Anthropic Settles with Authors After Allegations of Using Millions of Pirated Books to Train AI
Anthropic has reached a tentative settlement with U.S. authors who accused the company of using millions of pirated books to train its Claude AI, averting a December 2025 trial and potentially influencing how firms gather and license training data.
Key takeaways
- Class action: U.S. authors led by Andrea Bartz, Charles Graeber and Kirk Wallace Johnson alleged Anthropic used up to seven million pirated books to train Claude (class-action case).
- Legal stakes: Plaintiffs sought billions in damages; statutory damages for willful infringement can be as high as $150,000 per work (statutory-damages context).
- Court findings: Judge William Alsup said training on books can sometimes be “transformative” fair use — but use of pirated copies would be copyright violations (ruling overview, court findings).
- Settlement unknowns: Terms and payment breakdowns remain sealed until preliminary approval filings due Sept. 5, 2025 (publishersweekly report).
Main content
Background and settlement
San Francisco, California (Times Media Service) — Anthropic, an AI company backed by Amazon and Google, has reached a landmark settlement with U.S. authors who alleged the firm used millions of pirated books to train its Claude model. The tentative deal avoids a high-profile December 2025 trial and could shape how AI companies license and use books and other copyrighted works as training data (CreativeS Unite, Fortune).
Plaintiffs and allegations
A group of U.S. authors led by Andrea Bartz, Charles Graeber and Kirk Wallace Johnson filed a class-action lawsuit claiming Anthropic downloaded as many as seven million books from pirate websites to train Claude, later purchasing some titles but, they say, only after initial unauthorized copying had already taken place (CreativeS Unite).
Claims and potential damages
The complaint alleged mass copyright infringement and sought billions in damages. Legal analysts note that statutory damages for willful infringement can reach up to $150,000 per work, which underscores why plaintiffs pursued large-scale relief and why defendants face high exposure (Fortune).
Court rulings that set the stage
Before the settlement, U.S. District Judge William Alsup issued nuanced rulings in June 2025. He wrote that training an AI on books can, in some circumstances, be “transformative” — akin to a student reading to learn — supporting a potential fair-use defense. Crucially, Judge Alsup said the lawfulness of how copies were obtained matters: using pirated copies could constitute copyright violations (No Film School, Insurance Journal).
Settlement details: known and unknown
The parties say they reached a tentative agreement before the scheduled December 2025 trial. Specific terms — including the total payment amount, distribution plan among class members, and any non-monetary provisions — remain sealed for now. The court ordered both sides to file requests for preliminary approval by Sept. 5, 2025, which could reveal how funds would be allocated and whether the deal places limits on future litigation (CreativeS Unite, Insurance Journal, Publishers Weekly).
Allegations explained
Plaintiffs allege Anthropic assembled a massive dataset from pirate sites and used that content to train Claude to write, summarize and answer questions. The complaint acknowledges that Anthropic later purchased some of the titles, but the plaintiffs argue the initial copying from unauthorized sources was a legal violation and central to the claims.
Technical and legal context: “transformative” use and training data
Judge Alsup’s rulings raised the concept of “transformative” use — that an AI’s training purpose might change the original work enough to support fair use. However, the court emphasized that transformation does not override the legality of acquisition. As a result, AI firms face pressure to document lawful data acquisition and to be cautious about asserting broad fair-use defenses when training on copyrighted works (No Film School, Fortune).
Why this matters to the AI industry
While a settlement is not binding precedent, legal experts say the deal sends a strong market signal: litigation over training data can be costly, and firms may opt to license materials or otherwise harden data-governance practices to avoid trial risk. The agreement could affect active or pending suits against other companies, including OpenAI, Microsoft and Meta, and could influence norms around licensing, compliance, and model provenance (Fortune).
Responses from authors and publishers
The Authors Guild praised the tentative agreement as a warning to companies that use pirated or unlicensed material to develop AI models. Publishers’ trade groups said they are watching the case closely, viewing the outcome as potentially protective of writers’ and publishers’ rights. Anthropic has not publicly released the full settlement text (Publishers Weekly).
Implications for San Diego, California
- Local tech and startups: San Diego’s AI and biotech startups may adopt stricter licensing practices, raising costs and potentially altering project plans (Fortune).
- Universities and research labs: Institutions may revise policies on training data, with tech-transfer and legal offices offering tighter guidance on licensed versus scraped datasets (No Film School).
- Book community and creators: Authors, indie publishers and librarians may see the settlement as strengthening creators’ rights in the AI era (Publishers Weekly).
- Legal and business services: Law firms and consultants in San Diego may see higher demand for IP and compliance work as companies seek compliant data pipelines (Insurance Journal).
- Practical effects for consumers: Some AI features could become costlier or slower to appear, though training on licensed data may improve output reliability (Fortune).
Industry and legal experts to watch
Attorneys and AI policy analysts are expected to scrutinize the Sept. 5 preliminary-approval filings. Observers will look for how the settlement compensates authors, whether it imposes limits on future claims, and whether it contains clauses that might encourage similar settlements across other cases in the AI ecosystem (Fortune, Insurance Journal).
Next steps and outlook
Both parties must submit requests for preliminary approval by Sept. 5, 2025. Those filings may disclose payment amounts, distribution plans, and any future-release or release-of-claims terms. The December 2025 trial remains on the calendar until the court grants final approval; observers expect more documents and commentary to surface in early September (CreativeS Unite, Publishers Weekly, Insurance Journal).
Quote
“The settlement may not be precedent, but it is a powerful market signal that companies should treat training data acquisition and licensing seriously,” legal analysts say.
