Banning LLM-generated or assisted contributions to the GNU C Library

[Siddhesh Poyarekar]

On 2025-08-25 13:59, DJ Delorie wrote:
> "Andreas K. Huettel" <dilfridge@gentoo.org> writes:
>> And for programming that, AI technology is not ready yet.
> 
> I'm not a fan of AI but it's a tool, not a religion.  It has its place,
> and there are things it can do easier than we can.  I see no reason to
> not treat AI as any other new contributor - we should guide it and
> mentor it, but not trust it until it has earned that trust.  Let it do
> simple or mechanical changes at first, and build from there.  GLibC was
> "not ready" once too.

I hate to say this too, but I agree, banning a tool seems like 
overstepping.  I'm more in favour of clearly identifying AI-generated 
content so that...

> However, I'm also concerned about the copyright sitation wrt AI.
> Perhaps we could limit AI to non-copyrightable (i.e. small, trivial)
> changes?  Stuff like changing an internal API, updating copyright (yes I
> know we have a script for it, these are examples ;), etc.  Stuff that is
> possibly in the "trivial patches" category.

... we make sure that it does not make any legally significant 
changes[1].  That said though, I'd expect the GNU maintainers or FSF to 
add a legal perspective here for the community to follow.

> Otherwise, I'm inclined to trust the submitter to choose when to use AI
> and when not to; using AI as a starting point for a complex change could
> be a huge win, but the patch is still the responsibility of the
> submitter.  If a developer uses AI and fails their review, that's their
> problem (and their reputation), but let's not fail/stop them *because*
> of the AI.  It's the results that count, not the process.

However if a developer uses AI to even start a large change and then 
makes prodigious changes to the result before submitting, who owns the 
copyright for that change?  This is probably where things get hairy.

> Also, I would not be opposed to using AI on the review side - perhaps as
> a "note only" reviewer, or maybe reject-only, but it would not count as
> a "human reviewer" for approving a patch.  More eyes on a patch are
> good, even if those eyes are AI's.

Sure, but more like a non-blocking vote, even for rejections.

Thanks,
Sid

[1] https://www.gnu.org/prep/maintain/html_node/Legally-Significant.html