Ex-Meta Researcher Torrented Pirated Books, Deleted It 'Due to Litigation'

The admission that copyrighted books were torrented for AI training creates a direct link between unauthorized data acquisition and the development of major AI models.

A former Meta researcher, Stephen Roller, testified under oath that he torrented the Books3 dataset, which contains nearly 200,000 pirated books, while employed at Meta. This admission came during depositions for lawsuits filed by authors and publishers against Meta for allegedly using copyrighted books to train its Llama AI models. Meta has acknowledged using Books3 but disputes that it obtained the dataset via BitTorrent, suggesting Roller's downloads might have been for personal use. Roller reportedly deleted his copy of the dataset approximately a year ago, citing the ongoing litigation as the reason. The plaintiffs also seek internal Meta documents related to a folder titled "ML Books" and evidence of a Meta book club that may have involved pirated materials. Meta maintains that Roller's testimony does not confirm the dataset was downloaded for the company and claims it has produced substantial server logs, though plaintiffs assert some logs are still missing.

The testimony directly challenges Meta's defense against copyright infringement claims, suggesting the company's AI models were built, at least in part, using illegally obtained data. This significantly strengthens the plaintiffs' case and could have far-reaching implications for how AI companies source training data, potentially leading to stricter regulations and increased legal scrutiny on data acquisition methods. The mention of a "ML Books" folder and a "book club" using "pirated books" indicates potential broader issues within Meta concerning the use of copyrighted material beyond just AI training.

The ongoing legal battles and the admission of torrenting pirated books could set important precedents for the use of copyrighted material in AI development. If Meta is found liable, it may force a shift towards more licensed or ethically sourced datasets, impacting the cost and availability of data for training large language models. Companies might need to invest more in data provenance and copyright clearance, potentially slowing down the pace of AI development or forcing a focus on publicly available or explicitly licensed data.