[ai-control] Re: Proposed "AI System Inference" and "AI User Input" categories

[Leonard Rosenthol <[e-mail]>]

I have mentioned it many times, Farz.  As have others such as Nate and Laurent, and just earlier today, Sebastian.

The work item in question is Vocabulary​ document - it is not a protocol.   And the Vocabulary will be used in both online and offline scenarios, as noted in the Charter (https://datatracker.ietf.org/wg/aipref/about/) which clearly calls out support for “including preferences in content metadata” (which implies that such information can be retrieved offline).

Leonard

From: Farzaneh Badiei <[e-mail]>
Date: Friday, August 7, 2026 at 9:09 AM
To: Leonard Rosenthol <[e-mail]>
Cc: Nate Hake <[e-mail]>; Mirja Kuehlewind (IETF) <[e-mail]>; Alissa Cooper <[e-mail]>; [e-mail] <[e-mail]>
Subject: Re: [ai-control] Re: Proposed "AI System Inference" and "AI User Input" categories


EXTERNAL: Use caution when clicking on links or opening attachments.


Hi

You mentioned this "offline" aspect before Leonard. Can you please provide a concrete use case for the offline aspect?  I just can't imagine an "Internet" engineering body coming up with a protocol that targets assets' "offline".


My concern is that the conversations we’ve had on this topic, throughout our 2025 meetings and during Toronto and Vienna 2026, don’t seem to be reflected in this new conversation and text at all. Perhaps this text aims to prompt other stakeholders to reiterate the counter arguments for adding inference?

We’ve discussed this issue repeatedly under different names: AI use, substitutive use (inference and RAG included), AI output, “Lederhosen,” and Bradley’s substitutive-use proposal (draft-silver-aipref-vocab-substitutive-00: https://www.ietf.org/archive/id/draft-silver-aipref-vocab-substitutive-00.txt) Krishna's alternative approach... which to this day I still think is more promising...

In my opinion even Bradley’s draft is more balanced than this text because he considered the concerns that were raised and the text tries to exclude the user when assets are lawfully acquired, ( just to be clear we raised objections to that framing as well when it appeared and we still do, I am just pointing out other discussions that happened)

The Toronto 2026 discussions, where multiple arguments were made about the impact on the user, don’t appear to have been factored in either. Nor does the new text engage with any opposing positions or concerns raised during the Vienna meeting.

Since we have had these conversations so many times, resuming the discussion without first recalling the opposing views made is not fruitful.

I think instead of working on a text about "inference", we need to revisit our earlier discussions in our upcoming London meeting, see what has changed and what remains the same and hopefully the chairs will point us in a direction.



On Fri, Aug 7, 2026 at 7:56 AM Leonard Rosenthol <lrosenth=[e-mail]<mailto:[e-mail]>> wrote:
Mirja and Alissa - keep in mind that we are building a vocabulary that will then be used by various attachment methods - not all of which are “on the web” or even “online”.

And in those scenarios, users are the primary use case, not the automation/system.

Which is why we need to keep the “User initiated” flow in the discussion and can’t separate it.

Leonard

From: Nate Hake <[e-mail]<mailto:[e-mail]>>
Date: Friday, August 7, 2026 at 7:49 AM
To: Mirja Kuehlewind (IETF) <[e-mail]<mailto:[e-mail]>>
Cc: Alissa Cooper <[e-mail]<mailto:[e-mail]>>; [e-mail]<mailto:[e-mail]> <[e-mail]<mailto:[e-mail]>>
Subject: [ai-control] Re: Proposed "AI System Inference" and "AI User Input" categories


EXTERNAL: Use caution when clicking on links or opening attachments.


The desire to cover the full space, at least for my part, is because:

1) Generative AI is, in my opinion, a regressive technology that is being deployed in a way that is severely harming the world in many ways. I do not wish to contribute to it with my labor. Full stop. Public polling consistently shows many other people — possibly even majorities in many places — share my view. Suffice to say, “No AI” it is a common enough preference that we should allow it to be expressed.

2) I do not wish my creations to be used by generative AI, again full stop. Generative AI devalues and degrades creative assets each time it is used. Again, my opinion, but one I am far from alone in holding.

3) Any gap that is not expressible via a preference is likely to be exploited. And we are operating in a low to no trust environment.

4) For the reasons listed above, if there isn’t a way to say no, asset owners will be incentivized to put their assets behind lock and key — or to stop publishing altogether. We are already seeing this in the status quo with the proliferation of gated content. And we have seen incredible consolidation of web publishing in the last few years, as small and independent publishers are slowly being driven out of business in the current environment (AI is a technology that inherently leads to consolidation).

Giving assets owners a voice will encourage them to feel comfortable publishing on the open web.


