Additionally, any license that explicitly forbids the use in training an AI model would not be an open-source license, for the same reasons that you cannot prohibit commercial or military use with an open-source license.

@BartvanIngenSchenau: I understand the open source requirement is related to "field of endeavor" and not "business model". So the freedom must include using the library in open source AI training software, but it's not so clear that feeding the library source code into an AI as training input is a guaranteed freedom.

It absolutely is a derivative work, mechanically, but politics, not reality, will determine whether courts rule it as such. As FOSS folks, we should be fighting to ensure they do.

@BartvanIngenSchenau: Not true. You're confusing "use" with "prepare derivative works". GPL already forbids using the software as AI training material unless you plan to GPL the whole resulting model and release it in its "preferred form for modification". That has nothing to do with a field-of-endeavor use restriction, which would be like if a text editor said you couldn't use it as part of working at an AI company developing AI models.

@R..: Even if it is a derivative work, the vast majority of FOSS licenses have no limitations on the preparation of derivative works. They instead limit the distribution of such works, as explained in my answer.

@R..GitHubSTOPHELPINGICE It is not obviously a legally derivative work. If a human writer writes genre fiction they probably learned how to write that genre by reading lots of existing books, but whatever they write isn't automatically considered derivative. (It may be that AI is derivative because the AI can't really exercise creativity, but it's not immediately obvious that it is so--for example, there are many authors who write what could be called "boilerplate novels"...meaning that they seem very repetitive, but aren't necessarily derivative, legally speaking.)

@user3067860: That analogy (human learning to write) is based on a magical-thinking version of what "AI" is, not what the actual math is.

@BartvanIngenSchenau Wouldn't be open source according to the FSF... It would be open source according to what those words actually mean though...

@R..GitHubSTOPHELPINGICE I'm not saying ML is magical but it's not necessarily obvious that in a large learning model any single input contributes enough that the output can be considered a derivative work of that specific input. Especially not when you consider code where things like API calls, etc., are not copy-right-able. (E.g. every time I write an integral it is probably a derivative work of my intro-to-calculus textbook, but equations aren't copy-right-able so I'm off the hook for that one.)

@BenVoigt, to me, forbidding someone (through the license) to include the code in an AI training set is similar to forbidding them to include the code in a Visual Studio project. But note that I fully agree with the answer from Philip that it fully depends on how AI-generated code is seen in relation to its training data from a copyright law perspective.

@R..GitHubSTOPHELPINGICE What does copyright law say? Europe: DSM Directive Article 4: "Exception or limitation for text and data mining" states that text and data mining is legal (I assume other jurisdictions are similar). As OSS licenses cannot limit what is already allowed by law, the question (if AI learned models are 'derivative works') is essentially irrelevant.

@user3067860 Note that facts are not copyrightable, but creative expressions are. You are off the hook for remembering integrals from your intro-to-calculus textbook because they are facts. Likewise, AI is off the hook for remembering that math.floor calculates the floor and "quack" is usually a verb.

@BartvanIngenSchenau Regarding the prohibition of open-source code use for military purposes, I noticed that the glm 3d math library is using Happy Bunny Licence (modified MIT), which explicitly states: "Restrictions: By making use of the Software for military purposes, you choose to make a Bunny unhappy.". I guess that is as close as we get, then? :)

@ScottishTapWater "Wouldn't be open source according to the FSF... It would be open source according to what those words actually mean though" - It may also violate criteria 6 and 10 of the Open Source Definition, so I'm not sure I agree with your statement.

@user253751: The analogies to human memory are just fundamentally invalid here. Stop making them. These are the analogies non-technical judges will end up buying and screwing us all over with.

@user3067860: For any given fixed piece of training material from which the model is derived, it's not guaranteed or even likely that a randomly chosen output is a derivative of that work. However, there are lots of examples where output is verbatim, unquestionable copying of something unique and creative from the training material. This demonstrates both that the model is derivative and that the output is. This is why I call these models "stochastic plagiarism".

@R..GitHubSTOPHELPINGICE "This demonstrates both that the model is derivative" - I believe it can be that the model is covered under Fair Use due to its transformative nature (like Google Books) even if certain uses of some outputs wouldn't be (like copying pages obtained from a Google Books query into your own book). "copying of something unique " - Haven't yet seen any regurgitation examples that aren't widely reused snippets, but that can still be problematic.

@Dryt There's one well known example of Stable Diffusion doing a direct copy

@PhilipKendall Skimming paper so correct me if I'm wrong - but it looks like they use prompts from the "most-duplicated examples from the training dataset", so roughly the equivalent of widely reused code snippets. Result given those circumstances (and sorting "175 million generated images" by similarity to training set images) is honestly better than I would have expected. Not entirely without issue, but to me that seems a lower infringement rate than humans.

@R..GitHubSTOPHELPINGICE I'm curious, why would the analogy to human memory be harmful?

@Dryt: A derivative work that contains sufficient information-theoretic content to derive back from it large quantities of copyright-protected content, especially when the primary purpose for which it's being marketed is to "create content", is not fair use by virtue of being "transformative". It's a laughably bad attempt at circumventing copyright - or it would be laughably bad, if it didn't have billions of dollars behind trying to make it look serious.

@justhalf: Because it's a trick used to deceive people for the advantage of the parties peddling "AI". When someone tries to obscure the mechanism of something and make it look mysterious/magical/unknowable, they're either trying to entertain you (a magic show) or trying to dupe you. If you're asking why, mechanically, "AI" is not like human memory/lerning, that's a topic for a dissertation, but a few very obvious surface-level differences: it's copyable and it cannot be held accountable.

@R..GitHubSTOPHELPINGICE "that contains sufficient information-theoretic content to derive back from it large quantities of copyright-protected content, especially [..], is not fair use by virtue of being transformative" Doesn't seem generally true. Google Books/thumbnails/caches contain more than something like SD with an upper bound of less than a byte average per training image (but complicated by dupes). Maybe if you mean the especially term as mandatory, and the purpose of a content creation tool is judged not different to that of a training sample? Otherwise, purpose works in its favor

@R..GitHubSTOPHELPINGICE I'm more interested in the surface-level differences indeed. But what do you mean by copyable? And how does that increase the harm of using the analogy to human memory? (I ignored your first part for now since it seems like a rant if we haven't discussed the second part)

@user3067860 “If a human writer writes genre fiction they probably learned how to write that genre by reading lots of existing books” – yes, and there's no escaping the fact that AI can do something that is equivalent (or will soon be able to) as far as the information is concerned. But it's still different in its impact, because as long as a human does the learning-and-reproducing than there's a significant limit on both how soon the (non-)-derivates can be produced, how much such material can be produced, and how cheaply it can be done.

...That's why IMO a distinction must be made between when a human learns and bases work on it, and when an AI does it. The distinction can't be deduced from existing laws and therefore must be an axiom (i.e. politicians should write new laws that explicitly say there's a difference between humans or AI doing the same task).

Outlawing based on efficiency to ensure human labor remains a requirement seems a wasteful route to go down. Should translation tools, which were trained on human translations and can produce translations faster than any human, be illegal?