Client Alert – Summary and Strategic Analysis of Judge Chhabria’s Fair Use Ruling in Kadrey v. Meta

A copy of the alert can be found HERE.

Overview

On June 25, 2025, Judge Vince Chhabria of the Northern District of California issued a significant ruling in Kadrey v. Meta, granting summary judgment in Meta’s favor on copyright infringement claims relating to its use of books (including pirated works) to train the LLaMA family of large language models (LLMs). While the court allowed distribution-based claims to proceed, it held that the act of training the LLM on copyrighted text—without copying or outputting recognizable expression—constituted fair use. Crucially, the court emphasized the plaintiffs’ failure to prove concrete market dilution and deemed it dispositive.

This memo summarizes Judge Chhabria’s opinion, contrasts it with Judge Alsup’s earlier ruling in Bartz v. Anthropic, and offers practical takeaways for companies developing or licensing LLMs.

Key Divergence: How Courts Treated Copying from Unauthorized Sources

Strategic Takeaways for Clients

Source Legality Is Becoming a Split Issue:
– In some courts copying from shadow libraries may be a fair use bar per se.
– In others, courts may allow fair use arguments to proceed despite source illegality if plaintiffs fail to prove market harm.

Future-Proof Your Training Pipelines:
– Avoid any training that relies on data sourced from unauthorized repositories.
– Maintain provenance records and verify licensing paths.

Market Harm Still Central, but Not Always Decisive:
– In Meta, lack of evidence on market dilution was dispositive.
– In Anthropic, the court didn’t reach that issue for pirated works—because it didn’t have to.  It did not consider the market dilution theory advanced in the Meta case.

Note: Judge Chhabria’s suggestion that a use can be “highly transformative” yet still fail due to market dilution may not square with the Supreme Court’s guidance in Warhol v. Goldsmith. Warhol teaches that the concept of “transformativeness” must be evaluated in context, particularly with respect to whether the secondary use serves as a market substitute. Under this framework, substantial market harm may preclude a finding of “transformative” use altogether.

Judge Chhabria’s opinion places considerable weight on the plaintiffs’ failure to produce empirical evidence of market dilution, implicitly requiring them to quantify the causal effect of LLM training on book sales. While courts generally reject speculative damages, they do not demand pinpoint causation—particularly where the relevant data is controlled by the defendant. In the context of LLMs, any such harm is often inherently speculative and difficult to measure, especially when the outputs are not direct substitutes for the original works. The standard applied by the court risks shielding AI developers from liability not because harm is absent, but because it is difficult to trace with precision. Separately, the court appears to assume that books within a genre are largely interchangeable—a view that calls into question whether the damages theory itself is coherent, and one that runs counter to copyright’s foundational principle that each work embodies unique, protectable expression.

Some Further Thoughts for Publishers:

Judge Chhabria’s opinion highlights a concrete licensing breakdown that publishers should not overlook:

“Meta’s head of generative AI discussed spending up to $100 million on licensing. But as negotiations proceeded, Meta realized that licensing would be more difficult than anticipated. For one thing, publishers generally do not hold the subsidiary rights to license books for AI training. These rights are instead held by individual authors, and there is no organization for collective licensing of such rights. Even where publishers do hold AI training licensing rights, they do so regionally rather than globally.”

This observation underscores a missed commercial opportunity: Meta was ready to pay — but rights fragmentation blocked the deal. This should be a wake-up call for publishers seeking to preserve their value proposition in the AI era.

The Meta case confirms there is money on the table — but only for those prepared to claim it.

• Build Discoverability Tools: Make your titles and licensing terms easily searchable by AI companies through platforms or rights registries.
• Negotiate Author-side Consents: Where rights are held by authors, work with them proactively to secure nonexclusive licenses that could enable aggregate negotiation.
• Push for Collective Licensing Mechanisms: Consider banding together with other publishers to create an industry clearinghouse or licensing collective akin to ASCAP/BMI for music or CCC for academic texts.
• Standardize Clauses: Develop and adopt boilerplate AI-training clauses that clearly assign these rights — and preferably on a global basis.
• Audit Your Rights: Assess whether your contracts include rights for data mining, corpus training, or other AI-related uses. If not, consider updates in new agreements.

 Conclusion

The line between permissible AI training and actionable infringement is rapidly evolving. Whether the source of the training data is authorized may, in some courts, trump even a strong fair use defense. Companies should assume that using pirated training data creates legal exposure on multiple fronts—even if the final outputs are transformative and non-substitutive. Conservative compliance strategies will emphasize data provenance, licensing efforts, and risk modeling tied to market substitution evidence.

A copy of the order is available at https://chatgptiseatingtheworld.com/wp-content/uploads/2025/06/Judge-Chhabria-Fair-Use-decision-in-Kadrey-v.-Meta-June-25-2025.pdf?utm_source=substack&utm_medium=email

If you would like assistance in evaluating your organization’s copyright exposure, AI model training practices, or content licensing policies in light of this decision, please contact:

Lawrence R. Robins, Partner, Chair Brand Management Practice Group, at [email protected].