This article makes some reasonable points but a lot of bad ones, and ultimately even the "good" ones don't really support its conclusion.
> But rejecting code not for technical reasons (of which there are often many for vibecoded code) but for ideological reasons is unfortunate. [...] Such a policy — that contributing to a project because of the circumstances of the contribution’s creation, and not due to its quality — is directly contradictory to the spirit of free software.
Okay, what is FLOSS? Software published under free and open-source licenses. It's literally the definition of the concept, and thus a FLOSS project cannot accept contributions for which such licensing is impossible, regardless of technical quality. For example, a FLOSS project can't accept code copied out of a nonfree project, or copied from a project with an incompatible license. The provenance of the code has always been central to the whole endeavor.
When LLMs enter the picture, there are at least two problems.
1) LLMs were trained largely on copyrighted works, and it's unclear if the models themselves or their outputs should be considered to be derivative of their training data.
2) It's possible that LLM outputs can't be copyrighted at all, not as FLOSS or under any other license.
I won't try to get into the many arguments and opinions on both sides regarding those issues, but suffice to say they're contentious enough that, regardless of where an individual maintainer's opinion may fall, IMO it's definitely not wrong to refuse to accept LLM-generated code just to avoid needing to resolve them.
The "spirit of FLOSS" was never primarily about the technical quality of code, but about the freedom to use, study, modify, and redistribute software, although technical quality also benefits in some ways from the FLOSS ethos. I actually do care deeply about technical quality as well, so I have to mention that many of us have serious doubts about the overall effects of LLMs on software quality in the long term. Even if individual maintainers can maintain high standards for their projects, IMO promoting LLMs still creates long-term downward pressure on software quality that comes back to harm everyone.
I also failed to follow the logic of that quoted part.
Free software is the four freedoms you mentioned - use, study, modify, and redistribute. There is no obligation to accept anything, for any reason.
I don't think the argument that AI is a tool, combined with the disbelief that people would reject a tool due to "the circumstances of the contribution’s creation, and not due to its quality" is a strong argument for broad AI use, because we know examples were many people stopped using high quality tools due to the circumstances of its creation.
The Atlas of Topographical and Applied Human Anatomy was created by the Nazi Eduard Pernkopf and his four Nazi artist assistants. It is a "scientific and artistic masterpiece" ... created as the result of Nazi medical research on the corpses of executed political prisoners, and containing Nazi symbols in the original artwork which were airbrushed out for wider, post-war release. https://en.wikipedia.org/wiki/Eduard_Pernkopf
Sure, for some the book fundamentally a tool, but "[o]pponents have asserted that any use of the atlas makes the user complicit in Nazi crimes".
The Wikipedia entry ends with a quote from Pieter Carstens: "Beautiful anatomical drawings were created, but this was only made possible by the unethical and unlawful procurement of the anatomical remains of murdered victims of an evil Nazi regime–thus beauty and evil were fused. This fusion not only perverts and diminishes the status and content of the Pernkopf Atlas, but also explains why it should be rejected. / [It] should be permitted to show its duplicitous face only rarely and then for very good reason in the teaching of history, medical ethics and medical law so that its lessons will be learned and its history never repeated."
This is not a call for a ban of this tool (the atlas), but rather a call for highly limited use, for very specific purposes.
Since I think I would be one of those opponents to widespread use of the atlas, you can see how I would factor in "the circumstances of the contribution’s creation", even for a high quality contribution.
“Arguments that llms fundamentally a part of a system of political oppression fall flat against the case of open-weight models that can be run at home on (high-end) consumer hardware.”
Can anybody comment on the validity of this in 2026? Compared to Fable on a $200 monthly subscription, what is a $2400 investment in a home setup (including electricity) getting me in a year?
Qwen 27B is runnable for that price and it can produce patches that work using opencode as a harness. It's obviously not as capable as frontier models but it can do things if you provide the judgement.
I mean, spend a bit more to take into account the fact that deprecation on a graphics card (computer equipment) is five years and you can reasonably spend about 2400 * 5 = 12000 on a card which will get you an rtx pro 5000 blackwell with 96GB RAM which can run top-notch OSS models and fine-tune them for your purposes.
