AI Training and Fair Use

Why It Matters

The emerging jurisprudence lowers legal risk and licensing costs for AI firms, accelerating model development while still requiring careful fact‑by‑fact analysis.

Key Takeaways

• Courts generally deem unlicensed AI training data as fair use
• Transformative output favors fair use; direct substitutes do not
• Market‑harm analysis includes licensing and substitute‑work theories
• Fact‑specific fair‑use assessment remains essential for AI developers

Pulse Analysis

The wave of copyright lawsuits against AI developers has produced a surprisingly cohesive legal narrative: using large volumes of copyrighted material to train generative models often satisfies the fair‑use defense. Courts apply the classic four‑factor test, giving particular weight to whether the model’s output is transformative rather than a mere replica of the source works. In Bartz v. Anthropic, the Ninth Circuit highlighted that a model that produces novel text, even when trained on entire books, can be deemed transformative, whereas Thomson Reuters v. Ross showed that copying headnotes to build a competing product crosses the line into infringement.

A nuanced part of the analysis involves market impact. The Kadrey v. Meta decision introduced three market‑harm theories, distinguishing between direct substitution, licensing markets, and the emergence of new works that compete with the original. While the first two theories often tilt toward fair use, the third—new works that act as substitutes—could sway future rulings against developers if outputs closely mimic protected content. This layered approach forces AI companies to evaluate not just the quantity of copied material but also the commercial context of the resulting model.

Practically, the evolving case law offers AI firms a clearer risk framework, reducing the urgency to secure costly blanket licenses for every piece of training data. However, the fact‑specific nature of fair‑use determinations means that diligent documentation, transparency about data sources, and rigorous impact assessments remain essential. As courts continue to refine the third market‑harm theory, developers should monitor emerging precedents to stay ahead of potential liability and to shape responsible AI training practices.