The filter is based on quality, and humans remain responsible for the work they submit. One hundred percent reasonable policy. Wish it was like this everywhere.
The winner being a middle ground policy is not too surprising. I think the most surprising part is that all of the permissive-towards-LLMs options won over all the forbidding-LLMs options, and by non-trivial margins (>10% of the ballots).
That doesn't match my impression based on reading Planet Debian. The silent majority seems to be real in this case.
I'm also a bit surprised, going by the discussion on the debian-vote mailing list. I really felt the two extremes were very closely matched in popularity. With Debian's sane choice of voting system, that would still produce a sensible middle ground (like it did), but as you say the results do tilt significantly in favor of an LLM-liberal attitude. That last part surprised me.
(Disclosure: I proposed Option G – Debian is created by humans – and obviously favored it.)
The anti side definitely seems louder in public posts. But it seems to make sense - I largely stopped trying to discuss LLMs outside of groups where people are interested in usage. Most opposing views are either not new (they may be still valid though) or are based on outdated ideas - there's no point disturbing that angry nest. But I hang out in a few groups that simply share interesting experiences/news and use the LLMs where it makes sense. I don't expect anyone from those groups to be interested in evangelizing that position, but given a vote they'd all be counted as llm-liberal.
The anti side definitely seems louder in public posts.
I think the anti side wants to gain leverage right now. This result in a way is neutral: you can use it or you can opt against using it and Debian doesn’t take a stance. But for the people who are against AI this result is a win for the people who want to use it.
You can already see at least one maintainer resignation as a result of this vote.
I'm not sure a "neutral" position is really neutral when it comes to the use of AI. I think there's an asymmetry we need to acknowledge here. It's almost impossible to work with AI-wielding people without using AI yourself. So, in a way, being AI-neutral is like being doping-neutral in a competition.
I know open source is not a competition, so it's not a perfect analogy, but as a maintainer, if you don't have the liberty to close PR as "AI;DR", then you simply can't keep up, constantly wasting your time with PRs that will take 100x the time to articulate why it's not up to par than the time it took to create it.
Saying "Humans responsible for the work they submit" without any actual punishing consequences for submitting flawed code is also completely toothless, and based on personal experience, is equivalent to saying nothing.
For what it's worth, if I had been eligible to vote, I would have shied away from this point in particular:
Documentation and translation added as part of contributing to Debian
I can foresee real value from (responsible) use of AI for direct contributions in some areas.
(Disclosure: I made 162 commits to the open source project from which I resigned almost exactly three years ago. Almost all of these were for documentation.)
That doesn't match my impression based on reading Planet Debian. The silent majority seems to be real in this case.
I think Debian is the first major project to vote to not ban LLM usage? (Or at least I can't remember any others, the ones that ban tend to be bigger news...)
And despite doing so myself, I was pretty surprised that the two primary anti-AI options got ranked below "none of the above".
I think this is because most projects aren’t having a vote. If you say nothing, people will use LLMs.
So the projects that have a vote (or a ban for projects with a BDFL) tend to be the minority that have a large anti-LLM contingent that has enough influence to make a vote/ban happen.
That said, the Kernel has no ban, and this was explicitly stated by Linus.
The official announcement with the breakdown of votes is here, but option 5 has won.
To be pedantic for a second, that is explicitly not the official announcement. It even starts with "This message is an automated, unofficial publication of vote results." The project secretary will send the official version of the results to the debian-announce mailing list. But unless something has gone incredibly wrong, they'll be the same thing.
This is disappointing. The license concerns alone, disregarding all the societal and quality concerns, are enough to disqualify LLM's in my view. Is this really the same Debian that once forked Firefox to strip all the branding out due to licensing?
TBH, there were a lot of options with heavy overlap on both the for and against sides. I don’t really envy figuring out the results because of how redundant the options were.
It is a problem on the voter's side not results side.
Results-wise, the similar options are typically ranked close, so the question is just what is the more frequent order of ranking them. For Condorcet-based counts, it is groups of options that have widely different combinations of orders that are painful (like cycles with ABC/BCA/CAB close in popularity).
But each voter has to decide which of the similar options looks better. Or to toss a coin if it is really a coin-toss.
Or toss a coin properly (toss a coin, without looking think how you hope it has landed, don't look at the coin; alternatively toss a coin and decide if you feel regret or relief, if regret, turn it over).
