Meta’s AI Accused of Copying Books-Billions at Stake

Quick Summary

• Courts in the US and UK are deliberating on lawsuits concerning tech companies’ use of copyrighted books to train AI models. Billions of dollars are at stake.
• Researchers found that Meta’s Llama 3.1 70B model has memorised large amounts of copyrighted text,including full books such as Harry Potter,The Great Gatsby,and Orwell’s 1984.
• The Books3 dataset, containing nearly 200,000 copyrighted works, was widely used by AI developers; Meta faces allegations for breaching copyright in training models wiht this dataset.
• Researchers estimate that even minor infringements on the Books3 dataset could result in damages exceeding $1 billion for Meta.
• Methods tested a model’s ability to reproduce verbatim excerpts by prompting it with sections from source texts; results showed varying levels of memorisation across diffrent AI models, but Meta’s model exhibited particularly high verbatim recall rates.
• Legal opinions differ: US “fair use” doctrine offers broader exceptions for unlicensed usage compared to the narrower UK “fair dealing” concept.