Unless OpenAI can prove that the outputs are derived from legally vs illegally-obtained outputs, not sure that’s going to matter. And as far as I understand about their models, that’s effectively impossible.

Isn’t the burden of proof on the other side?

Not when OpenAI publicly declared they trained on pirated works. I can’t imagine “we can’t tell if this is the result of the illegal thing we did or not” is going to stand up very well, nor does it bode well for any refutation of the plaintiff’s depiction of their intent. Part of fair use consideration is commercial impact and when you steal a bunch of books to train your AI model, it’s hard to refute that the impact is not negative or that you didn’t intend commercial harm.

Please read more carefully. OpenAI never “declared they trained on pirated works.”

+1

However, the root of this thread quotes Meta admitting exactly that.

That's not a claim Silverman has to make. "You copied my book without license to do so" is sufficient.

Did they?

Do you think scraping huge swathes of the internet contains pirated works or not?

I see, they didn't.

Where does it say they scraped huge swathes of the internet and didn't look at the results?

> it is just as likely it saw any number of things about it

Is this based on inside information, or just the law of averages? Doesn't the fact that they openly admitted to having been trained on pirated books affect your priors?

They didn't, more conjecture

It seems to me that they are indeed admitting to using the pile/books3 dataset, which seems to contain Silverman’s book, at least.

See post https://news.ycombinator.com/item?id=36659041 for a summary.