Llama Memorizes Large Portions of Harry Potter by Meta

New Study Reveals Llama Model’s Extensive Memorization of Harry Potter and the Sorcerer’s Stone

Recent findings indicate that Meta’s Llama model has memorized Harry Potter and the Sorcerer’s Stone to such a degree that it can reproduce verbatim excerpts from an impressive 42 percent of the book. This revelation comes from a comprehensive study conducted by researchers affiliated with Stanford University, Cornell University, and West Virginia University. The researchers closely examined numerous books from the notorious Books3 dataset, which comprises pirated texts utilized for training Meta’s Llama models. This dataset is pivotal in a copyright infringement lawsuit against Meta, known as Kadrey v. Meta Platforms, Inc. The implications of this study could significantly impact other AI companies facing similar legal challenges.

The research highlights that the Llama 3.1 model demonstrates concerning levels of memorization, particularly with classic literature like Harry Potter and 1984. Specifically, the study reveals that Llama 3.1 can reproduce verbatim excerpts from the first Harry Potter book with a success rate of at least 50 percent, having memorized an astounding 42 percent of its content. Overall, Llama 3.1 can generate excerpts from 91 percent of the book, although the consistency varies. This finding raises critical questions about the boundaries of AI training data and copyright law.

“The degree of verbatim memorization from the Books3 dataset is more extensive than earlier reports indicated,” the research paper states. However, the authors note that the level of memorization is inconsistent across different models and books, and even within various sections of the same book. For instance, the study suggests that Llama 3.1 has only memorized a mere 0.13 percent of Sandman Slim, authored by Richard Kadrey, who is one of the primary plaintiffs in the ongoing copyright lawsuit against Meta. This discrepancy underscores the complexity involved in assessing AI models and their relationship with copyrighted material.

While some of the findings are alarming, it’s crucial to refrain from considering this study as definitive evidence for plaintiffs in AI copyright infringement cases. The nuances of each individual situation warrant careful examination and cannot be oversimplified into a singular narrative.

Timothy B. Lee, a noted journalist, commented on these results in his Understanding AI newsletter, stating, “These findings provide a crucial focal point for all parties engaged in the AI copyright debate.” He suggests that the contrasting results could lead to skepticism regarding the validity of grouping authors like J.K. Rowling, Richard Kadrey, and many others into a single mass lawsuit. This divergence may ultimately benefit Meta, as many authors may not possess the financial means to pursue individual legal actions against the tech giant.

The question of why Llama can replicate certain texts more effectively than others arises. Professor James Grimmelmann of Cornell University posits that the fame of Harry Potter plays a significant role. The book is extensively quoted, and it is highly likely that substantial excerpts have been included in the training data sourced from third-party websites. This phenomenon emphasizes the connection between a book’s popularity and its presence in AI training datasets.

Grimmelmann further asserts that this situation demonstrates that AI companies possess the ability to make decisions that either enhance or mitigate memorization within their models. This aspect is not merely an unavoidable characteristic of AI technology; rather, it is a factor under the control of those developing these systems.

Meta, alongside other AI developers, maintains that utilizing copyrighted materials for training purposes is justified under the doctrine of fair use, a complex legal principle. However, the extent of memorization observed in Llama’s performance could complicate the company’s defense of this argument in court.

Robert Brauneis, a law professor at George Washington University Law School, expressed in correspondence with Mashable that the findings of this study could indeed alter the copyright analysis. He indicated that the likelihood of large language models (LLMs) absorbing more information than previously assumed could ultimately weaken Meta’s position regarding fair use.

We have reached out to Meta for their perspective on the study’s revelations and will provide updates to this article should we receive their response.

Disclosure: Ziff Davis, the parent company of Mashable, filed a lawsuit against OpenAI in April, alleging copyright infringement related to the training and operation of its AI systems.