Re: Participation using AI in WG mailing lists and github repos

[Rohan Mahy <[e-mail]>]

Hi Mark,

> More concerningly, AI also makes it possible to participate with very
little (human) effort or understanding. At the extreme, one can point an
agent at a list, give it instructions, and step away. I (and others I've
talked to) believe we've already seen a few examples of this in the IETF.

That appears to be specifically the case. The account replies within
moments on every comment, Issue, or PR on our github repo, even at odd
hours in the timezone of the human.

> 1. Disclose AI use unless you've personally verified it and would stand
behind every word with your reputation.

I don't think that is quite enough. It implies that it is ok to generate a
deluge of AI content as long as you disclose that it is coming from an AI.
That's a DoS on the IETF even if at the bottom of a wall of text there is a
"Generated by Claude".
How about this:

*Human review*
It is never OK to have an AI agent reply to github or the mailing list
automatically. You need to personally read, understand and approve every
message you send to an IETF list. You need to personally, read, understand,
and approve every PR, Issue, comment, or reaction you submit to an IETF
GitHub repository.

*Disclosure*
Every new mailing list post on a new topic, and each new github PR or Issue
should include the following disclosure and assertion:
"This contribution was originally generated by <Model>[<version>]. I
personally reviewed every word."

> 2. It's OK to tell someone you're not reading what they wrote because it
appears to be undisclosed AI-generated content produced without sufficient
human guidance.

I'm ok with that, but I think we should come up with a way to say this
kindly. Perhaps this might be a place where we could ask our ombuds team to
suggest some appropriate text?

> The other thing we could do would be to work on an IETF-specific
CLAUDE.md / AGENTS.md file -- i.e., norms for the *agents* too.

I think writing the norms for agents would be excellent.

We should probably say some places where we think people can safely use AI
to participant, otherwise everyone will just ignore us:

   - It is fine to use AI to translate and/or summarize part of a
   discussion. It might be useful to ask the AI to suggest other possible
   meanings or conclusions.
   - It is fine to use AI to explain a machine-readable language (ex: "What
   does this ABNF or YANG module do?")
   - It is OK to use AI to translate a message from your language into
   English. You should read the sentence afterwards and make sure you
   understand it in English before you send it. Consider including the text in
   the original language, so another native speaker of your language can see
   if there is a discrepancy or a better translation. Make your sentences
   direct, and concise. Keep your post short. If possible use a small number
   of short examples that show the concept you are trying to convey.

Thanks,
-rohan


On Sun, Jun 7, 2026 at 5:58 AM Mark Nottingham <[e-mail]> wrote:

> Hi Rohan,
>
> Thanks for bringing this up; I've been thinking about doing so for a
> while. I trust you're aware of the discussion in RSWG about AI and
> authorship; I see this as a complementary but distinct discussion.
>
> My .02 -
>
> AI can certainly be seen as an assistive technology that levels the
> playing field for those who have language or other barriers to
> participation. Anything we do in this area should carefully consider impact
> there.
>
> More concerningly, AI also makes it possible to participate with very
> little (human) effort or understanding. At the extreme, one can point an
> agent at a list, give it instructions, and step away. I (and others I've
> talked to) believe we've already seen a few examples of this in the IETF.
>
> I'm concerned because the IETF's value as a community lies in its
> collective expertise, and that's a finite resource. Historically,
> participation has been gated on making enough effort to constructively
> participate -- understanding norms, reading drafts, discussing things with
> people, etc. That natural constraint on participants limited the amount of
> resources required to engage with newcomers, filter bad ideas from good,
> respond to questions, and so forth.
>
> AI, when used in certain ways, can flip that relationship: it becomes very
> easy to sketch out an idea, have AI develop it, and then use the IETF's
> interaction mechanisms to refine it by asking a series of persistent
> questions. It gets worse when the AI doesn't have any sense of the norms of
> discourse here and does things like being overly verbose, failing to
> concede points, and even getting aggressive (all of which we've seen
> examples of).
>
> If our capacity to take on new participants and ideas were infinite, this
> would be fine, but it's not; we have real capacity constraints in the
> organisation, measured in things like people's willingness to review
> drafts, read e-mails, engage with questions, and of course meeting time.
> The pool of experts that give the IETF its reputation -- and legitimacy --
> is not infinitely scalable, and needing to wade through too much slop is
> going to be an active disincentive for them.
>
> Obviously, how bad this gets depends on how prevalent these patterns of
> use become. Personally, I think we're early days here, and it's already
> been very distracting in some of the groups I participate in.
>
> What can we do about it? As a first step, I'd like us to consider
> introducing new norms:
>
> 1. Disclose AI use unless you've personally verified it and would stand
> behind every word with your reputation.
>
> 2. It's OK to tell someone you're not reading what they wrote because it
> appears to be undisclosed AI-generated content produced without sufficient
> human guidance.
>
> I suspect #2 might be controversial, and we'd need a test to see how it
> works. I think it's necessary, because the alternative is reducing our
> openness as a community and that's a much bigger change.
>
> The other thing we could do would be to work on an IETF-specific CLAUDE.md
> / AGENTS.md file -- i.e., norms for the *agents* too. Eg it could start by
> encoding the norms we have for human participants on the mailing list (stay
> within charter, assume a highly technical audience, be concise, respect
> chair decisions, don't re-litigate settled questions).
>
> Cheers,
>
>
> > On 6 Jun 2026, at 12:09 pm, Rohan Mahy <[e-mail]> wrote:
> >
> > Hi,
> > Do we have a policy that specifically mentions replying to WG mailing
> lists and/or github repos with obviously AI-generated content of
> questionable relevance?  I know we have tools for managing people who are
> disruptive, but the amplification using AI is more the issue here. The
> account in question does have a human user.
> >
> > Thanks,
> > -rohan
>
> --
> Mark Nottingham   https://www.mnot.net/
>
>