Anthropic allegedly acquired millions of pirated books to train its Claude AI models, then destroyed the copies after use. That sequence of events, now surfacing in ongoing litigation, has copyright lawyers asking a pointed question: does destroying the evidence of infringement help Anthropic legally, or make things considerably worse?
What the Allegations Say
Court filings and reporting from Yahoo News detail claims that Anthropic sourced a large corpus of copyrighted books through pirated channels, used them to train Claude, and subsequently deleted the files. Plaintiffs argue the destruction of those copies does not erase the underlying act of reproduction, which is the step that triggers copyright liability under U.S. law. The number of books involved reportedly runs into the millions, spanning fiction, nonfiction, and academic texts.
Key Facts
- Millions of books were allegedly obtained from pirated sources for Claude's training data.
- Anthropic reportedly destroyed the copies after the training process concluded.
- Copyright law treats the act of reproduction, not ongoing possession, as the point of potential infringement.
- The case is part of a broader wave of AI training data lawsuits targeting major model developers.
- Destruction of files could raise additional questions around spoliation of evidence.
The legal theory behind the plaintiffs' case centers on the reproduction right. When a company copies a book into a training pipeline, that copy exists, however briefly. Whether the file is later deleted has little bearing on whether that copying was authorized. Fair use arguments are available to Anthropic, and the company is expected to lean on them heavily, citing transformative use and the commercial value added by training a large language model. Courts have not yet settled how fair use applies to AI training at scale, making this case one that legal observers are watching closely.
"Destroying the files after training does not unring the bell. The reproduction already occurred, and that is the moment copyright law cares about."Copyright attorney quoted in Yahoo News coverage
Spoliation Risk Adds Another Layer
Beyond the infringement question, legal analysts point out that destroying files once litigation is foreseeable can constitute spoliation of evidence. If Anthropic knew or should have known that lawsuits over training data were coming, eliminating those files could invite sanctions or adverse inference instructions from a judge. The company has not publicly addressed the destruction timeline in detail, and it is unclear when the files were deleted relative to when the first complaints were filed.
This case fits a pattern that has emerged across the AI industry. Nearly every major model developer is facing some form of training data litigation, and the outcomes will shape how Claude's model family and comparable systems are built in the future. Anthropic has previously found itself navigating other sensitive data questions, including the incident where Claude conversation logs surfaced in Google Search results, a situation that required a public response about data handling practices.
The broader regulatory environment is also shifting. AI companies including Anthropic have been called to participate in policy discussions at the highest levels of government, with executives appearing at forums like the G7 summit in France to address questions about responsible AI development. Training data provenance is increasingly part of those conversations.
For now, the case moves through discovery. The court's eventual ruling on fair use, and any findings related to the destruction of the training files, could set precedent that reaches well beyond Anthropic. Publishers and authors who have been pressing for licensing frameworks will be watching closely to see whether litigation or legislation moves faster in establishing rules for how AI companies may use copyrighted material.
