OpenAI Knew Pirated Books Risked Authors' Livelihoods

The filings trace a pattern that starts well before the current generative AI boom. In May 2020, Jack Clark, then OpenAI's policy director and later an Anthropic co-founder, wrote internally that the company's work would lead to systems replacing people's labor and that OpenAI would likely ignore artists' concerns and ship anyway. By 2019, according to the documents, OpenAI had already shown Microsoft an early GPT-3 and disclosed that it had used Library Genesis (LibGen), described as one of the world's largest pirate libraries, for training.

The internal debate that followed appears to have focused less on legality than on optics. Sam McCandlish said his worry was how it would look — specifically an article on Hacker News accusing OpenAI of using copyrighted data from a questionable Russian site. Dario Amodei was said to have had some doubts about LibGen as a training set. By summer 2022, OpenAI ran Project Clear, deleting LibGen files; a June 15, 2022 Slack exchange shows research VP Bob McGrew deciding to purge LibGen from systems and storage because OpenAI was in the news so much.

The Authors Guild's CEO, Mary Rasenberger, argues the documents show OpenAI and Microsoft held writers and their work in astonishing disregard, choosing to repeatedly steal books rather than pay for them. The case's outcome may hinge on whether a court reads these messages as deliberate infringement or as the ordinary internal debate of a fast-moving startup, and on how much weight it gives to the executives' own stated expectations about the harm to authors.