Now let’s all wait for a big check coming from Anthropic/OpenAI/Google/whatever and see if the “neither endorsed nor prohibited” policy makes any difference whatsoever.
Nothing out of the ordinary. Would it be a surprise if any of these companies sponsored a big open source project to later publish a cool press release like "$AI_COMPANY donates $BIG_AMOUNT to $OPEN_SOURCE_PROJECT to advance the use of AI on open source"?
If that open source project happens to be Debian itself, and they don’t object to the use of their trademark, wouldn’t that at least fall within a gray area wrt the “neither endorse nor prohibit” policy?
This shouldn’t be as controversial as the downvotes indicate and although I do acknowledge that
(i) my original comment was not a full blown opinion article on the possible pros and cons;
(ii) I don’t know jack about how does the Debian project deal with this kind of donations; and
(iii) I’m not making any distinction between money, tokens, or compute in general but, honestly, and I don’t think I should anyway,
it is very obvious to me that the policy in and of itself is good for nothing.
matheusmoreira | 8 hours ago
The filter is based on quality, and humans remain responsible for the work they submit. One hundred percent reasonable policy. Wish it was like this everywhere.
Thanks, Debian.
gspr | 22 hours ago
The results are perhaps more easily visualized as a directed graph: https://vote.debian.org/~secretary/gr_llm/
ralfj | 19 hours ago
The winner being a middle ground policy is not too surprising. I think the most surprising part is that all of the permissive-towards-LLMs options won over all the forbidding-LLMs options, and by non-trivial margins (>10% of the ballots).
That doesn't match my impression based on reading Planet Debian. The silent majority seems to be real in this case.
gspr | 19 hours ago
I'm also a bit surprised, going by the discussion on the debian-vote mailing list. I really felt the two extremes were very closely matched in popularity. With Debian's sane choice of voting system, that would still produce a sensible middle ground (like it did), but as you say the results do tilt significantly in favor of an LLM-liberal attitude. That last part surprised me.
(Disclosure: I proposed Option G – Debian is created by humans – and obviously favored it.)
viraptor | 15 hours ago
The anti side definitely seems louder in public posts. But it seems to make sense - I largely stopped trying to discuss LLMs outside of groups where people are interested in usage. Most opposing views are either not new (they may be still valid though) or are based on outdated ideas - there's no point disturbing that angry nest. But I hang out in a few groups that simply share interesting experiences/news and use the LLMs where it makes sense. I don't expect anyone from those groups to be interested in evangelizing that position, but given a vote they'd all be counted as llm-liberal.
mitsuhiko | 14 hours ago
I think the anti side wants to gain leverage right now. This result in a way is neutral: you can use it or you can opt against using it and Debian doesn’t take a stance. But for the people who are against AI this result is a win for the people who want to use it.
You can already see at least one maintainer resignation as a result of this vote.
intelfx | an hour ago
enobayram | 36 minutes ago
I'm not sure a "neutral" position is really neutral when it comes to the use of AI. I think there's an asymmetry we need to acknowledge here. It's almost impossible to work with AI-wielding people without using AI yourself. So, in a way, being AI-neutral is like being doping-neutral in a competition.
I know open source is not a competition, so it's not a perfect analogy, but as a maintainer, if you don't have the liberty to close PR as "AI;DR", then you simply can't keep up, constantly wasting your time with PRs that will take 100x the time to articulate why it's not up to par than the time it took to create it.
Saying "Humans responsible for the work they submit" without any actual punishing consequences for submitting flawed code is also completely toothless, and based on personal experience, is equivalent to saying nothing.
fuzzy | 17 hours ago
I sped through https://www.debian.org/vote/2026/vote_002#textg.
For what it's worth, if I had been eligible to vote, I would have shied away from this point in particular:
I can foresee real value from (responsible) use of AI for direct contributions in some areas.
(Disclosure: I made 162 commits to the open source project from which I resigned almost exactly three years ago. Almost all of these were for documentation.)
With respect
kghose | 15 hours ago
It’s a hopeful sign that technologists are slightly in favor of technology.
Halkcyon | 14 hours ago
It's a dire sign that the technologists don't have any morals as a majority in that community.
singpolyma | 8 hours ago
This is an outrageous thing to say about one of the most staunchly moral group of technologists on the planet.
