On 07-Jun-26 09:04, Joel Halpern wrote:
This seems to now be looking enforce disclosure of the use of AI.
That seems extremely difficult.
Even for a participant who wants to do "the right thing", If they have
been using AI assistance during the document development and refinement
process, I doubt even they know how much of the text was AI originated,
or even whether a significant part was.
Possibly, although I think an author will know if a LLM actually
*originated* an idea. I have already marked this statement in the draft
as an open issue, and I don't think we're going to get much further
in this discussion.
And the enforcement mechanisms you suggest below seem to rely on the RPC
or the IESG being able to tell which parts of a document were generated
by AI. Which as far as I can tell is somewhere between difficult and
impossible. I would not even be surprised if, as some people read more
AI generated text they start unconsciously writing n the same style.
Not only that; LLMs are literally getting better every day, so very idea
of a distinctive LLM "style" will soon be overtaken by events.
I don't mind a note suggesting that if folks know that an AI generated
significant portions of the text in an I-D, they should include that
fact in the acknowledgements. But any step beyond that seems to be
asking for trouble. And I am doubtful of the value of that request.
I wonder if the RSWG chairs have formed a view of where rough consensus
might lie. (I know this is not a WG draft but it seems to have attracted
quite a lot of interest.)
Brin
Yours,
Joel
On 6/6/2026 4:55 PM, Brian E Carpenter wrote:
John,
On 07-Jun-26 03:16, John R Levine wrote:
On Sat, 6 Jun 2026, Brian E Carpenter wrote:
Sorry about that - My question was "do people reading text want to
know if it
was generated by a LLM?"
It looks like we're talking past each other. People reading text could
want a lot of things, and that might well be one of them. You're
proposing we create a new rule that requires people to state whether
they
used AI in their contributions.
I think the questions I asked a few messages ago are reasonable so if
you
want that rule, could you go back and address them?
Assuming this is what you mean:
Nothing in any version of the draft suggests changing anything,
except requiring disclosure. I don't really understand why that is
regarded as undesirable or onerous.
It's intrusive, vague and unenforcable. How much AI do I have to
use before I have to disclose?
The draft says "If, however, a substantial part of the document was
created by AI..."
"Created" is clear enough, but obviously "substantial" is subjective, a
matter of human judgement, and I don't see how we can ever get away from
that, as in many other judgement calls in the RFC process.
If I lie, what are the penalties?
Potentially, non-approval by the stream approving body. (There might be a
different answer in a copyright dispute, but that's not in scope here.)
Who is going to enforce it?
Potentially, the stream approving body, if the text is suspicious.
However, this is ethics, not a hard rule, so it really only works
by self-enforcement.
Do they know they've signed up to do that?
Yes, if they adopt the policy.
Regards/Ngā mihi
Brian
--
rswg mailing list -- [email protected]
To unsubscribe send an email to [email protected]
