A study finds GPT-4o seems to know books behind a paywall

Researchers tested if OpenAI models recognize books behind a paywall. OpenAI had no licence to train on them. Such tests can point to access violations. They used 34 books from a publisher, with 13,962 paragraphs, that they got legally. GPT-4o got a score of 0.82 on a scale where 0.50 is chance. The authors say the result is preliminary.

How it works

The team used DE-COP, a test of if a model recognizes text from its training data. The new part is the use of non-public books. This lets the test point to access violations, not only to training use.

The numbers

• GPT-4o: 0.82, with a confidence interval of 0.60 to 0.96.
• GPT-3.5 Turbo: 0.54.
• GPT-4o Mini: 0.56.

GPT-4o and GPT-4o Mini have the same training cutoff of October 2023. Because of this, the difference is not only a change in language with time.

What is not known

The authors name 3 limits. The sample of books is small. The control texts after the cutoff can have contamination. The models are different in dimension. The test shows recognition. It does not show how OpenAI got the books.

Insight: Only outside tests can check what a model learned

The signal is in GPT-4o, not in GPT-3.5 Turbo or GPT-4o Mini. The authors say the results agree with rules that make makers disclose training data, and with licence markets. This reads as a difference in control: the maker holds the data record, and publishers can only test from the outer side. Leaders who use these models must tell vendors to give the terms on training data.