Post B9wBlkMb8QZPaQ6Fs0 by tiotasram@kolektiva.social
 (DIR) More posts by tiotasram@kolektiva.social
 (DIR) Post #B9vHlHYeddsgoYwP0C by skjeggtroll@mastodon.online
       0 likes, 0 repeats
       
       @freshstart @neil I would think that just based on license concerns alone would necessitate a policy of "No source-code or binary files may be generated wholly or in part by an LLM."If _you_ didn't write it, or if you don't have a license to relicense it, you can't in good order publish it under an open-source license. (Or, for that matter, a closed-source license.)
       
 (DIR) Post #B9vHlHqjYQLJidOqBM by tiotasram@kolektiva.social
       0 likes, 0 repeats
       
       @skjeggtroll @freshstart @neil it's worse than this actually.We know that LLM code sometimes includes fragments of training data verbatim. We know much training data is licensed such that verbatim copies requires attribution. So we know that LLM code sometimes violates these attribution requirements.There is no practical way to check whether a given section of LLM produced code violates an attribution requirement. How are you going to search against all lines of code in the training database? Do you run that search for each line of code?Even worse: two different people using the same LLM, even with different prompts, may generate the same or substantially similar code. Whoever tries to slap a license on that code second is in violation of the first person's license, assuming a world in which such code is licensable at all. These people probably don't know of each others' existence at first. This has already happened, see:https://blog.terrygodier.com/2026/08/09/mea-culpa-dark-hours.htmlTo accept any LLM-generated code or documentation is to say that you are fine with license violations. If, as Debian did, you try to thrust the impossible burden of ensuring license violations do not occur onto each contributor, I'm not sure what you're saying, but your policy isn't sincere.
       
 (DIR) Post #B9vHlI72ZnO2XD1rbE by lxo@snac.lx.oliva.nom.br
       0 likes, 0 repeats
       
       FYI, there's a serious flaw in the following argument:Whoever tries to slap a license on that code second is in violation of the first person's licensethat's not how copyrights and licensing work.  if two or more people independently come up with the same original creative expression, each one gets a separate copyright on it, and can license it accordingly.  (if they use LLMs, there's uncertainty as to getting copyrights at all, but that's a different issue)you're probably thinking of patents, that involve a race even between independent inventions.  but patents and copyrights are so different that expecting similarities leads to confusion.CC: @skjeggtroll@mastodon.online @freshstart@hachyderm.io @neil@mastodon.neilzone.co.uk
       
 (DIR) Post #B9wBlkMb8QZPaQ6Fs0 by tiotasram@kolektiva.social
       0 likes, 0 repeats
       
       @lxo @skjeggtroll @freshstart @neil Okay, that's what the law says, but how is the second person going to fare against a lawsuit if the first person claims infringement?If the second person succeeds on an argument that "using an LLM is like a clean-room implementation" then open-source is non-functional, legally, as anyone can easily ignore any license by asking an LLM to duplicate relevant functionality and then claiming plausibly that they hadn't seen the other work before.If, as you're suggesting, there's an "I came up with it independently" defense, does the burden of proof fall on the accuser or the accused? Normally burden always falls on the accuser, but if we assume open-source development on both sides, the accuser can trivially prove (if they published faster) that they came up with the idea first and that their code was theoretically accessible to the accused when the accused was developing their version, and that the accused used lines of code verbatim which are in their version. The accuser can probably even show that their code was likely in the training data of any LLM that the accused used.I don't actually know the law here, but if that's not enough to throw a presumption onto the accused and the accuser actually has to prove intent or something, it again seems like open-source licenses offer negligible protections against infringement (maybe this is actually the case). If the accuser on the basis of showing that the accused used verbatim copies of their public code can shift the burden of proof onto the accused, how is the accused going to prove they actually came up with the idea independently, especially when they used an LLM so it's not really their idea?Maybe some kind of twisted precedent *will* be established that this situation is fine actually and anyone accused of copyright infringement who is using an LLM can just claim their invention is independent despite having literal copies of another work in it... I don't want to be the person testing that legal theory against a startup with a 6-figure legal budget, let alone a big tech firm. The big tech firms regularly do patent software, after all.On the other side of things out it turns out that legally LLM-generated code cannot be licensed or patented at all, then it can't be open source and this should not be acceptable to Debian.But let's think about the moral level too. Do I want to be using the sometimes-steals-code machine and then get into a situation where it looks like I stole someone's code? No. What actually happened in this case was not that the slower sloperator mounted an arcane legal defense and everyone was okay to let the two apos coexist. Instead he backed off, apologized to everyone, deleted the project and promised never to use LLMs to generate code again. That seems like the best case for reputational damage. So even if the legal reality permits some really gross stuff (let's face it, this is the norm actually) unless you're a massive corporation who doesn't care about reputation, a legal technicality doesn't make this situation all good.
       
 (DIR) Post #B9wBlkf21tJcVaiybQ by lxo@snac.lx.oliva.nom.br
       0 likes, 0 repeats
       
       FTR, I didn't mean to join the entire discussion, just to address a minor point about copyright within your earlier post.  I don't really know what the proof standards and requirements look like, I just know that independent works, no matter how similar, aren't supposed to infringe each other's copyrights.  this fact has nothing to do with LLMs.  using LLMs may bring about doubts about authorship, provenance, and independence, but those don't favor claims of infringement between works made supposedly-independently with LLMs.  which is not to say that they can't both be found to infringe on other preexisting works used for training the model; there's plenty of uncertainty there.CC: @skjeggtroll@mastodon.online @freshstart@hachyderm.io @neil@mastodon.neilzone.co.uk
       
 (DIR) Post #B9wCAfhxa4faO4ydAO by lxo@snac.lx.oliva.nom.br
       0 likes, 0 repeats
       
       I'm not sure what you're getting at.  my point was about copyright in general, so it applies to copyleft as well, inasmuchas copyleft is built on copyrightCC: @tiotasram@kolektiva.social @skjeggtroll@mastodon.online @neil@mastodon.neilzone.co.uk