It's just a computer. An expensive one, but a computer nonetheless. Unless OSS authors are going to be so fundamentalist as to insist the computer hardware also be free as in beer, then the argument is really just a dogmatic one
It seems undeniable that LLM "vibe coding" presents an existential threat to F/OSS, its legal underpinnings, and its ethos.
Richard Stallman's vision involved an exploitation of copyright to ensure that software would become free and remain free, and get even freer as we went along. The GPL and its derivatives have been, for decades, a really potent weapon to wield against closed-source, corporate interests, and proprietary software. Especially in the form of GNU/Linux: the kernel and its major distros are ultimate bastions of GPL software.
Creative Commons relies on copyright to work. BSD, MIT, Apache, all the F/OSS source licenses, and permissive licenses: they all rely on copyright. So how are LLMs going to help here?
If LLM output remains uncopyrightable, then we're in trouble. It seems that only a minority of jurisdictions confer any rights to the "prompters" or "vibe coders" of genAI works. If any given contributor to an F/OSS project cannot own, copyright, or assign rights from their works, then the project cannot remain "free".
Oh, sure, any F/OSS can become public domain, and purely public domain works are indeed "free software" but they can't be GPL'd and they can't be subject to any additional "freedoms" such as requiring attributions. Public domain software can be subsumed into trade secrets and closed-source projects. What company will want to contribute public domain code to any public project that they've written with LLMs... they will simply go closed-source.
I think any F/OSS dev should be concerned, at this point, with LLM contamination of code bases. Many projects are wisely clamming up and prohibiting AI-generated code, to the extent possible, but that's sort of impossible, isn't it? Contributors will simply claim ownership and claim manual output, when it wasn't, and who will prove them wrong?
This article makes some reasonable points but a lot of bad ones, and ultimately even the "good" ones don't really support its conclusion.
> But rejecting code not for technical reasons (of which there are often many for vibecoded code) but for ideological reasons is unfortunate. [...] Such a policy — that contributing to a project because of the circumstances of the contribution’s creation, and not due to its quality — is directly contradictory to the spirit of free software.
Okay, what is FLOSS? Software published under free and open-source licenses. It's literally the definition of the concept, and thus a FLOSS project cannot accept contributions for which such licensing is impossible, regardless of technical quality. For example, a FLOSS project can't accept code copied out of a nonfree project, or copied from a project with an incompatible license. The provenance of the code has always been central to the whole endeavor.
When LLMs enter the picture, there are at least two problems.
1) LLMs were trained largely on copyrighted works, and it's unclear if the models themselves or their outputs should be considered to be derivative of their training data.
2) It's possible that LLM outputs can't be copyrighted at all, not as FLOSS or under any other license.
I won't try to get into the many arguments and opinions on both sides regarding those issues, but suffice to say they're contentious enough that, regardless of where an individual maintainer's opinion may fall, IMO it's definitely not wrong to refuse to accept LLM-generated code just to avoid needing to resolve them.
The "spirit of FLOSS" was never primarily about the technical quality of code, but about the freedom to use, study, modify, and redistribute software, although technical quality also benefits in some ways from the FLOSS ethos. I actually do care deeply about technical quality as well, so I have to mention that many of us have serious doubts about the overall effects of LLMs on software quality in the long term. Even if individual maintainers can maintain high standards for their projects, IMO promoting LLMs still creates long-term downward pressure on software quality that comes back to harm everyone.
I also failed to follow the logic of that quoted part.
Free software is the four freedoms you mentioned - use, study, modify, and redistribute. There is no obligation to accept anything, for any reason.
I don't think the argument that AI is a tool, combined with the disbelief that people would reject a tool due to "the circumstances of the contribution’s creation, and not due to its quality" is a strong argument for broad AI use, because we know examples were many people stopped using high quality tools due to the circumstances of its creation.
