but we’d first have to have models trained without violating the rights of all the authors of the works in the training data. Like if you trained a model only on GPL software, then the model and its output should also be GPL.
I’m not so convinced on that one. Copying code without permission, yes, that’s illegal. Distributing copied code without permission, also illegal. But processing code (copied with permission) into weights, that doesn’t feel to me like a violation. If I have a legit copy of something, I should be able to distribute computed statistics about that copy.
Reproduction of the material we agree is against the law.
A model is very obviously a derivative work, as it has 0% original content input. Under fair use law, it also very obviously fails all four tests.
Computing the letter frequency has 0% original content input, as well, but is definitely not a derivative work. Do you disagree?
I’m not so convinced on that one. Copying code without permission, yes, that’s illegal. Distributing copied code without permission, also illegal. But processing code (copied with permission) into weights, that doesn’t feel to me like a violation. If I have a legit copy of something, I should be able to distribute computed statistics about that copy.
EFF points out some hazards: https://www.eff.org/deeplinks/2026/10/rewriting-copyright-law-generative-ai-would-hurt-everyone
Sure, but it’s not just statistics like “30% of lines contain the letter E”. It reproduces the material, sometimes verbatim.
A model is very obviously a derivative work, as it has 0% original content input. Under fair use law, it also very obviously fails all four tests.
Reproduction of the material we agree is against the law.