Re: Dealing with LLMs in IETF discussions draft

--00000000000042b7060657df19f2
Content-Type: text/plain; charset="UTF-8"

Hi,

This is a really important topic and an incredibly complex one. I think the
document already does a good job outlining both: a) concerns that apply
well beyond and directly within the work the IETF does, and b) the
beneficial impact of using AI assists for translation, triage and
organizing thoughts and ideas.

The preliminary conclusion that "humans originate" while "AI executes" is a
fine ideal. I think the IETF has a number of innate self-defense
mechanisms that may help even though AI slop will surely be a persistence
nuisance. That said, I think those self-defense mechanisms might need some
form of bolstering, hence why I think the idea of this draft is a good idea.

A few of those self-defence mechanisms:

- The IETF holds 3 large meetings per year -- and agentic bots don't yet
have access to holographic projection :)
- Because AI cannot reason, a few rounds with a bot without a human in the
loop makes it pretty clear that there's no further technical substance to
extract (if any)
- I've observed that publishing an RFC to the standards track is one that
has a tremendous amount of diverse oversight

On major new risk that I didn't see discussed is how easily an agentic bot
can be persuaded to align with a particular frame of argument in its
pursuit to be agreeable with whoever most recently spoke with it. You don't
even need fancy tricks when you can just apply the right phrasing to
manipulate their "Reinforcement Learning from Human Feedback" (RLHF) loops.

Given that AI can't originate novel technical insight (the way a human
can), and as the IETF aims to count "issues" instead of "noses", weeding
out these kinds of 'participants' is possible. But no doubt these agentic
systems will present ongoing concerns against the consensus process.

All that to say, my guess is that longer-term solutions will require the
IETF will need to find ways to apply increased vigilance without turning
away newcomers, so a new mailing list sounds like a good start. As long as
that mailing list isn't used to name and shame, which is a whole other risk.

Cheers,
Nathanael



On Thu, 30 Jul 2026 at 19:21, Stephen Farrell <stephen.farrell=
[email protected]> wrote:

>
> Hi all,
>
> Myself and Chong Feng have posted an I-D [1] that (from two
> pretty different perspectives) aims to progress discussion
> of how we might better handle LLM text in IETF discussions.
>
> I think this'd be a fine topic for broad discussion. I do
> not know what would be the correct mailing list for such a
> discussion, but also think it doesn't warrant waiting for
> another round of dispatchery before being discussed. (Nor
> would any dispatch outcome above 'new mailing list' suit.)
>
> So, I'd be v. interested in comments on this and/or on it
> being moved to some other mailing list for discussion. If
> no existing list suits, I'd suggest [email protected]
> (which could be contracted to [email protected]:-) as a name
> for a new list, but whatever works is fine.
>
> Meanwhile, we'd be v. interested in any comments here or
> on the github repo [2].
>
> And of course, this is a v. rough thing, not intended to
> be prescriptive or authoritative, so have at it:-)
>
> Cheers,
> S.
>
> [1] https://datatracker.ietf.org/doc/draft-fengfar-led/
> [2] https://github.com/sftcd/led/
>
>

--00000000000042b7060657df19f2
Content-Type: text/html; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable

