Re: Dealing with LLMs in IETF discussions draft

[Stephen Farrell <[e-mail]>]

Hiya,

On 31/07/2026 02:59, George Michaelson wrote:
> The discussion is important but something about the shape of the dialogue
> as proffered didn't work for me.

As-is, the draft really just provides two probably-conflicting
perspectives with a bit of an attempt to see if those could end
up producing a common set of recommendations.

> I think it should be a discussion. I think perhaps the way to approach it
> is a discussion about a ML or WG charter exploring issues.

Yep. I think a mailing list would be the right thing. I do
worry a bit that such a list would only have the same ten
people who care about IETF administrivia. (Or that plus a
pile of AI enthusiasts, not sure which'd be worse:-)

> I feel (as an example) that there is a fundamental lack of honesty
> preferring synthetic input as discourse in a standards directed document,
> where I do not feel the same problem with a protocol spec or data structure
> in some notation which was machine generated, if (for example) i can see
> the prompt. I find declared use of translation agents more acceptable than
> wall of text with no proffer of source.

I'd start from close to there too. OTOH, the translation
thing may be a real win for those not fluent in English so
I don't think we can ignore that. (And laziness too;-)

> My concerns are not entirely the same as the ones you put in this draft.
> Hence my sense the intent is good, but I am unsure this is the form.

Again yep. A mailing list to thrash things out seems like
the right idea. We figured a draft to bash at might help
with that.

Cheers,
S.

> No hat, but as an observation the LLC Board (of which I am a member) has
> asked participants in open board meetings not to bring AI adjuncts with
> them into the conversation or use their note taking and transcripts as a
> record.
> 
> George
>