Meta AI model can replicate half of Harry Potter text

The research could have big implications for generative AI copyright lawsuits.

Meta AI study reveals model can reproduce half of Harry Potter

Recent research shows that a Meta AI model can reproduce nearly half of the text from Harry Potter and the Sorcerer's Stone. This raises important issues as AI companies face lawsuits for using copyrighted material to train their models. The study conducted by researchers from Stanford, Cornell, and West Virginia University tested how easily five AI models, including those from Meta and Microsoft, could generate text from a widely used collection of books known as Books3. Notably, the findings suggest that the Llama 3.1 70B model, released in July 2024 by Meta, is particularly adept at reproducing Harry Potter excerpts, which could significantly impact ongoing legal battles over copyright infringement.

This study adds a new layer to the ongoing discourse about copyright in the age of AI. With AI models showing such a high ability to reproduce copyrighted text, it challenges the notion of what constitutes fair use. It also amplifies the urgency for a clearer legal framework governing AI training data. This could reshape the legal risks faced by AI developers and lead to a more stringent examination of how models are trained in the future.

Highlights

The findings of this study may set a precedent for future copyright cases involving AI-generated content.