Meta's Llama 3.1 AI memorises large portions of Harry Potter
A new study found that Meta's Llama 3.1 AI model memorized and could repeat more than 42% of "Harry Potter and the Philosopher's Stone" word-for-word—way more than any other model tested.
This challenges the idea that AI doesn't really copy books, and it brings up some big copyright questions.
How well did the models perform?
Researchers split up 36 books, prompting the models with part of a passage to see if they'd spit out the next bit exactly.
Llama 3.1 (with its massive 70 billion parameters) remembered way more text than Meta's earlier version or rival AIs from Microsoft and EleutherAI, especially when it came to popular titles like "The Hobbit" or "1984."
Trouble ahead for companies using these models
If an AI can memorize whole chunks of copyrighted books, that could mean trouble for companies using these models—maybe even breaking copyright law.
As co-author Mark Lemley put it, this raises significant legal concerns, especially since there are reports Meta trained on pirated e-books.
The study shows we might need new rules for how AI learns from creative work in the future.