Discussion
So Debian adopted a resolution on the use of “generative AI”:
https://www.debian.org/vote/2026/vote_002#texte
Condorcet spoke: of all the options put to the vote, it must be a good compromise, an average of where the community stands.
How much of a compromise is it really, though?
🧵
Their amoral contributors have voted to deliberately ensloppify.
The message of this resolution is that it’s neither “pro” nor “anti”: it’s neutral, promoting “responsible use”.
In fact, it looks like an example of “extreme centrism”: “Debian neither endorses nor prohibits the use of generative AI”, “Contributors are expected to exercise appropriate care”, “[t]his resolution therefore affirms that generative AI is neither exempt from nor subject to special rules”.
The narrative seems to be: “AI” is just another tool, nothing changes here—move along.
Under the guise of neutrality, what it actually does is open the gates to genAI-produced contributions. This, is political.
The “responsible use” narrative shifts responsibility from the group to the individual. It’s not unlike the tobacco industry’s strategy—or indeed that of the “AI” industry.
A funny example: the text recognizes that “the legal status of material produced by generative AI systems remains the subject of ongoing discussion in many jurisdictions” (a euphemism) and then states that Debian will “rely on the judgment and responsibility of its individual contributors”—as if individual contributors would know better than lawyers in this domain.
Disregard of legal issues is more of a detail though.
While the text states that “such tools can substantially improve the productivity of contributors” (which is a subject of debate), what’s telling is what the text doesn’t even mention: the “externalities”, the environmental and social impact of these services.
And that’s probably the crux of the split among free software people on this matter: are our development efforts made in a vacuum?
It’s not new: whether, for instance, to use BitKeeper for Linux development led to similar debates.
The difference here is that LLMs are not even “just a tool” in the sense BK was: the lock-in is going to be much deeper, in our organizations and in our brains. LLMs threaten the ideal of knowledge sharing and the social fabric we’ve built.
I’m not a Debian contributor and not a user anymore but I’ve always viewed Debian as a lighthouse for free software and for collective self-determination.
I’m afraid about the signal this vote is sending to the free software community at large; I’m afraid of it moving the Overton window further towards disregard for humans and for the climate change emergency, in a selfish and myopic way that we’ll regret.
Faint hope ✊:
The blunt, unsatisfactory decision that has emerged from #Debian is sorta least common denominator that means something to someone but something else to somebody else... Yet maybe an attempt at a statement: "_irresponsible_ or _non-reponsible_ use of #LLM/#GenAi is not acceptable."
Now would be time for #Debian and #FLOSS communities at large to lead the way and determine: what would be "_responsible use of #GenAI_"?
Is the use of LLM/GenAI "responsible" if:
- it boils rivers?
- it belongs to evil US tech overlords?
- its model is proprietary?
- its training dataset is not public? (then can anything from it be Free/Libre software?)
I actually put the last bit in parenthesis, in relationship to training data... but in all truth i think this is maybe _the_ most fundamental question Free/Libre #FLOSS software users, authors, supporters and overall communities will need an answer to, in relation to #LLM/#GenAI:
If some code is output by [machine], can one:
- claim to be the sole author of it and chose which license to apply to it?
- declare it free/libre software, IF one cannot prove that it is NOT based, at least in part (training data), on other software that may be released by its authors under incompatible license, or plainly proprietary?
Note: i am not trying to make it impossible through argumentation.. I could get convinced that that free/libre LLM models, available along with their training dataset (so one could potentially re-train them and to some extent get some reproducibility of the model?), ran on one own's infrastructure (or community-based, mutualised, etc. ie. not owned by some evil billionaire), with a focus on energy-efficiency, frugality maybe, could indeed be used for good, and provide quite some services with some menial/repetitive tasks (boiler plate, transcription, tests, etc.) and others, under strict validation/audit by humans, in free/libre software dev.
Whereas I do respect your approach, and I am sure you try to bring sensical points of view
- no doubt about that in general
But there is this other detail that breaks the constructive flavour of your criticism: "it boils rivers?"
If you were serious, I think you would have indicated some sort of energy cost scale, something more open to discussion than "boiling rivers"
My 2 cents
Install bonfire.cafe
Get the full app experience
