Re: Dealing with LLMs in IETF discussions draft

Mikael Abrahamsson <[email protected]> Fri, 31 Jul 2026 13:17:01 +0200 (CEST)
Newsgroups gmane.ietf.general
Organization People's Front Against WWW
Message-ID <[email protected]>
On Fri, 31 Jul 2026, Stephen Farrell wrote:

> I think this'd be a fine topic for broad discussion. I do

Yes, I 100% agree.

I've used AI lately to perform document reviews and to sum up what might 
have happened leading up to the document. It's great at finding typos. 
It's good at correlating. It sometimes guesses (quite human actually). One 
document I asked it to perform an analysis on merging two drafts. It spent 
30 minutes summing up email discussions, read both drafts, summed up 
possible ways to merge them, and produced an overall good overview of the 
situation.

I use AI daily in my work to perform some tasks. It's like any tool, 
everything from amazingly good, to huge... errr.. disappointment.

What I think the IETF should encourage draft authors to use AI to find 
editorial problems. Take the draft, run it through an LLM, and ask it to 
find typos, reference problems etc. It's excellent at this. Perhaps IETF 
should even have this as part of the submission process, that before a new 
version is submitted, it's run through an editorial AI review. We use it 
at work as part of git workflow, have an AI comment on the MR diff. 
Sometimes it finds things. Sometimes it's meh. Still overall net positive 
in my experience.

What I think is the real problem, is someone taking up everybodys time 
with obvious AI generated wall of text of unfinished "solution". This has 
happened historically with human-only input, but it's become so much 
easier with the AI tools. One good example is all the "new IP protocol 
that's better than IPv6 because it's backwards compatible" when it's 
really not. I think we've had 4 of those this year, from reading it, all 
of them heavily AI generated without any notice that this was the case.

Overall, I agree with what's in the document. As with any tool, it can be 
used for good or bad. It's a tool, being used by a human.

Just as we have CoC for abusive behavior, we need to have CoC for AI use, 
and people who misuse the tools need to be told as much, and perhaps even 
blocked from participating.

-- 
Mikael Abrahamsson    email: [email protected]