The certification Anthropic signed is the real precedent in this $1.5B case

The lawyers scaled back their fee request. The money is still historic.

Anthropic's $1.5 billion settlement with authors and publishers — agreed in August 2025 and now awaiting final court approval — is the largest reported copyright class action settlement involving AI training data in history. The legal fee dispute that played out this week is a secondary story. The primary story is what the settlement requires, why it matters, and what it signals for every AI company that built its models on ingested text.

What Anthropic agreed to

The settlement covers claims that Anthropic trained Claude and other models on hundreds of thousands of pirated books. Beyond the monetary compensation — more than $3,000 per copyrighted work — Anthropic agreed to two things that go beyond a simple payment:

First, it agreed to destroy pirated datasets. Second, it agreed to certify that those works were not used in its commercial AI models. This is a compliance mechanism dressed as a legal term, and it is the part that will matter beyond this specific case. When a company settles a training data claim, the certification requirement creates a paper trail that plaintiffs' lawyers and regulators can follow. If the certification is false, that's a separate liability.

The fee dispute and what it reveals

The lawyers initially sought $300 million in fees, with $75 million designated for three other firms that were not appointed as class counsel. U.S. District Judge William Alsup — who gave preliminary approval before going on inactive status at the end of December — called this improper in a December 23 memorandum: the appointed counsel could not "appoint someone else" to share responsibility for the class. Anthropic separately objected to the fee structure, according to Reuters.

After that pushback, Susman Godfrey and Lieff Cabraser lowered their bid to $187.5 million, or 12.5% of the settlement fund, with a filing that explicitly disclaimed any obligation to share that amount with the other three firms, Reuters reported. Jay Edelson of Edelson, one of those three firms, said he was "incredibly proud" of his firm's work preparing for trial.

The fee dispute is not incidental. It signals how valuable these cases are perceived to be. Plaintiffs' lawyers who take on AI companies on contingency know that discovery in these cases can be extraordinarily sensitive — training data, dataset compositions, internal communications about data sourcing. The $187.5 million number is a signal about the expected value of similar cases going forward.

The judge and the hearing

U.S. District Judge Araceli Martinez-Olguin is now overseeing the case and scheduled to consider final approval at an April 23 hearing. The Alsup-to-Martinez-Olguin transition is worth watching. Alsup is known for his hands-on approach to complex tech cases — he presided over the Google Street View wi-fi collection case and the Oracle-Google copyright dispute. Martinez-Olguin will be making the final call on whether the settlement is fair to the class, adequate compensation, and whether the compliance terms are enforceable.

What this means for builders and investors

Every foundation model company trained on text faces some version of this exposure. The settlement does not resolve the underlying legal question — whether training on copyrighted text is itself infringement — but it does establish a floor for what resolution looks like when training data is demonstrably pirated. The "destroy and certify" mechanism is likely to appear in future settlements and may become a de facto standard for AI copyright resolution.

The bigger picture: $1.5 billion is a large number, but it is a fraction of what a fully litigated case could have cost if the court had found that training on copyrighted books without a license was per se infringement. The settlement price reflects uncertainty on both sides, not a determination that the training was lawful.