AI Models Exposed: The Hidden Memorization of Copyrighted Books

Researchers at Stanford and Yale reveal that popular AI models, including OpenAI's GPT, Anthropic's Claude, Google's Gemini, and xAI's Grok, have stored and can reproduce large portions of copyrighted books. This finding directly contradicts the claims made by AI companies that their models do not memorize or store training data.

Key Findings of the Study

The study, published on Tuesday, shows that when prompted strategically, these AI models can reproduce near-complete texts of well-known books such as Harry Potter and the Sorcerer’s Stone, The Great Gatsby, 1984, and Frankenstein. Additionally, they can generate thousands of words from other books like The Hunger Games and The Catcher in the Rye.

Thirteen books were tested, and varying amounts of text from each were reproduced by the AI models. This phenomenon, known as 'memorization,' has been a contentious issue in the AI industry, with companies consistently denying its occurrence on a significant scale.

Industry Denials and Legal Implications

In a 2023 letter to the U.S. Copyright Office, OpenAI stated, 'models do not store copies of the information that they learn from.' Similarly, Google claimed, 'there is no copy of the training data—whether text, images, or other formats—present in the model itself.' Other major players, including Anthropic, Meta, and Microsoft, have made similar assertions.

The new research could have massive legal implications for the AI industry. It may lead to billions of dollars in copyright-infringement judgments and potentially force products off the market. This also challenges the industry's explanation of how AI works, often described in terms of learning and understanding rather than storing and accessing data.

Technical Explanation: Lossy Compression

Many AI developers use the term 'lossy compression' to describe the process more accurately. This concept is gaining traction outside the industry, with a recent German court ruling against OpenAI using this term. The judge compared the AI model to MP3 and JPEG files, which store data in a compressed form, losing some detail but retaining the essence of the original content.

From a technical perspective, lossy compression means that AI models ingest text and images and output approximations of those inputs. This description is less appealing to AI companies, which prefer the metaphor of learning and understanding to promote the idea of AI making novel scientific discoveries and undergoing continuous improvement.

Impact on the Industry

The findings raise serious questions about the foundational claims of the AI industry. The potential for legal action and the need for more transparent practices could reshape how AI companies operate and how they communicate with the public and regulatory bodies.

References

• AI's Memorization Crisis - The Atlantic