I'm waiting for AGPL to become AIGPL: If you train a model with some or all of the licensed work, you agree that the weights of that model constitute a derivative work, and further for the weights, as well as any inference output produced as a result of those weights to be bound by the terms of the license. If you run a model with the licensed work in part or in full as input, you agree that any output from the model is bound by the terms of the license.

You can't change the law with a license agreement and redefine what constitutes a derivative work. If that was possible, people could have done it pre-LLMs.

also how would you prove it was in the training set? re: your last sentence, the licensed work wasn't in the input in the chardet example ("no access to the old source tree")

Sure, a license can't create new legal understanding of "derived work", but I think the intent of what Splinelinus said still works: a license outlines the terms under which a licensee can use the licensed Work. The license can say "if you train a model on the Work, then here are the terms that apply to model or what the model generates". If you accept the license, those terms apply, even if the phrase "derived work" never came up. I hope there are more licenses that include terms explicitly dealing with models trained on the Work.

Also, for comparison, both GPL and LGPL, when applied to software libraries (in the C sense of the word), assert that creating an application by linking with the library creates a derived work (derived from the library), and then they both give the terms that govern that "derived work" (which are reciprocal for GPL but not for LGPL). IANAL but I believe those terms are enforceable, even if the thing made by linking with the library does not meet a legal threshold for being a derived work.

Yeah, that's possible, but seems to me more about contract law and creating an EULA for the code, than it is about copyright-derived enforcement. maybe 'copyleft' stuff will move in that direction.

it's barely tangential to the topic but worth pointing out, I don't think there's firm legal consensus on your library point, that is just the position of the FSF that that's true. IANAL tho. https://en.wikipedia.org/wiki/GNU_General_Public_License#Lib...

This is also my thinking. A(ffero)GPL does something similar by saying a user of an API to AGPL code is bound by the AGPL license. You can always choose not to use the code, and not to use the license.

For the parent comment on discoverability, I honestly don't know. Some models list their data sources, others do not. But if it came down to a dispute it may be that a court order could result in a search of the actual training data and the system that generated it.

For the second case of derived work through context inclusion, it may end up in a similar situation with forensic analysis of the data that generated some output.

Agree. But then, the test suite was the input (chardet). So, is the test suite creative or functional in nature? And does the concept of fair use apply globally?

Bingo. I can see this is a possible future, and probably desirable scenario for anyone with preference for free software.

So write it! Shouldn't be much extra to add to the AGPL licence?

I think you're seeing a magic trick that's been around forever: https://www.youtube.com/watch?v=o-KDuOvMAfQ&t=80s

This article is kind of vague on that tbf: To conclude, we observed no credible evidence for a beneficial effect of L-dopa (vs. Haloperidol) on reinforcement learning in a reward context, as well as the proposed mechanistic account of an enhanced striatal prediction error response mediating this effect.