Re: AI disclosure [was: AI slop "contributions" to IETF working groups]

[Brian E Carpenter <[e-mail]>]

On 16/02/2026 15:20, Christian Huitema wrote:
> 
> On 2/15/2026 11:28 AM, Brian E Carpenter wrote:
>> On 16/02/2026 07:53, Michael Richardson wrote:
>>>
>>> John Levine <[e-mail]> wrote:
>>>       > With copyrights, unlike patents, if you create something
>>> similar to another
>>>       > work, but independently of the other work, that is OK. Back in
>>> the PC era there
>>>       > were elaborate "clean rooms" used to write PC BIOS code
>>> equivalent to the IBM
>>>       > PC's but without looking at the original. But if an LLM is
>>> trained on something,
>>>       > and you use output from that LLM, is that enough of a
>>> connection? How much do we
>>>       > want to spend to find out?
>>>
>>> How much do we want to risk if things turn out poorly for the LLMs?
>>> It seems to me that the lowest risk thing for the IETF to do is to
>>> decline
>>> "help" from these plagirism systems.
>>
>> How can the IETF do that? I can't get past the idea that all we can do
>> is ask
>> contributors to state whether a contribution is AI-generated.
> 
> Define generated?

As far as I-Ds are concerned, the question I suggested was:
  1. Was AI used in the production of this draft? (Y/N).
That's binary. We can't reasonably ask for details.

Whether AI usage matters is (like an IPR disclosure) a judgment call.
I think we'll have to leave that to the normal process. If it produces
a better RFC, that's great. If it produces rubbish, we already know
how to deal with it.

As far as emails, reviews, etc., are concerned, I think we probably
need to keep it equally simple - request people to add a rider such
as "AI contributed to this text." After that, it's in the hands
of the reader.

    Brian

> 
> Yes, some LLM output is entirely prompt based. The workflow that I
> observe is that the human author will start with a prompt, and then
> iterate, progressively refining the prompt until getting an acceptable
> result. After how many iterations of prompt/observe/refine is the result
> still "AI generated"?
> 
> But that's certainly not the only workflow. Another one would be for the
> human author to write a draft, and then ask the LLM to improve it. Is
> that AI generated"?
> 
> Yes another workflow is to do something like that paragraph by
> paragraph, resulting in a mix of human written and partially AI
> generated paragraphs. Is that "AI generated?"
> 
> And then you have the workflow in which the human author starts typing
> the beginning of a sentence, and the AI helper proposes a continuation
> that the author may or may not accept. One could argue that each
> sentence in the document is either written or approved by the author. Is
> that still "AI generated"?
> 
> We have a valid concern about copyright and plagiarism. I think we
> should express us in these terms, rather than the blunt approach "is
> this AI generated?"
> 
> -- Christian Huitema
> 
> 
> 
> 
> 
>>
>> It does seem to me that we need something added to the Note Well ASAP.
>> Is the
>> IESG paying attention?