Meta’s AI Fair Use Win, but Anthropic Shows the Limits

Two judges in the Northern District of California have issued the first detailed decisions on whether copying books to train a large language model constitutes fair use. Judge Vince Chhabria’s order in Kadrey v. Meta Platforms, Inc., No. 3:23-cv-03417-VC (N.D. Cal. June 25 2025) granted Meta summary judgment on the plaintiffs’ test claim, while Judge William Alsup’s opinion in Bartz v. Anthropic PBC, No. 3:24-cv-05417-WHA (N.D. Cal. June 23 2025) found fair use for lawfully obtained books but sent claims based on pirated copies to a jury. Together, the opinions provide clearer guardrails for sourcing training data.

What convinced Judge Chhabria

The Kadrey plaintiffs focused on a single question: Could the initial ingestion of their books ever be considered fair use? Judge Chhabria said yes. He emphasized the absence of evidence that a licensing market for AI training already exists, found the model’s use to be highly transformative, and noted no record of market harm. Because the plaintiffs chose not to develop the rest of their claims in discovery, Meta prevailed for now.

Why Judge Alsup drew a line

The record in Bartz was wider. Anthropic admitted it had created a central library containing both purchased e-books and files downloaded from pirate sites. Judge Alsup held that using legitimately acquired books was “exceedingly transformative” and covered by fair use, but that retaining and copying pirated works was not. He set a damages trial for December 2025 to determine the cost of that misstep.

Comparing the two opinions

Although both judges view large-scale ingestion of text as transformative, the cases diverge on evidence and provenance. In Kadrey the authors offered little discovery and could not show any realistic licensing market; in Bartz expert testimony and internal emails painted a fuller picture and revealed blatant use of pirate libraries. Judge Chhabria emphasized the plaintiffs’ evidentiary burden and left policy questions to Congress. In contrast, Judge Alsup, drawing on his extensive software docket, likened machine learning to human reading billions of words, yet made clear that willful infringement crosses the line. The practical message is that lawful sourcing matters as much as technical transformation.

Looking ahead

Neither ruling decides whether AI outputs that mimic an author’s style infringe that author’s rights, and both judges suggested that a stronger evidentiary record might shift the fair use balance. For now, the opinions signal that companies ingesting only lawfully obtained material and documenting a genuine transformation process have a credible fair use defense, while those relying on pirate sites face significant risk. Plaintiffs are likely to test these boundaries again as licensing markets for training data mature.