Re: On rejecting LLM-based contributions (Was: On publicly shaming people)

Aryan Karamtoth <[email protected]> Thu, 15 Jan 2026 22:34:33 +0530
Newsgroups	gmane.linux.debian.devel.project
Message-ID	<[email protected]>
On 15/01/26 8:16 pm, Lucas Nussbaum wrote:
> On 15/01/26 at 00:49 +0100, Matthias Geiger wrote:
>> IMO it's high time we had a GR banning so-called AI (read LLMs) at the very
>> least for project contributions, and I would suggest just to ignore mail
>> that looks like it was written by a machine in the meantime.
> On the other hand, since this discussion and GR did not happen yet, I
> think that it's unfair to assume that the result would be the rejection
> of AI-assisted contributions. I suspect that things would be a bit more
> nuanced.
>
> Two data points :
>
> 1/ Back in June I volunteered to explore if we could get AI companies to
> sponsor Debian with LLM access (for Debian development). This went
> nowhere so far, unfortunately, but several DDs contacted me privately
> (for example during DebConf) to express interest. So clearly some DDs
> think that AI-assisted contributions are worth exploring in the context
> of Debian.
>
> 2/ Most of my personal Free Software contributions over the last year
> have been AI-assisted to some degree. Usually that degree is that I let
> the agent write an initial version of the code, and then I review it,
> and then either rework it manually or prompt the agent to rework it until
> it matches my expectations, or a mix of both. When they were
> contributions to projects where I don't consider myself the maintainer,
> I think I was always clear that the contributions were AI-assisted.
> My feeling is that contributing that way made me produce more and in a
> more interesting way, because I was able to focus more on what the code
> achieves, than on not-so-interesting implementation details. Also it
> makes addressing boring bugs in legacy code a bit more fun.
> So if Debian had a policy of banning AI-assisted contributions, I would
> probably just decide to contribute elsewhere.
>
> I fully recognize that the AI world is far from perfect, and raises many
> questions. There's some moderate hope with open weight models and free
> frontends. Things are moving very fast, so it's hard to predict what
> will happen over the next months.
>
> In the end it's again a question of where we draw line.
> - Should we make a difference between AI-assisted and AI-generated
>    contributions?
> - What about AI-assisted contributions to upstream projects?
> - If it's about non-free tools, what about bug reports generated by
>    services such as Coverity Scan?
> - if it's about environmental impact, what about large-scale QA checks
>    and CI?
>
> Lucas
>
I think one thing that people often forget is that you cannot hold an 
LLM accountable if anything breaks. If a human makes some kind of 
mistake in the code and someone points it out, the person can easily 
recall the steps he has done, discuss with the others and fix the bug.

But LLMs very often go into hallucinating loops.

Overtime, there will be a situation where humans cannot understand the 
code written by LLMs and one day the LLM breaks something and it falls 
into a hallucination loop. The result? a critical broken piece of 
software that's hard for anybody to fix.

I agree with Matthias. Debian isn't a corporate company to care about 
implementing LLM usage across the organisation to make shareholders 
happy. We're a volunteer organisation.

-- 
Regards,

Aryan Karamtoth,
Debian Maintainer

Homepage: https://spaceports.in
Matrix: @SpaciousCoder78:matrix.org
XMPP:[email protected]

GPG Fingerprint: 7A7D 9308 2BD1 9BAF A83B 7E34 FE90 07B8 ED64 0421
OpenPGP_0xFE9007B8ED640421.asc (application/pgp-keys, 3.9 KB)
OpenPGP_signature.asc (application/pgp-signature, 840 B)