Hacker News Reader

Comment by EMIRELADERO | Hacker News Reader

What if the whole pipeline (scraping for training, the training itself, model distribution, and use to generate derivative works) is found to be a fair use (at least in American law)? Licenses wouldn't mean anything at that point, since it's because of copyright that they can make you accept them in the first place

View on HN

Then every employee of openai should scrape the weights of GPT-4 and train their own neural nets. That would not be a derivative work and also be free use under this logic.

Obviously this isn't tested in the courts (for now) but it is my opinion that model weights aren't even copyrightable.

I don’t think we need a court case for that. Model weights aren’t copyrightable by the trainer in the US. Copyright only protects works of human authorship, and training a network is not authorship.

If the model weights encode other copyrighted works literally enough then they may be copyrighted by the author(s) of the works, that is a gray area. But the training process itself is not authorship and doesn’t confer copyright.

That's not how fair use works. This kind of use would facially fail three of the factors for fair use; it's not be transformative, it copies the original work in its entirety, and it harms the commercial market for the original work.

how much do the weights have to change for it to be transformative? Edit: you arrive at the same question as taking an image with img2img and running it through a diffusion variant - keeping the composition, colors etc. but no the details

Much of the case law about the "transformative" factor focuses on "new meaning or expression", but it's about visual art, which is generally very difficult to reason about w.r.t. copyright. I think the example to look to for technology is Authors Guild v. Google, where "transformative" is more about non-expressive purpose, and it was considered transformative to copy a bunch of books to produce a search functionality, since the search functionality (which only displayed snippets) was a transformative purpose compared to the underlying creative expression in the books.

Fair use doesn't enter into it if weights aren't copyrightable. They're machine generated by stochastic gradient descent. There's no human hand setting the weights. We won't know until something ends up before the USCO.

It seems credible to me to suggest that the model weights are a trade secret, but aren't copyrightable. There's lots of stuff that could be in that category for other companies.

The employees would still have a contractual responsibility about their use of the model weights.

Reply on news.ycombinator.com