The recently proposed $1.5 billion settlement between Anthropic—one of the most prominent artificial intelligence (AI) developers—and a class of authors and publishers was poised to be the largest copyright settlement in U.S. history. But Monday, the federal judge overseeing the case rejected the deal, leaving unresolved questions about liability, damages and the future of AI copyright law.
Although the size of the proposed payout drew headlines, the factual and legal backdrop of the dispute remains just as remarkable. Unlike most AI copyright cases, which hinge on whether training on lawfully acquired books qualifies as “fair use,” this case turned on Anthropic’s admitted reliance on pirated digital libraries. That unusual fact pattern simplified the infringement claims and magnified the company’s potential exposure.
Factual Background: From Innovation to Litigation
Anthropic, founded in 2021 by former researchers from another AI company, positioned itself as a leader in safety-aligned large language models (LLMs). Its flagship model, Claude, quickly gained market traction, supported by billions in investment from established tech giants.
In building Claude’s training corpus, Anthropic engineers allegedly assembled a “research library” of books, many not obtained through licenses or purchases. Instead, they were downloaded from piracy networks like Library Genesis (LibGen) and Pirate Library Mirror (PiLiMi). This revelation had significant consequences. Courts continue to wrestle with whether training on lawfully purchased works is “transformative” fair use, but reliance on stolen content left Anthropic with far fewer defenses.
Judge Alsup’s Mixed Ruling
In June 2025, Judge William Alsup of the Northern District of California issued a landmark ruling:
- Fair Use Applies to Purchased Books: Training on books Anthropic lawfully acquired was deemed transformative under the Supreme Court’s decision in Campbell v. Acuff-Rose, 510 U.S. 569 (1994) and not a substitute for the market of the original works.
- No Fair Use for Pirated Books: Using pirated ebooks was infringement, plain and simple. Fair use could not apply to works Anthropic never lawfully obtained.
This mixed decision simultaneously strengthened the broader fair use defense for AI developers and carved out clear liability for piracy.
The Proposed Settlement: Building the $1.5 Billion Figure
Before Monday’s rejection, the parties had presented a settlement structured on a per-work basis, with damages set at $3,000 per infringing title across an estimated 500,000 pirated works. The calculation was straightforward:
500,000 × $3,000 = $1.5 billion
If more pirated titles were identified, Anthropic would have been required to pay an additional $3,000 per book.
Why $3,000? The exact thinking behind this number has not been revealed, but it does fall within the statutory range under 17 U.S.C. Section 504(c) (from $750 to $30,000 per work infringed and up to $150,000 for willful infringement). Second, though, is the “common-ness” of the figure: $3,000 is a commonly awarded figure—above the minimum but well below catastrophic levels. It is also likely that the parties worked to determine a number that balanced meaningful recovery for authors against the risk of bankrupting Anthropic.
The $1.5 billion fund was to cover author claims, administrative costs, attorney’s fees capped at 25% (~$375 million) and service awards to class representatives. Net payouts to authors would have been reduced accordingly.
Why the Judge Rejected the Deal
Monday, Judge Alsup declined to approve the settlement. Although court observers are still analyzing his detailed reasoning, several themes are apparent on a first read. First, the court has concerns over the fairness of the settlement. Courts must ensure class settlements are fair, reasonable and adequate. The court questioned whether $3,000 per work appropriately valued the claims given statutory maximums orders of magnitude higher. Second, Judge Alsup raised his sensitivity over precedent setting—the concern that such a formula risked setting a de facto benchmark for future AI copyright disputes—something the court may have deemed premature. Third, with AI copyright doctrine still unsettled, the court appeared reluctant to cement a global compromise that could shape the trajectory of ongoing cases against others.
Unusual Facts of This Case
Even without an approved settlement, this case stands apart from others. Most AI copyright suits hinge on unsettled fair use arguments; this case involved straightforward piracy. Furthermore, with 500,000 works at issue, statutory damages at $30,000 per work could have exceeded $15 billion, and at $150,000 per work would have been upwards of $75 billion. The impact of a large-scale award could have significant impacts not only on Anthropic, but the entire AI industry. Moreover, the settlement (had it been approved) would have required Anthropic to destroy pirated files, underscoring that this was a cleanup of past conduct, not a forward-looking license.
Implications for Future AI Copyright Cases
The collapse of the Anthropic settlement keeps uncertainty alive. Developers cannot rely on a simple damages formula, and plaintiffs may feel emboldened to demand higher payouts given the statutory ranges at stake. The case still clarifies one critical principle: fair use arguments remain viable for lawfully purchased works, but piracy is indefensible.
What remains unclear is how courts will value claims tied to lawful acquisitions—a vastly more common issue than outright piracy. Without the $3,000 per work figure locked in, there is still no benchmark for calculating “royalties” for AI training on legitimate sources. The lawyers involved are, no doubt, feverishly working to resolve that uncertainty.
Main Takeaways
The rejection of the Anthropic settlement is itself a watershed moment in AI copyright law. It underscores that while courts may tolerate training on lawfully acquired material, they will not excuse reliance on pirated content—and they are not yet ready to endorse sweeping financial frameworks for resolving AI copyright claims.
Instead of providing closure, the decision prolongs uncertainty. AI companies still face existential risks from aggregated statutory damages, and authors remain without clarity on compensation benchmarks. For developers, the lesson is unchanged but sharper: what matters is not just what your model learns, but how you sourced the data—and until courts provide firmer guidance, the compliance path will remain precarious.
