Re: AI slop "contributions" to IETF working groups

[George Michaelson <[e-mail]>]

I guess I was more focussed on the "breach" sense that people submit
documents on a basis the work is proffered in good faith in a certain
condition and so if we're asking questions which they tick NO and we
subsequently have reason to think they were YES, the problem here isn't the
quality of the solution, the problem is that somebody was misleading.

For IPR, we have strong law issues. So, its a well dug trail road.  For AI,
we don't have well dug lines. Thats my concern because the people, the
process side of this, have to make decisions.

I don't disagree machine based methods are coming to do validation, checks,
maybe even code. I don't like this when it's referred to the giant
statistical machine-in-the-sky but we've had "this schema has been checked
by xxx" for a long time now. I can live with post-hoc review and checks by
machines.

What concerns me, is what we do when we ask people to say something, and
then discover they didn't answer truthfully.

In emails, I've already seen people make a reasoned case that for EFL
speakers there is a role for systems which do better than their otherwise
poor english skills. I don't personally believe we're ready to be tolerant
of that in normative language contexts, and I am a lot more confident we
don't want to be led to critique which has no human reasoning in it, just a
machine review.

-G (No Hats)

On Tue, Feb 10, 2026 at 11:18 AM Brian E Carpenter <
[e-mail]> wrote:

> On 10-Feb-26 16:31, George Michaelson wrote:
> > Not to disagree with anything said, what would people want from our
> mechanistic processes when a breach is discovered, especially when it's
> discovered late in process like during AUTHOR48
>
> George, if our WG Last Call/Shepherd review/IETF Last Call/IESG review
> sequence doesn't find an AI-induced error, I'm not sure we should blame the
> AI. I don't see any difference from human error, by the time we send a
> draft off to the RFC people. All I'm saying is that it should be
> acknowledged up front and verified by a human before posting.
>
> In any case, AI isn't automatically bad. For example, I could imagine
> designing a protocol and then asking an AI to look for race conditions**.
> As others have said, there are many valid uses of AI.
>
> ** We probably aren't quite there yet, but see:
> https://chatgpt.com/share/698aaede-0438-8002-bc1f-a3b3b7411c52
>
> So having an AI verify our designs is probably coming Real Soon Now.
>
>      Brian
>
> >
> > If you want specific outcomes like a stall, or a reject then it's going
> to need careful definition and probably implies other process cost like
> appeals.
> >
> > G (no hats)
> >
> > On Tue, 10 Feb 2026, 10:28 Cheng Li, <c.l=[e-mail]
> <mailto:[e-mail]>> wrote:
> >
> >
> >     Agree with Bob and Brian that we need to request a disclosure of
> using AI.
> >
> >     I do not like to read an AI generated document, especially, without
> human checking, feeling unprofessional and uncomfortable.
> >
> >
> >
> >     Thanks,
> >     Cheng
> >
> >
> >
> >
>  ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
> >
> >     李呈 Li Cheng
> >     Mobile: [phone](中国电话)
> >     Mail: [e-mail] <mailto:[e-mail]>
> >
> >     *发件人：*Brian E Carpenter <[e-mail] <mailto:
> [e-mail]>>
> >     *收件人：*[e-mail] <mailto:[e-mail]> <[e-mail] <mailto:
> [e-mail]>>
> >     *时 间：*2026-02-10 04:14:21
> >     *主 题：*Re: AI slop "contributions" to IETF working groups
> >
> >     On 08-Feb-26 14:55, John Levine wrote:
> >     > It appears that Bob Hinden  <[e-mail] <mailto:
> [e-mail]>> said:
> >     >> I like the idea of requiring a disclosure that an Internet Draft
> was written with AI tools.  I agree probably not for emails.
> >     >
> >     > I expect that the people whose drafts most suffer from being
> written by an LLM will lie about it, either because
> >     > their English is too poor to understand the rules, or they have
> perverse incentives like an an employer that
> >     > gives them a bonus for publishing an RFC.
> >     >
> >     > Then what?  I'm not saying we shouldn't ask, but I'm saying we
> need to be prepared for noncompliance.
> >
> >     Of course, but if we make it a submission requirement - check the
> "AI used" box - there's really no harm, no shame.
> >
> >     Actually there are perhaps two boxes to check:
> >
> >     1. Was AI used in the production of this draft? [yes/no]
> >     2. If yes, did the authors personally verify the AI-generated
> material? [yes/no]
> >
> >     The attachment was generated by ChatGPT.
> >
> >          Brian
> >
>