Re: [Discussion] AI-generated content within Debian

Aryan Karamtoth <[email protected]> Thu, 19 Feb 2026 09:41:13 +0530
Newsgroups	gmane.linux.debian.devel.project
Message-ID	<[email protected]>
On 19/02/26 05:38, Pierre-Elliott Bécue wrote:
> (I wrote this markdown style, and I'm too lazy to convert it to text)
>
> # The infamous LLM discussion
>
> So, I'm starting this discussion publicly because a heated discussion
> started privately, and this is no private topic. The discussion started
> because of the new DFSG team's NEW queue website, which has been (to
> some extent I don't personally know) developed with the assistance of an
> agentic coding tool.
>
> I'd like to summarize where we all collectively are, where Debian is
> currently, and the different pros/cons/arguments I read and heard in the
> past two years. This obviously won't be exhaustive, it's a starting
> point.
>
> This is not an opinionated post, I am in an uncomfortable cognitive
> dissonance on the matter, so it's rather a snapshot of my brain on the
> topic.
>
> To be frank, I personally don't know where I stand. I think I'm neither
> for nor against AI-generated code, but I am aware that currently, it's
> not possible to give a simple and trivial ruling. If some specific
> questions worth an answer are asked, I'll reply, but otherwise I have
> the very intent to not post after this mail. The topic, its
> ethics/sociologic/technologic ramification is exhausting, and I'd rather
> spend my time doing funny stuff.
>
> I might at some point in this text (I'm writing it linearly so I don't
> know how much time I'll take to write it and what the end will look
> like) offer an idea of a policy on the matter. But don't expect from me
> to say if it's a good idea or not.
>
> I do ask everybody, DDs, DMs, DCs, bystanders, to refrain from flaming.
> I know this wish has little chances to be successful, but at least I
> will have tried.
>
I raised this point earlier in a mailing list and would like to bring it 
up again:

For a non-profit Linux distribution whose main focus is stability and 
security and with a release philosophy like "we ship when its ready", I 
don't think LLMs go hand-in-hand with Debian's vision.

LLMs are sure useful for certain tasks but the problem is that everyone 
fell for the corporate marketing saying it can replace substantial 
amount of human effort needed in software development. They're not the 
jack of all trades that the corporates claim them to be.

Sure, maintaining a distro as huge as Debian takes a lot of time and 
effort but we as volunteers, agreed to devote some time towards the 
project to keep the project in a good shape. We aren't working in a 9-5 
job forcibly with strict deadlines and targets to offload our work to a 
LLM.

We're also forgetting about the hallucination problem that's still 
persistent with LLMs. I used ChatGPT back when it first came out in 2022 
and often use Google Gemini in 2026 here and there for some personal 
tasks and the hallucination problem still persists. These tools are 
still impractical and flawed for any serious work unless you're fine 
with low quality slop code.

We're also forgetting about the licensing. Most of the training material 
that was used to train LLMs was illegally scraped by violating copyright 
laws. When we generate a code snippet using a LLM, we have no clue what 
licensed code it used to generate the code.

-- 
Regards,

Aryan Karamtoth,
Debian Maintainer

IRC: SpaciousCoder78
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, 6.5 KB)
OpenPGP_signature.asc (application/pgp-signature, 840 B)