Meta and Mistral AI cofounder accused of copyright infringement for AI training
According to court filings in 'Kadrey v. Meta Platforms Inc.', Guillaume Lample, now cofounder of Mistral AI, reportedly ordered the download of 70 terabytes of copyrighted books from a pirate library while working at Meta in 2022, citing that competitors OpenAI, Google, and DeepMind used similar sources. The Library Genesis data was not ultimately used for LLaMA, but questions remain about training data for Mistral AI's models.
What happened
Guillaume Lample, cofounder of Mistral AI and former researcher at Meta, reportedly ordered the download of 70 terabytes of data from Library Genesis, a pirate library containing copyrighted books and scientific articles, during his time at Meta in 2022.
Internal Meta communications published in the 'Kadrey v. Meta Platforms Inc.' lawsuit show Lample responding to a researcher's objection to using pirated works by stating 'Everybody uses LibGen. That's what OpenAI does with GPT3, that's what Google does with Palm, that's what DeepMind does with Chinchilla. So we're going to do it too.'
Download logs confirmed Lample's role in the operation. Though the Library Genesis data was reportedly not ultimately used for the first version of LLaMA published in February 2023, that model relied on Books3, a dataset containing copyrighted works.
Reported impact
The allegations raise questions about the training data used for Mistral 7B, Mistral AI's first model launched in September 2023, and whether protected copyrighted material was used in its development.
Index analysis
The incident is documented in publicly available court filings from the 'Kadrey v. Meta Platforms Inc.' case, which encompasses over 70 lawsuits against major AI actors filed by creators alleging copyright infringement in model training.
Governance controls
Meta documents indicate that a researcher at the company flagged the use of pirated materials as a 'red line,' suggesting internal awareness of ethical and legal concerns, though the download proceeded.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.