I guess in the same way we accept that the open source & licensed code was used to train the models, or in the same way we accept that Google / Meta / Anthropic and all the others use copyrighted material to get the degree of intelligence that it needs.

If you extend this thought and this rationale to humans as well, we shouldn't be allowed to "get smarter" by consuming copyrighted content (movies, books) or anything really - as we might end up referencing them.

I think copyright laws are being bended a lot here - and I don't think (as a copyright holder myself) I'm too concerned about that - if this then allows humanity to move forward. It's a philosophical topic really, more than a legal one

Same reason/way projects like ReactOS/Wine and friends exist in a world where source code from early 21st century Windows is literally sitting in a bunch of 6+ year old Github repositories, I guess. Or any number of other proprietary codebases. Legal limbo, effectively.

I'm not saying that's an equivalent problem to LLM taint, but it's not exactly unrelated, either. It may be the case that all LLM output is tainted, but it's also likely the case legally that you can't meaningfully prove that contributions from random people to FOSS aren't tainted.

Legally proving that an LLM hasn't pulled from proprietary software is probably in the same ballpark as proving some Wine contributor hasn't looked at any of those Github repos.

I dunno how ReactOS/Wine manages that threat/exposure. But I'm pretty sure both have fairly liberal contributor policies. They're not Compaq reversing the IBM PC BIOS.

Yes, the difference here is that most projects don't have to accept that (level of) risk, while ReactOS/Wine don't have a choice. But they do still trudge on regardless.