<div dir=3D"ltr">Hi,<br><br>This is a really important topic and an incredi=
bly complex one. I think the document already does a good job outlining bot=
h: a) concerns that apply well beyond and directly within the work the IETF=
 does, and b) the beneficial impact of using AI assists for translation, tr=
iage and organizing thoughts=C2=A0and ideas.=C2=A0<div><br></div><div>The p=
reliminary conclusion that &quot;humans originate&quot; while &quot;AI exec=
utes&quot; is a fine ideal. I think the IETF has a number of innate self-de=
fense mechanisms=C2=A0that may help even though AI slop will surely be a pe=
rsistence nuisance. That said, I think those self-defense mechanisms might =
need some form of bolstering, hence why I think the idea of this draft is a=
 good idea.<br><br>A few of those self-defence mechanisms:=C2=A0<br><br>- T=
he IETF holds 3 large meetings per year -- and agentic bots don&#39;t yet h=
ave access to holographic=C2=A0projection :)<br>- Because AI cannot reason,=
 a few rounds with a bot without a human in the loop makes it pretty clear =
that there&#39;s no further technical substance to extract (if any)</div><d=
iv>- I&#39;ve observed that publishing an RFC to the standards track is one=
 that has a tremendous amount of diverse oversight<br><br>On major new risk=
 that I didn&#39;t see discussed is how easily an agentic bot can be persua=
ded to align with a particular frame of argument in its pursuit to be agree=
able with whoever most recently spoke with it. You don&#39;t even need fanc=
y tricks when you can just apply the right phrasing to manipulate their &qu=
ot;Reinforcement Learning from Human Feedback&quot; (RLHF) loops.</div><div=
><br></div><div>Given that AI can&#39;t originate novel technical insight (=
the way a human can), and as the IETF aims to count &quot;issues&quot; inst=
ead of &quot;noses&quot;, weeding out these kinds of &#39;participants&#39;=
 is possible. But no doubt these agentic systems will present ongoing conce=
rns against the consensus process.=C2=A0</div><div><br></div><div>All that =
to say, my guess is that longer-term solutions will require the IETF will n=
eed to find ways to apply increased vigilance without turning away newcomer=
s, so a new mailing list sounds like a good start. As long as that mailing =
list isn&#39;t used to name and shame, which is a whole other risk.</div><d=
iv><br></div><div>Cheers,</div><div>Nathanael<br><div><br></div><div><br></=
div></div></div><br><div class=3D"gmail_quote gmail_quote_container"><div d=
ir=3D"ltr" class=3D"gmail_attr">On Thu, 30 Jul 2026 at 19:21, Stephen Farre=
ll &lt;stephen.farrell=3D<a href=3D"mailto:[email protected]">40cs=
[email protected]</a>&gt; wrote:<br></div><blockquote class=3D"gmail_q=
uote" style=3D"margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,2=
04);padding-left:1ex"><br>
Hi all,<br>
<br>
Myself and Chong Feng have posted an I-D [1] that (from two<br>
pretty different perspectives) aims to progress discussion<br>
of how we might better handle LLM text in IETF discussions.<br>
<br>
I think this&#39;d be a fine topic for broad discussion. I do<br>
not know what would be the correct mailing list for such a<br>
discussion, but also think it doesn&#39;t warrant waiting for<br>
another round of dispatchery before being discussed. (Nor<br>
would any dispatch outcome above &#39;new mailing list&#39; suit.)<br>
<br>
So, I&#39;d be v. interested in comments on this and/or on it<br>
being moved to some other mailing list for discussion. If<br>
no existing list suits, I&#39;d suggest <a href=3D"mailto:humanorwhat@ietf.=
org" target=3D"_blank">[email protected]</a><br>
(which could be contracted to [email protected]:-) as a name<br>
for a new list, but whatever works is fine.<br>
<br>
Meanwhile, we&#39;d be v. interested in any comments here or<br>
on the github repo [2].<br>
<br>
And of course, this is a v. rough thing, not intended to<br>
be prescriptive or authoritative, so have at it:-)<br>
<br>
Cheers,<br>
S.<br>
<br>
[1] <a href=3D"https://datatracker.ietf.org/doc/draft-fengfar-led/" rel=3D"=
noreferrer" target=3D"_blank">https://datatracker.ietf.org/doc/draft-fengfa=
r-led/</a><br>
[2] <a href=3D"https://github.com/sftcd/led/" rel=3D"noreferrer" target=3D"=
_blank">https://github.com/sftcd/led/</a><br>
<br>
</blockquote></div>

--00000000000042b7060657df19f2--