Why Meta Could Encounter the Same ‘AI Copying’ Challenges as OpenAI and Microsoft

Meta’s latest AI model, Llama 3.1, has alarmingly high rates of text replication from copyrighted works, particularly the first Harry Potter book, which the AI reportedly memorized about 42%. A study by researchers from Stanford, Cornell, and West Virginia University found that Llama 3.1 reproduces 50-word passages quite frequently, far more than earlier models, like Llama 1, which memorized only 4.4% of the same text. This raises significant concerns regarding copyright infringement, as authors and publishers are increasingly resisting unauthorized use of their works. The training methods may involve repeated exposure to specific texts or inadvertent inclusion of excerpts from various sources, leading to learning behaviors that favor memorization over generalization. As legal scrutiny increases, Meta, along with companies like OpenAI—facing a lawsuit from The New York Times—might confront serious challenges relating to compliance with copyright laws.

Source link