OpenAI and Microsoft Executives Knew Book Piracy Was Illegal, Unsealed Court Filings Reveal

The discovery phase of a federal lawsuit is where the polished corporate narrative goes to die, replaced by the blunt reality of internal Slack messages. Newly unsealed filings in the copyright battle against OpenAI and Microsoft have just dragged those private conversations into the public record, revealing exactly what executives were saying behind closed doors. Documents released in Manhattan federal court show that leadership at both companies internally acknowledged their artificial intelligence models were trained on pirated books, fully aware that the technology would directly threaten the livelihoods of human authors.[1][4]

The disclosures stem from consolidated copyright litigation brought by the Authors Guild, alongside prominent writers including John Grisham and George R.R. Martin, as well as news publishers like The New York Times. The plaintiffs are seeking summary judgment, arguing that the tech giants committed willful copyright infringement by ingesting millions of protected works to build systems like ChatGPT.[1][5]

Central to the authors' case is the revelation that early versions of OpenAI's language models were trained on data from Library Genesis, or LibGen—a widely known shadow library that hosts pirated books. It turns out the engineers building the future of text generation knew exactly where their material was coming from. According to the unsealed briefs, OpenAI engineers Tom Brown and Ben Mann internally described the data source as "sketchy AF" back in 2019.[4][5]

The internal communications show a deliberate effort to obscure the origins of this training data. When Dario Amodei, then a senior researcher at OpenAI and now the chief executive of Anthropic, asked if it was problematic to name the training corpora "Books1" and "Books2" without disclosing their source, Mann replied that the description was "deliberately vague since it's LibGen."[5]

The filings also highlight executives' awareness of the economic impact their products would have on the publishing industry. In May 2020, OpenAI policy director Jack Clark cautioned that improved language models would increasingly displace genre fiction writers, predicting the company would proceed despite pushback from artists.[4]

Two years later, the internal ambition had only grown. OpenAI research lead Tarun Gogineni outlined goals for autocompleting stalled literary series—a wry nod to famously delayed fantasy epics. When discussing authors' complaints that their work was being stolen to create AI-generated competition, Gogineni reportedly didn't mince words, dismissing the concerns as "acceptable economic disruption."[4]

Microsoft, which heavily invested in OpenAI, was also aware of the practices. The documents indicate that Microsoft executives, including former Chief Technology Officer Kevin Scott and co-founder Bill Gates, were informed about the use of LibGen as early as April 2019 during a walkthrough of an early GPT-3 model.[4]

Microsoft's own leadership expressed alarm at the sheer scale of the data ingestion. Brent Hecht, Microsoft's director of applied science, didn't hold back in his internal assessments. He described the mass scraping of copyrighted material as an "astonishing theft of unprecedented proportions" and suggested it could be the "largest theft of labor in human history."[1]

OpenAI eventually initiated a cleanup effort in the summer of 2022. Then-vice president of research Bob McGrew directed the deletion of the LibGen-derived datasets, noting that removing the material from their systems would be "very valuable for legal reasons" as the company's public profile grew.[5]

The tech companies maintain that their training methods are protected under the legal doctrine of fair use, arguing that the models analyze information to generate transformative new responses rather than simply reproducing the original texts. They also note that the employees involved in creating the LibGen datasets are no longer with the company and that current models do not rely on that specific data.[1][5]

The unsealed documents do not resolve the underlying legal question of whether training artificial intelligence on copyrighted material constitutes infringement. However, the internal admissions provide the plaintiffs with direct evidence of corporate intent, shifting the weight of the upcoming trial as both sides await a ruling on their competing motions for summary judgment.[1][2]