Judge William Alsup expressed broad concerns about a proposed $1.5 billion settlement deal between Anthropic PBC and a number of author-plaintiffs, questioning whether the agreement adequately protects class members and Anthropic. At a September 8 preliminary hearing, Judge Alsup denied the parties’ motion for approval without prejudice, later issuing an order postponing approval until more information is provided. The order delays a September 5 settlement agreement between artificial intelligence developer Anthropic and a group of authors who claimed the company illegally used their copyrighted works to train its large language model (LLM).
Background
Authors Andrea Bartz, Charles Graeber, and Kirk Wallace filed a class action lawsuit against Anthropic in the Northern District of California in August 2024, alleging that the company had used their works without permission to train its AI assistant, Claude. The case, Bartz v. Anthropic, 3:24-cv-05417 (N.D. Cal.), joined several other lawsuits targeting AI developers for allegedly infringing on copyright-protected materials during the LLM training process.
In June of 2025, U.S. District Judge William Alsup delivered a ruling that would shape the eventual settlement. Judge Alsup ruled that Anthropic’s training of its AI systems on copyrighted material was protected by the fair use defense, holding that Anthropic’s use was transformative under the first factor. However, Judge Alsup drew a critical distinction regarding works Anthropic downloaded from piracy websites, ruling that their inclusion in the company’s training library was “inherently, irredeemably infringing,” even if those pirated copies were used solely in a transformative manner. Anthropic’s digitization of books it had legally purchased in print was also deemed a fair use, as the digital versions served as direct replacements for the physical copies.
Original Terms and Structure
The original settlement would have required Anthropic to pay a minimum of $1.5 billion, averaging out to approximately $3,000 per infringed work used by the company. As the parties continue to finalized the complete list of infringed works, Anthropic agreed to add $3,000 to the final sum for each work beyond the first 500,000, meaning the final settlement could have potentially exceeded the announced $1.5 billion settlement.
The settlement also required Anthropic to destroy its LibGen and PiLiMi datasets containing pirated works within 30 days of final judgment, thereby removing the infringing materials from its training infrastructure.
The original agreement functioned as a past release settlement, meaning that it only covered infringement by Anthropic that occurred before August 25, 2025. The settlement also did not extend to claims relating to future reproduction, distribution, or creations of derivative works, leaving the door open for copyright owners to sue Anthropic if Claude continues to generate content that infringes on their copyrighted works.
Finally, the release only applied to works appearing on a specific list that the parties have agreed upon, rather than providing blanket protection for Anthropic.
The Court’s Criticisms
In a pre-hearing order, Judge Alsup wrote that he was “disappointed that counsel have left important questions to be answered in the future,” including which works the settlement covered, the class members to be involved, the claim form’s design, and the processes for notification, allocation and dispute resolution. The order expressed skepticism with respect to the parties’ proposed timeline and emphasized that these issues must be addressed before the agreement will receive preliminary approval.
During the subsequent hearing, Judge Alsup cautioned that class members often receive inadequate compensation once monetary settlements are reached. The judge relayed the need for robust notice procedures and an agreement that would shield Anthropic from future litigation over the same copyrighted works, stating that “When [Anthropic] pay[s] that kind of money, they’re going to get the relief in the form of a clean bill of health going forward.”
Judge Alsup also ordered the parties to prepare a claim form requiring copyright owners to opt in. Any work whose owner declines to opt in will be excluded from the settlement entirely.
The parties must submit a finalized list of works, currently numbering around 465,000, by September 15.
Takeaways
Although the settlement approval process has hit a temporary snag, the broad impact of an eventual settlement on copyright holders and generative AI developers is still unfolding. While Judge Alsup’s fair use ruling protects LLM training on copyrighted material under certain circumstances, he found that Anthropic’s downloading of pirated works was still fundamentally infringing. Contrastingly, in Kadrey v. Meta, 3:23-cv-03417 (N.D. Cal.), Judge Vince Chhabria ruled that training LLMs on pirated copyrighted works may still qualify as fair use depending upon the level of transformative use. Given the current uncertainty, more AI developers and media companies may opt for licensing agreements to train LLMs, following the example of the May 2025 deal between The New York Times and Amazon.
Assuming a settlement is approved, the case sidesteps a jury trial on damages that would have tested the limits of statutory damages under copyright law, which can reach up to $150,000 per infringed work in cases of willful infringement. In the most extreme scenario, Anthropic faced potential liability exceeding $1 trillion.
Ultimately, the first major copyright settlement against an AI developer is expected to deliver wins for both sides. AI companies secured another favorable ruling confirming that using copyrighted material to train LLMs qualifies as a transformative use under the fair use doctrine. Meanwhile, copyright holders will likely receive an historic settlement payout along with a holding that a defendant’s transformative use of pirated works does not shield it from infringement claims.
Although AI copyright litigation remains in its early stages, the outcome, whatever its final form, will help strike a temporary balance between technological innovation and the rights of creators. However, this balance is far from settled and likely to be refined, amended, and potentially upended as future cases bring new facts and tougher questions.