The Atlas of Topographical and Applied Human Anatomy was created by the Nazi Eduard Pernkopf and his four Nazi artist assistants. It is a "scientific and artistic masterpiece" ... created as the result of Nazi medical research on the corpses of executed political prisoners, and containing Nazi symbols in the original artwork which were airbrushed out for wider, post-war release. https://en.wikipedia.org/wiki/Eduard_Pernkopf
Sure, for some the book fundamentally a tool, but "[o]pponents have asserted that any use of the atlas makes the user complicit in Nazi crimes".
The Wikipedia entry ends with a quote from Pieter Carstens: "Beautiful anatomical drawings were created, but this was only made possible by the unethical and unlawful procurement of the anatomical remains of murdered victims of an evil Nazi regime–thus beauty and evil were fused. This fusion not only perverts and diminishes the status and content of the Pernkopf Atlas, but also explains why it should be rejected. / [It] should be permitted to show its duplicitous face only rarely and then for very good reason in the teaching of history, medical ethics and medical law so that its lessons will be learned and its history never repeated."
This is not a call for a ban of this tool (the atlas), but rather a call for highly limited use, for very specific purposes.
Since I think I would be one of those opponents to widespread use of the atlas, you can see how I would factor in "the circumstances of the contribution’s creation", even for a high quality contribution.
“Arguments that llms fundamentally a part of a system of political oppression fall flat against the case of open-weight models that can be run at home on (high-end) consumer hardware.”
Can anybody comment on the validity of this in 2026? Compared to Fable on a $200 monthly subscription, what is a $2400 investment in a home setup (including electricity) getting me in a year?
Qwen 27B is runnable for that price and it can produce patches that work using opencode as a harness. It's obviously not as capable as frontier models but it can do things if you provide the judgement.
60+ GB VRAM, unless you quantize. If you quantize sure, but then it will not really be comparable to what others are using.
A model that doesn't change under your feet every few weeks. Great if you need to consistently process small data tasks over the course of months.
I mean, spend a bit more to take into account the fact that deprecation on a graphics card (computer equipment) is five years and you can reasonably spend about 2400 * 5 = 12000 on a card which will get you an rtx pro 5000 blackwell with 96GB RAM which can run top-notch OSS models and fine-tune them for your purposes.
It's just a computer. An expensive one, but a computer nonetheless. Unless OSS authors are going to be so fundamentalist as to insist the computer hardware also be free as in beer, then the argument is really just a dogmatic one
The elephant in the room here is that there are no FLOSS LLMs. Renders most of the rest of this argument moot.
It seems undeniable that LLM "vibe coding" presents an existential threat to F/OSS, its legal underpinnings, and its ethos.
Richard Stallman's vision involved an exploitation of copyright to ensure that software would become free and remain free, and get even freer as we went along. The GPL and its derivatives have been, for decades, a really potent weapon to wield against closed-source, corporate interests, and proprietary software. Especially in the form of GNU/Linux: the kernel and its major distros are ultimate bastions of GPL software.
Creative Commons relies on copyright to work. BSD, MIT, Apache, all the F/OSS source licenses, and permissive licenses: they all rely on copyright. So how are LLMs going to help here?
If LLM output remains uncopyrightable, then we're in trouble. It seems that only a minority of jurisdictions confer any rights to the "prompters" or "vibe coders" of genAI works. If any given contributor to an F/OSS project cannot own, copyright, or assign rights from their works, then the project cannot remain "free".
Oh, sure, any F/OSS can become public domain, and purely public domain works are indeed "free software" but they can't be GPL'd and they can't be subject to any additional "freedoms" such as requiring attributions. Public domain software can be subsumed into trade secrets and closed-source projects. What company will want to contribute public domain code to any public project that they've written with LLMs... they will simply go closed-source.
I think any F/OSS dev should be concerned, at this point, with LLM contamination of code bases. Many projects are wisely clamming up and prohibiting AI-generated code, to the extent possible, but that's sort of impossible, isn't it? Contributors will simply claim ownership and claim manual output, when it wasn't, and who will prove them wrong?
local models, yes. Cloud models, no.