AI Regulation Tracker  /  Litigation

Anthropic agrees to a reported $1.5 billion settlement over pirated books used to train Claude

A federal court held that training AI on lawfully acquired books was fair use, but that storing pirated copies was not. The resulting settlement puts a public dollar figure on training-data infringement that vendors and their customers now have to price.

By Anthony Guerriero  ·  Reviewed by The Leveraged Years Editorial Desk  ·  Published July 7, 2026  ·  Last updated July 7, 2026

Anthropic has agreed to resolve a class action brought by book authors for an amount reported at $1.5 billion, reported as one of the largest copyright recoveries on record. The case, Bartz v. Anthropic PBC, No. 3:24-cv-05417, was filed in the Northern District of California by authors Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson. They alleged that Anthropic downloaded millions of copyrighted books from pirate libraries, including LibGen and the Pirate Library Mirror, to train its Claude models. The number matters less for what it says about one company than for what it hands to every lawyer who reviews an AI contract: a public price for training on infringing data.

The ruling that split training from storage

In June 2025, Judge William Alsup issued a summary judgment order that separated two questions courts had been treating as one. Training a model on books Anthropic had lawfully acquired, he held, was fair use, describing the use as transformative. But acquiring and retaining a permanent library of pirated copies was a different act, and that act was not fair use. The distinction is the load-bearing point for professionals. The court did not bless AI training in general, and it did not condemn it in general. It drew the line at how the underlying data was obtained and kept.

What the settlement covers

Following the ruling and the certification of a class, the parties reached a settlement. Public summaries put the fund at a reported $1.5 billion covering roughly 500,000 titles, which several commentators have translated to a reported figure of about $3,113 per work before fees and costs. Those per-work numbers are computed from the reported totals, so treat them as estimates rather than confirmed court-ordered amounts. The settlement also requires Anthropic to destroy the pirated datasets it downloaded, according to the settlement materials. The court granted preliminary approval in September 2025, and final approval remains pending in 2026.

Why this reaches beyond Anthropic

For in-house counsel, the value is the benchmark, not the defendant. Until now, the cost of training-data infringement was speculative. A concrete per-work figure, even a reported one, lets a legal team model worst-case exposure across a corpus of allegedly infringing works and compare that number against the indemnity a vendor is offering. Many AI vendor agreements cap IP indemnification at a fraction of the deal value or the fees paid. A cap measured in six figures looks different next to potential liability measured against hundreds of thousands of works. The benchmark also reframes the negotiation. A vendor that once treated an uncapped or high indemnity as an unreasonable ask now has to weigh that request against a settlement its own industry produced. For the customer, the ask is no longer abstract, and the counterparty knows it.

The defenses that got weaker

Two familiar contract and litigation positions took damage here. The first is "the data was publicly available." Availability on a pirate site is not lawful acquisition, and the court's storage holding turned on how the copies were obtained. The second is "we did not know." The ruling and settlement together signal that a party building or deploying models cannot assume good faith about data provenance will insulate it. For a company buying AI services, that means the provenance of a vendor's training data is now a due diligence item, not an assumption.

What this ruling does not do

It does not hold that AI training is inherently infringing. The fair use finding for lawfully acquired books cuts the other way, and it remains a single district court order that other courts may weigh differently. It does not create a statutory per-work damages rule. And a settlement is a negotiated resolution, not a merits judgment on damages, so the reported figures are not a court's valuation of each work. The durable takeaway is narrower and more useful: data sourcing is where copyright exposure concentrates, and contracts should allocate that risk explicitly.

For counsel and procurement, the practical response is to open the AI agreements already in force, read the indemnity language against this benchmark, and ask whether the vendor warrants the provenance of its training data at all.

Browse the full AI Regulation News tracker

Informational analysis for working professionals, not legal advice. Confirm how any rule applies to your situation with qualified counsel.