Study: Meta AI model can reproduce almost half of Harry Potter book (arstechnica.com)

A recent study conducted by researchers from Stanford, Cornell, and West Virginia University has revealed that Meta's Llama 3.1 70B AI model can reproduce significant portions of the text from "Harry Potter and the Sorcerer’s Stone," raising questions about copyright infringement in AI training. The research examined five popular open-weight models, highlighting that Llama 3.1 is disproportionately adept at generating 50-token excerpts from the novel, suggesting that text reproduction is not a fringe behavior but a notable risk for copyright holders. This finding is significant for the AI and ML community as it underscores the ongoing legal dilemmas surrounding AI models' training on copyrighted material. As lawsuits intensify against AI companies for the use of proprietary content, understanding the capabilities and shortcomings of these models is crucial. The implications of this research could bolster the arguments of plaintiffs in these cases, while also challenging AI firms to improve their practices to ensure compliance with copyright laws, thereby shaping the future landscape of AI development and ethics.