Unless you literally programmed every component yourself, you can’t get away from some level of LLM involvement, even if the distro itself bans it. There is no 100% certain way to check code for LLM use. So if a developer doesn’t disclose it, and the maintainers accept the code, it’s in the official codebase now.

There’s a deeper problem though, which is a problem of demarcation. What counts as an “AI contribution” to you? Obviously vibe-coding is. But what about more subtle uses?

What if a developer uses an LLM to read the code and suggest a way to make a certain contribution, but it never actually generates any code, the developer implements the LLM’s suggestion completely by hand. Does that count?

What about if a developer hand writes 95% of the code themselves, but that last 5% they aren’t as confident in, so they have the LLM generate it. Does that count?

What if an extremely experienced developer uses an LLM to generate 100% of the code, but spends hours going over ever single line making sure it is correct and making little tweaks before submitting the code?

What if a developer writes the initial code 100% by hand, and only uses an LLM to go through and optimize the formatting and performance before submitting it?

What if a developer writes 100% of the code by hand, but uses an LLM to generate a bunch of unit tests and documentation?

These are just some of the ways that LLMs can be involved in the process of software development, and most of these wouldn’t have any real chance of being caught by a maintainer doing some sort of scan to detect LLM usage.

I’m not saying these uses are or aren’t justifiable/acceptable, but my point is it’s often not clear what people consider to be “over the line”. Almost no experienced devs in the FOSS world are advocating for the 100% vibe-code/token-maxing approach. But lots of them are comfortable with using LLMs in similar ways to what I listed earlier.

The tough truth is that you’re almost certainly already using software that has LLM involvement at some level. Short of going full Amish mode, or using completely pre-LLM software in some kind of air gapped setup, you’re not going to be able to avoid it.

The linux kernel is compromised. Lets use NetBSD!

No. It’s not practical.

No matter how you feel about ai or how you think it is wrong, the increased tempo of bug fixes and security patches necessitates either manpower beyond what is feasible for volunteer projects or the use of ai tools.

It’s an arms race and you can’t opt out without just missing the patch schedule.

Are Anti-AI Linux distributions practical?

It’s not possible to have a 100% no-LLM-coding-involved Linux distro anymore now that the kernel itself is welcoming LLM-assisted contributions, but it’s still possible to say that everything which the distro produces itself will not use AI.

How can AI commits be responsible?

What Debian means is: by reading what it produces and not just blindly accepting it but having humans review it carefully. The quality of LLM-assisted software is highly variable. Experienced people like Linux and Debian developers are much less likely to publish very bad slop than inexperienced developers who are newly able to make a thing work well enough to ship it.

I am concerned about the extent of Debian’s embrace of LLMs but I still have a lot more confidence in them than most other distros (especially small hardcore-anti-AI distro projects lacking sufficient developer capacity to even keep up with security updates).

I expect we’ll be seeing lots more things like this non-upstreamable-because-AI hardware enablement and distros like Nura (formerly postmarketOS) will either either adopt more nuanced LLM policies or be forked.

I am doubtful about if there will be any serious distros with hard LLM bans in a few years from now. But also, perhaps the bubble will have burst by then, so hosted inference won’t be so cheap and the use of LLMs will probably be quite different.

the primary reason is AI is a thief

This is not the primary reason for most people, I don’t think. I think the fact that it increases complexity while reducing human understanding is the biggest problem; it’s possible (easy even) now to produce mountains of good-enough-by-many-metrics source code which the person doing the prompting has very little comprehension of. The fact that it is also a plagiarism machine is far less of a problem than this. (There are a litany of other problems too of course, several others of which I would also rate as worse than the plagiarism aspect.)

thief

If you advocate for further expansion of “copyright” oppression, you deserve microsoft windows.

Is most of the training data for popular LLM tools stolen? Yes, but the algorithms are not so from a theoretical perspective there’s nothing inherently wrong with developers using AI (which itself is a broad term covering frontier models to simple neural net classifiers) but it’s a matter of their tools which afaik is the responsibility of the developer. (Ex. Devs aren’t banned from using VS Code even though Microsoft is not a good company) (Ex. Dont upload user PII to online tools just to convert a pdf you’re responsible if they steal user data)

So really responsible AI means developers accept liability for any theft or plagarism these tools commit, just as they would if they plagarized code themselves which is less problematic for developers than introducing precedent for restricting developer tooling.

In theory you could revert the commits that are labelled AI-generated but given the scale and application of LLMs that’s a lot of code to rewrite that was already reviewed by maintainers for a lot of important processes and will have likely undergone far less scrutiny and correctness tests. Until AI companies are held responsible for training data that’s stolen and their generated outputs are held to legally be plagarism there is negative incentive to rewrite ai-generated code.

Sure but why?

the primary reason is AI is a thief.

That’s the least important reason to not use AI. All property is theft, especially copyright!

There is a difference between people using AI to help write/maintain code that they understand, and slopcoding unmaintainable vibe projects.

https://www.debian.org/vote/2026/vote_002#texte

The Debian Project nevertheless expects that all contributions submitted to Debian, regardless of how and with which tools they were produced, satisfy the same standards of quality, correctness, maintainability, and legal compliance. The use of a generative AI tool does not diminish the contributor’s responsibility for the work they submit. Contributors are expected to understand, review, test, and, where appropriate, modify AI-assisted output before incorporating it into Debian. Blindly accepting or uploading AI-generated material without appropriate human review is inconsistent with Debian’s established development practices. We enourage [sic] our contributors to disclose whether a contribution was made with AI assitance[sic], but do not require them to do so.

the only good to come of it is the destruction of the concept of intellectual property, the tools of capital won’t serve us

AI companies and AI users either do not care or don’t know/understand the value of software freedom. I only used Debian from Mint [Mint being based on Ubuntu, Ubuntu based on Debian] but they used to appear as a champion of software freedom. It was somewhat understandable when they choose to no longer ~seperate proprietary drivers/codec/software but, whoever runs Debian now, I think their true colors are showing with AI.

I avoided using Debian as I thought it would be too hard to get up to date drivers! Gentoo has binaries for most of what I need and compiles the rest for me. It may not be as difficult to use as you think. However, trying to install Gentoo would definitely be… an experience 🫣