Meta Wins Legal Battle in Generative AI War: Fair Use Analysis in Kadrey v. Meta Platforms, Inc.

DEREK FAHEY, ESQ.

In its June 2025 ruling in Kadrey v. Meta Platforms, Inc., the U.S. District Court for the Northern District of California considered whether Meta’s use of copyrighted books to train its Llama large-language models constituted fair use. The court granted summary judgment for Meta, but the result depended heavily on weaknesses in the plaintiffs’ evidentiary presentation rather than a blanket approval of AI training practices.

The Four Fair-Use Factors

1. The purpose and character of the use
2. The nature of the copyrighted work
3. The amount and substantiality used
4. The effect on the potential market for the original work

Factor One: Purpose and Character

The court found Meta’s use highly transformative. The books were created to be read, while Meta used them to train a system capable of responding to prompts and performing computational tasks. Although Meta’s use was commercial, the transformative purpose weighed strongly in its favor.

The court did not ignore Meta’s acquisition of works from “shadow libraries,” but treated that issue as secondary without evidence tying the conduct to broader infringing activity.

Factor Two: Nature of The Works

The plaintiffs’ novels, plays, memoirs, and nonfiction works were highly expressive and entitled to strong copyright protection. This factor favored the authors, although the court noted that it generally carries less weight when the works have already been published.

Factor Three: Amount Used

Meta copied the works in full. The court nevertheless found the amount reasonable in relation to the transformative purpose because full-text ingestion supported the training process. Meta’s post-training safeguards and the plaintiffs’ inability to show substantial regurgitation also influenced the analysis.

Factor Four: Market Effect

The court treated market effect as the most important factor. The plaintiffs did not establish that Llama reproduced substantial portions of their works, and the court rejected a claimed licensing market for AI training as circular.

The plaintiffs’ strongest potential theory was that generative AI could flood the market with competing works and reduce demand for human-authored content. The court acknowledged that this theory could be viable but found the evidence too generalized and insufficiently tied to the plaintiffs’ specific markets.

Final Takeaways

The ruling does not provide AI developers with blanket protection. Future copyright plaintiffs will need detailed, work-specific evidence of substitution and market harm. AI developers should continue to consider licensing, technical safeguards, and transparency regarding training sources.