You can disagree without stooping to this kind of attack.
sunshowers | 7 hours ago
kghose | 3 hours ago
Ouch. Anyway…
legoktm | 15 hours ago
I think Debian is the first major project to vote to not ban LLM usage? (Or at least I can't remember any others, the ones that ban tend to be bigger news...)
And despite doing so myself, I was pretty surprised that the two primary anti-AI options got ranked below "none of the above".
hyperpape | 15 hours ago
I think this is because most projects aren’t having a vote. If you say nothing, people will use LLMs.
So the projects that have a vote (or a ban for projects with a BDFL) tend to be the minority that have a large anti-LLM contingent that has enough influence to make a vote/ban happen.
That said, the Kernel has no ban, and this was explicitly stated by Linus.
fuzzy | 12 hours ago
The FreeBSD Project is currently surveying.
https://www.reddit.com/r/freebsd/comments/1vthk9s/comment/p4uwki6/
fuzzy | 12 hours ago
The FreeBSD Project is currently surveying.
https://www.reddit.com/r/freebsd/comments/1vthk9s/comment/p4uwki6/
AviKav | an hour ago
For those unfamiliar with Debian voting I found the text version an easier introduction.
https://vote.debian.org/~secretary/gr_llm/results.txt
fuzzy | 19 hours ago
I think I love you. But not as much as I love the fact that responsibility prevails.
gspr | 19 hours ago
To be clear, the page I point to is by the project secretary, Kurt Roeckx. He definitely deserves love for his tireless work :)
fuzzy | 18 hours ago
I hereby share the love. Thanks to all concerned.
I see a comparable image – https://people.debian.org/~lucas/gr-2026-002/results.png – linked from https://lists.debian.org/debian-vote/2026/08/msg00362.html with a footnote from Lucas Nussbaum,
Yogthos | a day ago
sensible
legoktm | 15 hours ago
To be pedantic for a second, that is explicitly not the official announcement. It even starts with "This message is an automated, unofficial publication of vote results." The project secretary will send the official version of the results to the debian-announce mailing list. But unless something has gone incredibly wrong, they'll be the same thing.
williewillus | an hour ago
This is disappointing. The license concerns alone, disregarding all the societal and quality concerns, are enough to disqualify LLM's in my view. Is this really the same Debian that once forked Firefox to strip all the branding out due to licensing?
calvin | 14 hours ago
TBH, there were a lot of options with heavy overlap on both the for and against sides. I don’t really envy figuring out the results because of how redundant the options were.
k749gtnc9l3w | 8 hours ago
It is a problem on the voter's side not results side.
Results-wise, the similar options are typically ranked close, so the question is just what is the more frequent order of ranking them. For Condorcet-based counts, it is groups of options that have widely different combinations of orders that are painful (like cycles with ABC/BCA/CAB close in popularity).
But each voter has to decide which of the similar options looks better. Or to toss a coin if it is really a coin-toss.
Or toss a coin properly (toss a coin, without looking think how you hope it has landed, don't look at the coin; alternatively toss a coin and decide if you feel regret or relief, if regret, turn it over).
[OP] danlamanna | 11 hours ago
Related LWN discussion https://lwn.net/Articles/1091231/.
[OP] danlamanna | a day ago
tmcb | 20 hours ago
Now let’s all wait for a big check coming from Anthropic/OpenAI/Google/whatever and see if the “neither endorsed nor prohibited” policy makes any difference whatsoever.
ralfj | 19 hours ago
What are you trying to insinuate?
tmcb | 16 hours ago
Nothing out of the ordinary. Would it be a surprise if any of these companies sponsored a big open source project to later publish a cool press release like "$AI_COMPANY donates $BIG_AMOUNT to $OPEN_SOURCE_PROJECT to advance the use of AI on open source"?
If that open source project happens to be Debian itself, and they don’t object to the use of their trademark, wouldn’t that at least fall within a gray area wrt the “neither endorse nor prohibit” policy?
This shouldn’t be as controversial as the downvotes indicate and although I do acknowledge that (i) my original comment was not a full blown opinion article on the possible pros and cons; (ii) I don’t know jack about how does the Debian project deal with this kind of donations; and (iii) I’m not making any distinction between money, tokens, or compute in general but, honestly, and I don’t think I should anyway, it is very obvious to me that the policy in and of itself is good for nothing.
nikaro | 10 hours ago
Google is listed there: https://www.debian.org/partners/
tmcb | 16 hours ago