Ex-Meta Researcher Torrented Pirated Books, Deleted It 'Due to Litigation' (torrentfreak.com)

Meta is facing multiple lawsuits over the alleged use of a pirated dataset, known as Books3, to train its Llama AI models. Key figures including authors Sarah Silverman and Ta-Nehisi Coates are involved in the cases, which have been consolidated under Judge Vince Chhabria in California. Recent testimony from a former Meta researcher, Stephen Roller, revealed that he torrented the Books3 dataset while employed at Meta, contradicting the company's previous claims that it did not obtain the dataset through illicit means. This development raises significant questions regarding Meta's data acquisition practices, particularly given the court's 2022 decision that allowed AI training on books under the fair use doctrine, leaving the legality of the torrenting as the primary issue. The implications of these lawsuits extend beyond Meta, highlighting potential ethical and legal concerns within the AI/ML community regarding data sourcing. The fact that employees may have accessed pirated materials not only questions the integrity of training models like Llama 1 and Llama 2 but also suggests broader cultural issues within tech companies concerning the use of shadow libraries. As the situation unfolds, with additional publishers and authors joining the plaintiffs, it could redefine the boundaries of fair use in machine learning and set precedents for how AI models are trained in the future, particularly concerning copyright and intellectual property rights.