Unsealed Briefs Reveal OpenAI and Microsoft Knowingly Used Pirated Books

Unsealed court documents reveal that OpenAI and Microsoft executives knowingly used pirated book datasets to train AI models despite recognizing the legal and economic risks to authors. Internal messages show the companies anticipated that their technology would replace human writers and cause significant job losses in the creative industry. The filings also detail 'Project Clear,' a 2022 effort by OpenAI to delete evidence of their reliance on illegal sources like LibGen to manage public 'optics.'

Key Points

• Internal communications show OpenAI executives knew they were using pirated books from 'sketchy Russian websites' like LibGen to train AI models.
• OpenAI leadership admitted that their technology would likely make human authors unemployed and planned to ignore artist concerns to release products anyway.
• Microsoft was allegedly aware of the use of pirated datasets as early as April 2019 during a presentation to Bill Gates and CTO Kevin Scott.
• OpenAI initiated 'Project Clear' in 2022 to delete evidence of LibGen usage from their systems to avoid negative public perception.
• The lawsuit alleges that OpenAI viewed the destruction of the writing profession as 'acceptable economic disruption' in pursuit of their research mission.

Sentiment

The overall sentiment is critical of the companies and mostly sympathetic to the article's core accusation that the use of pirated books was knowing and ethically troubling. Hacker News does not show full consensus, because many commenters push back on copyright maximalism and question whether training should require permission or payment. Still, the dominant mood is distrustful of OpenAI and Microsoft, with the legal debate functioning more as a dispute over remedy and theory than as a defense of the companies' conduct.

In Agreement

• OpenAI personnel appeared to understand that LibGen and similar sources were legally and reputationally risky, making the conduct look deliberate rather than accidental.
• A commercial AI system trained on shadow-library books is not equivalent to an individual reading books, because the system scales into a market product that may substitute for authors' work.
• Companies with OpenAI and Microsoft's resources should have licensed books or negotiated compensation instead of relying on pirated corpora.
• Efforts to remove or obscure mentions of pirated datasets suggest awareness of wrongdoing and make the corporate conduct look worse.
• The filings reinforce broader concerns that AI companies externalize costs onto writers and the cultural commons while privatizing the gains.

Opposed

• Model training is closer to learning from reading than to republishing, so requiring payment to every author whose work informs a model is conceptually unworkable.
• The Authors Guild is presenting an advocacy narrative, and internal worries about optics do not alone prove legal liability.
• Copyright law may be a poor fit for large-scale statistical learning, especially when models do not simply distribute copies of the books.
• Publishers and rights holders helped create piracy incentives by making legal access fragmented, expensive, or unavailable.
• Some commenters argue that information copying is an unavoidable feature of the digital era and that attempts to fully police training data may be impractical or harmful.