Meta's Llama 3.1 AI model has made headlines for memorizing 42% of 'Harry Potter and the Sorcerer’s Stone', raising serious questions about the AI industry's claims of 'fair use'.

This significant memorization rate marks a stark increase from Llama 1, which only retained 4.4% of the same book, indicating a troubling trend in AI memorization capabilities.

Possible reasons for the high memorization rates include repeated training on the Books3 dataset and the inclusion of text from third-party sources like fan forums and reviews.

The findings come from a recent study published in May 2025 by researchers from Stanford, Cornell, and West Virginia University, who examined multiple open-weight models, including those from Meta.

Stanford law professor Mark Lemley noted that the model's internal structure could legally be considered a copy of parts of the book, complicating the legal landscape for AI outputs.

There is speculation that changes in Meta's training methods may have unintentionally worsened the memorization issue, further complicating the company's legal and operational challenges.

Reports suggest that Meta's CEO Mark Zuckerberg approved the use of pirated content for training, despite internal warnings regarding its legality.

The U.S. Copyright Office has indicated that there is a strong argument that the internal weights of AI models can be viewed as infringing copies of the training data.

These results raise concerns about potential legal liabilities for copyright infringement, as the AI model itself may constitute a copy of copyrighted works.

Meta is facing legal challenges not only in the U.S. but also from French publishers and authors who have filed claims for copyright breaches related to AI model training.

Amid these controversies, Meta is under pressure, having lost key talent and experiencing delays in the development of its ambitious Llama 4 model.

To alleviate financial strains from AI development, Meta is seeking co-funding from competitors like Amazon and Microsoft for its Llama project, referred to as the 'Llama Consortium'.