Study Reveals Meta’s Llama 3.1 Model “Memorized” 42% of Harry Potter: Insights into AI Learning | Technology News

A recent study by researchers from Stanford, Cornell, and West Virginia University reveals that Meta’s Llama 3.1 AI model is more likely to reproduce copyrighted text from the Harry Potter series than competing models. Llama 3.1, which includes 70 billion parameters, has memorized over 42% of “Harry Potter and the Philosopher’s Stone,” allowing it to accurately reproduce excerpts at a greater than 50% probability. The research evaluated five open-weight models, finding significant differences in verbatim text memorization, challenging claims by companies like OpenAI that such memorization is rare. The findings raise concerns about potential copyright infringements as Meta faces lawsuits for utilizing copyrighted materials without permission. The study suggests that fair use determinations must consider not only the training data but also how that data is integrated into AI models. Legal complications may arise, particularly for popular works, complicating efforts by authors to seek redress against AI companies.

Source link