Judge Pauses Anthropic's $1.5B Author Settlement, Demands Detail on Payouts and Fees

Judge Araceli Martínez-Olguín has put the brakes on what would be the largest copyright class action settlement in U.S. history, declining to grant final approval to Anthropic's $1.5 billion deal with a class of book authors and demanding more detail on how the money will actually reach claimants and how much will go to plaintiffs' attorneys. The pause, set out in a recent order from the Northern District of California, leaves the landmark Bartz v. Anthropic settlement in legal limbo just weeks before payouts were expected to begin flowing to writers whose books were used to train Anthropic's Claude models.

The settlement, first announced in late 2025, was supposed to close one of the most closely watched chapters of the generative AI copyright wars. Authors Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson sued Anthropic in 2024, alleging the company had downloaded and ingested pirated copies of their books from sources like LibGen and Z-Library to train Claude. After the late Judge William Alsup issued a partial summary judgment ruling that Anthropic's training use was transformative but its acquisition of pirated copies was not protected fair use, the parties moved quickly to settle rather than face a damages trial that plaintiffs' counsel projected could exceed $1 trillion in statutory exposure.

What's in the deal

Under the proposed settlement, Anthropic would pay $1.5 billion into a common fund covering an estimated 480,000-plus copyrighted works. Class members would receive a per-work payment estimated at roughly $3,000 per book, subject to verification of ownership and registration status. The deal also requires Anthropic to delete the pirated training datasets it acquired through unauthorized channels, while permitting it to retain models already trained on that data — a compromise that became a flashpoint during the objection period.

Plaintiffs' lawyers had requested attorney fees of around 25 percent of the fund — roughly $375 million — plus expenses, an award that would itself rank among the largest fee recoveries in copyright history.

Why the judge hit pause

According to filings reviewed this week, Judge Martínez-Olguín, who inherited the case after Judge Alsup's retirement, declined to issue final approval at the May fairness hearing. Instead, she ordered class counsel to submit supplemental briefing on several issues that surfaced during the objection period:

• Per-claimant payout mechanics. The judge wants more detail on how the claims administrator will verify ownership of works, handle disputed registrations, and treat works with multiple rights holders such as ghostwritten titles, co-authored books, and works subject to reverted publishing contracts.
• Attorney fees. Martínez-Olguín signaled concern that a 25 percent fee award on a $1.5 billion fund may not be justified under the Ninth Circuit's lodestar cross-check, particularly given how quickly the case settled after the partial summary judgment ruling. She asked counsel to submit detailed time records and a fee calculation showing the multiplier implied by the requested award.
• Notice and opt-out rates. The judge requested updated data on how many class members received direct notice, how many opted out, and how the parties propose to handle late-filed claims.
• Treatment of retained models. Several objectors argued the settlement effectively lets Anthropic keep the fruits of the infringement by allowing continued use of models trained on pirated data. Martínez-Olguín did not rule on that objection but asked the parties to address it more fully on the record.

Objections that gained traction

The objection period drew unusually substantive pushback for a class settlement of this size. A coalition of 28 writers — separately suing Anthropic in a follow-on action over similar conduct — filed a brief arguing that class treatment in this posture allows AI companies to extinguish high-value copyright claims at fire-sale rates. They told the court that statutory damages of up to $150,000 per willfully infringed registered work make the settlement's roughly $3,000 per book figure look like pennies on the dollar for authors with timely registrations.

Other objectors flagged the lack of injunctive relief preventing Anthropic from acquiring training data through similar channels in the future, and the absence of any audit mechanism to verify that the pirated datasets are actually deleted as promised.

What it means for the AI copyright landscape

The pause matters far beyond Anthropic. The Bartz settlement has been treated by the industry as a template — a number that other AI labs and rights holders are quietly using as a benchmark in their own negotiations. If Martínez-Olguín ultimately approves the deal as written, expect a wave of similar class settlements in the OpenAI, Meta, and Microsoft cases. If she requires material changes — higher per-work payments, broader injunctive relief, or stronger deletion provisions — the leverage in every pending AI training case shifts toward plaintiffs.

The timing is particularly awkward for Anthropic, which is simultaneously defending a new class action filed last week by 28 authors over Claude training data, and which has asked a separate court to grant summary judgment in its favor in another copyright dispute. A delay in final approval also delays the release of class members from their claims, meaning Anthropic remains legally exposed on the same conduct it thought it had bought peace on.

For the broader generative AI industry, the order is a reminder that the easy path of settling fast and moving on is not always available when courts take their gatekeeping role seriously. Federal Rule of Civil Procedure 23(e) requires judges to find a class settlement "fair, reasonable, and adequate" — a standard courts have been applying with increasing rigor in technology cases involving novel rights and asymmetric information.

What happens next

Class counsel has been ordered to file supplemental briefing in the coming weeks. Anthropic has indicated it will support the original deal terms. A renewed fairness hearing is expected later this quarter, with no firm date set as of publication. If the judge ultimately denies approval outright, the parties would either need to renegotiate or return to litigation — a prospect that could revive the threat of a damages trial that triggered the settlement in the first place.

For authors waiting on payments, the practical effect is delay. The claims administrator has continued accepting submissions, but no funds will be distributed until the court signs off. For Anthropic, the cost of capital on $1.5 billion in escrow continues to accrue while the legal status of its training pipeline remains unresolved.

The case is Bartz et al. v. Anthropic PBC, No. 3:24-cv-05417 (N.D. Cal.).