Re: Dealing with LLMs in IETF discussions draft

On Jul 31, 2026, at 7:59=E2=80=AFAM, Vittorio Bertola =
<[email protected]> wrote:
>=20
>> Il 31/07/2026 03:20 CEST Stephen Farrell =
<[email protected]> ha scritto:
>>=20
>> Meanwhile, we'd be v. interested in any comments here or
>> on the github repo [2].
>=20
> My first comment after a quick read: the draft only marginally touches =
what I see as the main current problem created by LLM users, that is, =
overwhelming non-AI-assisted participants with overlong drafts and list =
messages that they can't keep up with. We just had an example in DNSOP, =
which in the end forced the chairs to moderate away the author, as a lot =
of people could not keep up with his (its?) overlong and over-frequent =
messages.
>=20
> Transparency on the use of AI will not solve this problem. Perhaps, =
things would become more manageable if all participants used AI to parse =
the messages, but not everyone can do that today; such a factual =
requirement would introduce discriminations among participants depending =
on:
> - level of skill in this specific technology
> - availability of time and resources to set up AI assistance for IETF =
work
> - availability of money to pay for significant model usage (like =
parsing and summarizing all messages from high volume mailing lists)
> - availability and quality of LLM/AI services and software in their =
country of origin

Maybe it will be different in some years, but for now, I think we want =
discussions to be *human communication*. When you write, you write for =
another human. Your job in writing is to convey ideas effectively to =
peers whom you assume are reading directly without AI assistance.

Even without AI, people write messages that are far too long and very =
suboptimal for communicating ideas to peers. Sometimes I push back: =E2=80=
=9CI didn=E2=80=99t read that because it was too long=E2=80=9D. I=E2=80=99=
ve seen chairs push back too.

So maybe one of the principles here is: IETF is for humans?

LL