular situation is likely to result in AI copyright infringement, the legal landscape remains unsettled. At this point, each copyright infringement case is being considered as a stand-alone case specific basis but the key takeaways for content creators, copyright holders, and AI developers are:

• The fair use defense is highly fact-dependent in the context of AI systems.
• Market harm—actual or potential—is now the pivotal factor in determining infringement.
• Evidence matters: Plaintiffs must move beyond theoretical arguments and provide concrete proof of how AI outputs impact the market for their works.
• The transformative nature of AI training is not a panacea; courts will scrutinize whether the outputs compete with or substitute for the originals.

As more cases proceed to trial and appellate courts weigh in, the question of whether there is AI copyright infringement will continue to evolve which may reshape the balance between technological innovation and the rights of creators.

If you have specific questions or are seeking legal advice on copyright infringement, fair use, or to discuss how your business can mitigate risk when using AI systems, contact our Intellectual Property team.