***
[https://travellemming.com/wp-content/uploads/Nate-Square-Photo.jpg]
Nate Hake
Founder
TravelLemming.com<https://travellemming.com/>

On Fri, Aug 7, 2026 at 18:27 Mirja Kuehlewind (IETF) <[e-mail]<mailto:[e-mail]>> wrote:
I’m not fully sure why there is this desire to cover the whole space…? Generally,  I think there will always be cases that were not considered when design these term. Further in this case, I think it would be more important carve out direct user actions as that seems actually go beyond the intended scope of this group. Having an overly broad definition would simply lead to a mismatch between content owner expectation and real use (given a human user will also not be aware of these preference).

Also how  could/should the Ai User Input be implemented practically? A user copying some content from a website might not be ware of the preference. How does an AI know to which page the content belongs and that there is a preference? User provided links are different but such a preference would then just lead back to this stupid pattern that we had in the past, where the AI asks the user to past the content in…

Mirja



On 7. Aug 2026, at 13:09, Nate Hake <[e-mail]<mailto:[e-mail]>> wrote:

I am certainly happy to take the discussion in a particular order if the chairs find it helpful. That said, the fundamental challenge is that some of us strongly want the ability to express a preference as to *all* inference. So a system category alone would be woefully insufficient from my perspective.

To achieve full coverage, any contours we draw around what constitutes “system” would need to have corresponding categories that cover whatever is not “system.”

So I do frankly think it will be difficult to bifurcate the discussion cleanly as these proposed categories are really meant to work in tandem.



***
[https://travellemming.com/wp-content/uploads/Nate-Square-Photo.jpg]
Nate Hake
Founder
TravelLemming.com<https://travellemming.com/>

On Fri, Aug 7, 2026 at 17:46 Alissa Cooper <[e-mail]<mailto:[e-mail]>> wrote:
From a process perspective it might be helpful to focus the discussion on system-directed inference and see if the group can find any sort of agreement on that category, and then come back around to the user-directed inference question. It would be helpful to see if progress can be made on the former without having to wade through the contention on the latter.

Alissa


On Aug 1, 2026, at 4:55 PM, Greg Lindahl <[e-mail]<mailto:[e-mail]>> wrote:

> Rights owners have expressed their need for a preference regarding AI inference opt-out since the start of this IETF project

That's true -- and we have discussed it extensively. Let me remind you of an example from my past.

I had a search engine in 2010. For many reasons we needed to implement "safe search", so that users who had not explicitly expressed a preference about naughty websites would not be surprised by the occasional naughty photos or text website.

The standard implementation, back then, was to run every webpage in the response through AI inference. If the porn score was above a threshold, and if 80%+ of the top 10 results were above the threshold, then we would show the naughty results -- else we would suppress them.

Without the ability to do AI inference, there's no good way to do safe search. If 1% of the web opted out of inference, the best solution is to never show those websites in search results.

This is probably not what you expected.

And this isn't a new issue, this is a situation from 16 years ago.

-- greg (formerly CTO of blekko)


On Thu, Jul 30, 2026 at 7:10 AM Laurent Le Meur <[e-mail]<mailto:[e-mail]>> wrote:
Hi Farzaneh,

Rights owners have expressed their need for a preference regarding AI inference opt-out since the start of this IETF project, and even before. Such preference/control is already present in multiple opt-out models (C2PA, RSL ...).

Section 3.2 covers the fact that AI preferences may not be followed by users and client systems. Therefore, the specification leaves the evaluation of their impact on accessibility, translation, and other user-beneficial features to the client systems and their users. Not us.
Thanks to section 3.2, the document you reference is now mainly of interest to developers of client systems.

Best regards
Laurent LE MEUR


Le 30 juil. 2026 à 15:02, Farzaneh Badiei <[e-mail]<mailto:[e-mail]>> a écrit :

One thing that warrants a long and in-depth discussion is that since the beginning of this working group, some of us time and time again have warned against having an inference category especially for the consumer. It can have an impact the use of accessibility and translation features along with many more issues. Meredith, Lila, Jo and myself have time and time again raised these concerns in different shape and form. I raised them last week as well.

We had submitted this draft to illustrate some of the impact. https://www.ietf.org/archive/id/draft-farzdusa-aipref-enduser-00.html

In my opinion, we need to take a step back and before we discuss any text, we should discuss the addition of inference, whether the group agrees with it and under what circumstances the group agrees with adding inference.



--
ai-control mailing list -- [e-mail]<mailto:[e-mail]>
To unsubscribe send an email to [e-mail]<mailto:[e-mail]>
--
ai-control mailing list -- [e-mail]<mailto:[e-mail]>
To unsubscribe send an email to [e-mail]<mailto:[e-mail]>

--
ai-control mailing list -- [e-mail]<mailto:[e-mail]>
To unsubscribe send an email to [e-mail]<mailto:[e-mail]>
--
ai-control mailing list -- [e-mail]<mailto:[e-mail]>
To unsubscribe send an email to [e-mail]<mailto:[e-mail]>

--
ai-control mailing list -- [e-mail]<mailto:[e-mail]>
To unsubscribe send an email to [e-mail]<mailto:[e-mail]>