Re: Thoughts about LLM usage

Adrian Bunk dijo [Wed, Aug 12, 2026 at 12:52:35AM +0300]:
>> You seem to be saying that a maintainer would not be allowed to ask
>> the question, "Did you use AI", and I don't think option 5 disallows
>> the asking of that question.  The contributor might or might not
>> answer the question, but there's also nothing that says that a
>> maintainer is forced to accept a contribution if someone answer the
>> question one way or another --- or not at all.
>>...
>
>It is problematic when "or not at all" is a disclosure.
>
>When a potential employer in Germany asks an applicant in a job
>interview an impermissible question (e.g. about pregnancy, religion,
>sexual orientation or previous convictions), the job applicant has
>the legal right to tell a lie.

The GR is a position statement and an expectation of behavior, but it is
not a legal document. For what is worth, nothing really binds us to the
Debian Consitution other than our own good will. I would at least be _very_
surprised if we were forbidden to ask our colleagues if they used a given
tool.

I read the requirements being discussed in this GR as having a way to
_publicly disclose_ LLM usage when committing or proposing a contribution.

>Can Debian force GSoC contributors to disclose AI assistance when
>a GR states that we "do not require them to do so"?

GSoC contributors are to be checked both by their mentors and by the GSoC
program. And they have to comply with rules set by both.

>(...)
>~ 99% of my uploads are NMUs.
>Would it be OK today to upload NMUs without disclosing AI usage?

FWIW I agree with you using an LLM to prepare an NMU — and then
_double-checking a full diff_, and becoming _personally_ responsible for
any breakage it introduces.

We should be able to use whatever tools we can, but whatever goes in the
distribution signed by [email protected] is _my sole responsibility_, and I
cannot point fingers at any LLM for doing something wrong.