# Judge Approves Anthropic's $1.5 Billion Settlement Over Pirated Books Used to Train Claude

The payout works out to roughly $3,000 per book across about 500,000 titles, and it follows a ruling that training on lawfully acquired books is fair use.

- Published: 2026-07-23T06:03:52.480Z
- Canonical: https://polylog.news/ai/2026-07-23/judge-approves-anthropic-s-1-5-billion-settlement-over-pirat
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [Polylog editors](https://polylog.news), [TechCrunch](https://techcrunch.com/2026/07/20/anthropics-landmark-1-5b-copyright-settlement-is-approved/), [Marketplace](https://www.marketplace.org/story/2026/07/21/anthropic-pays-15-billion-to-settle-a-copyright-case)

A United States federal judge has given final approval to a settlement in which Anthropic will pay roughly $1.5 billion to authors and publishers whose books were downloaded from pirate repositories and stored in the company's internal training corpus. District Judge Araceli Martínez-Olguín called the class-action deal "meaningful relief," according to [TechCrunch](https://techcrunch.com/2026/07/20/anthropics-landmark-1-5b-copyright-settlement-is-approved/) and [Marketplace](https://www.marketplace.org/story/2026/07/21/anthropic-pays-15-billion-to-settle-a-copyright-case). The agreement covers more than 482,000 works at about $3,000 each, and about 91% of eligible titles have already been claimed. A Russian-language summary of the decision circulated the same day on [AI ML Big Data](https://t.me/ai_machinelearning_big_data/10570).

The structure of the payout matters as much as the total. Anthropic already paid $300 million into the fund after preliminary approval, with a further $300 million due shortly and two later tranches of $450 million on the first and second anniversaries. Plaintiffs' counsel described it as the largest known copyright recovery, a characterization the court did not dispute.

The legal logic is narrower than the number suggests. The court had earlier held that training models on books a company lawfully obtained is fair use. Anthropic's exposure came from a separate act: retaining more than seven million pirated copies in a central library that was not necessarily used for training at all. In other words, the liability attached to acquisition and storage, not to the act of learning from text.

## What this means

The settlement prices a specific liability that every frontier lab shares: the provenance of pretraining data, not the legality of training itself. Labs that sourced corpora from pirate repositories now face a quantified downside of roughly $3,000 per infringing work, which raises the value of licensed data deals and clean-room sourcing and lifts the effective cost of a pretraining run. OpenAI, Google, Meta, and any lab that scraped books or code without provenance records are the exposed parties, and the channel is litigation cost plus the need to re-license or re-collect data.

## What to watch

- Whether the next wave of plaintiffs (music publishers, news organizations, code owners) cites this dollar figure as a benchmark, which would signal that per-work damages are becoming a standard settlement currency.
- Whether labs begin disclosing data provenance in model cards or licensing terms, a sign that clean sourcing is turning into a competitive and legal requirement rather than an afterthought.