Mistral AI Accused of Using Copyrighted Works to Train Large Language Model
Mediapart reported that Mistral AI reportedly used copyrighted books, songs and other protected works without permission to train its Mistral Large 3 language model. Testing showed the model could reproduce significant portions of works including Harry Potter and Le Petit Prince, suggesting they were included in its training data.
What happened
Mediapart published an investigation on 23 February 2026 claiming that Mistral AI used copyrighted works including Harry Potter, Le Petit Prince, and other protected content to train its models without authorization.
Testing by Mediapart and CNRS researcher Paul Bouchaud found that Mistral Large 3-2512 reproduced 35% of Harry Potter, 58% of Le Petit Prince, and 25% of The Hobbit when prompted paragraph by paragraph. The model also readily produced opening passages from 1984 and A Song of Ice and Fire.
Reported impact
The allegations undermine Mistral AI's stated position as different from American tech giants regarding copyright protection, and suggest the company has engaged in unauthorized data scraping despite prohibitions.
Index analysis
Researchers at universities including Stanford and Yale have validated testing methods that use a model's ability to reproduce protected texts as a strong indicator of their presence in training data. Oxford professor Luc Rocher stated that reproduction of 25-58% of book content indicates the model was trained on the complete work.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.