Recent searches
Search options
if the models have been trained on copyleft code, mustn’t all the code they generate also be copyleft?
@interfluidity great point! The question is… who has deep enough pockets to see the claim through?
@interfluidity does this go for code produced by a human who has ever looked at copyleft code?
(I'm not trying to deny there are differences; just trying to think through what they are...)
@interfluidity I suspect the right* answer is that just like with code produced by humans, it depends on the details, and the model output may or may not be a derivative work depending on its actual relationship to the prior code
*(the answer I prefer, and the one I would guess is most consistant with the law but shrug)
@djc it’ll be interesting! i think courts have said that generative-ai images are not copyrightable, because they have no author. that’s a big break from an analogy with human education and fair use! humans say “this is my work, sure, it is influenced by stuff i’ve seen, but it’s original plus ‘fair use’” but with ai there is (in law) no author whose work can be original, the work can be nothing but derivative of what it has been trained on.
@interfluidity interesting
That seems wrong to me but I agree with the way you're reasoning from what courts have said
@interfluidity Well, that's in effect the claim of stablediffusionlitigation.com, regarding pictures -- that the output from these systems is derivative of every bit of training data that went into them, with all the legal implications. Not every lawyer agrees; some would say that to determine whether one thing is a copy of another, for copyright purposes, you look at the things, not the process which generated them.
@interfluidity (In software, companies sometimes use "clean room" development to prove that new software they're writing *can't* have been a copy of an older program they're trying to emulate. But the requirements of the law are not that strict; if a programmer *does* read the GPLed code for the game Quake, that doesn't make any other game first-person shooter they write GPLed derivative code, just because they'd seen the Quake release.)
@rst yes. for humans we presume there is a ghost in the machine: what we create is “ours”, not just merely derivative of what we may have seen. so far, though, for generative AI courts have said there is no ghost, no author beyond transformation of the inputs. copyright can’t be enforced on generative AI images, because there is no author. there is no one whose “fair use” can disencumber inputs, perhaps even very glancing inputs.