glibc AI policy

[Andrew Pinski]

On Thu, Apr 16, 2026 at 5:33 AM Carlos O'Donell <carlos@redhat.com> wrote:
>
> On 4/15/26 7:14 AM, Florian Weimer wrote:
> > * Carlos O'Donell:
> >
> >> On 4/14/26 2:09 PM, Florian Weimer wrote:
> >>> * Carlos O'Donell:
> >>>
> >>>> On 4/13/26 8:52 AM, Florian Weimer wrote:
> >>>>> For the record, I'm not happy with requiring Assisted-by:.  If those
> >>>>> companies want us to advertise their products, that's something they
> >>>>> should pay for.  It also raises the question whether we need to track if
> >>>>> users paid models or the free version, which may have different
> >>>>> ownership terms of generated materials.  Without such information, it
> >>>>> would be misleading to use the trailers for any project-specific purpose
> >>>>> (which is why I think it's mostly about advertising).
> >>>>
> >>>> Sorry, when I reviewed your earlier patch I tried to express that I did
> >>>> not think that we needed to write the name of the actual tool used, but
> >>>> I failed to express that clearly. When I said "LLM" I meant it literally
> >>>> in that we could use "Assisted-by: LLM" and that would be sufficient
> >>>> for now.
> >>>
> >>> I can live with that.  Regarding outcomes, I would like to see a
> >>> third one: We should not fall into the trap a dehumanize
> >>> contributions because we could be wrong.
> >>
> >> May you please expand on this a bit more?
> >
> > I don't reviewers to actively look for LLM usage in non-labeled
> > contributions because it's bad to accuse someone of being an LLM when
> > they are not.
>
> I agree completely.

I actually disagree here. In fact see below.

>
> If they did not label their work "Assisted-by: LLM" then the contribution
> must be considered to come from the author.

A violation of this needs to be handled in the same way as a CoC violation.

>
> We can work to add a section to the "Contribution Checklist" and more
> generally ask new contributors to review the entire list in a
> non-judgemental way.

NO. It should very much be in a judgemental way. It should say missing
disclosing that you used a LLM is a CoC violation fullstop.

>
> > So we can say “tool-generated contributions must be labeled as such”,
> > but we can't really speculate publicly whether a particular contribution
> > was written by hand or not.
>
> What I can do is make it clear with my actions that behaviour that includes
> calling the contributor a liar is unacceptable and will be considered
> harassment under the Code of Conduct.

I disagree with respect to this. In fact the burden of proof if a
patch is NOT LLM generated should be on the submittor rather than a
reviewer mentioning this seems LLM generated.
Yes there might be cases where someone pushes too hard on that point
and yes that would be covered under the CoC. But we have already seen
in other projects that is not the case. There has been an effort in
hiding LLM usage happening.
A reviewer needs the ability to also say this has gone back and forth
and this is not going anywhere; if this is not currently covered in
general this needs to be there.

Thanks,
Andrew


>
> There might be situations where someone is identified as not being truthful,
> but I'd like those situations to be raised directly with the Code of Conduct
> Committee for review (https://sourceware.org/glibc/wiki/CoC/Committee).
>
> These concerns aren't new, but "Assisted-by: LLM" brings them to the forefront
> in a whole new light.
>
> Does that align with what you're thinking?
>
> --
> Cheers,
> Carlos.
>