[ai-control] Re: WG Last Call: draft-ietf-aipref-vocab-03 (Ends 2025-09-18)

[Chris Needham <[e-mail]>]

Hi Mark, Suresh, all,

Overall, I don't think this draft is ready to progress and further work is needed, and I encourage the group to continue to iterate on this draft.

I also think that the work in this group has been influenced by (valid) concerns that, because the EC's AI Code of Practice mentions "any subsequent version" of RFC 9309, the group's drafts may create legal obligations. I would encourage the chairs to reach out to policymakers to ensure they are well informed of the status of discussions here.

My specific concerns:

1. Automated processing is too broad a scope

The scope, set out in the Abstract and Introduction, "a vocabulary of preferences regarding how automated systems process digital assets" seems overly broad. As a publisher the usage I'm most concerned about is processing in AI systems, and in particular generative AI. Introducing a scope for all automated systems risks unintended consequences for other uses that are unrelated to AI. So I suggest more clearly scoping the document to AI related usage.

Following from the above, the Automated Processing Category defined in Section 4.1 should either be revised (the original Text and Data Mining was preferable, because it's more specific) or removed entirely.

2. Missing category for RAG and Grounding usage

Without a category that covers RAG and grounding usage, the vocabulary does not meet publisher needs. This need was raised at the outset of this work, in the IAB workshop in Washington DC in 2024, and included in the WG charter scope.

In addition to reducing web traffic to original content, such usage introduces the possibility of creating misleading or biased statements, an issue that impacts news publishers in particular, risking undermining their credibility. Research published by the BBC goes into this in more detail, demonstrating that AI-generated answers summarise and present content in a way that can introduce factual inaccuracies and biases that are not present in the original published content.

https://www.bbc.co.uk/aboutthebbc/documents/bbc-research-into-ai-assistants.pdf

3. The Search Category is too broad

While I appreciate the effort that has gone into crafting this definition, as it stands it still allows all the generative AI usage that publishers find problematic. If adopted, publishers will be in the position of having to allow the Search Category, because they cannot realistically opt-out of traditional (i.e., non generative AI search).

Given that major search engines are rapidly integrating generative AI features, a generative AI response to a user query that includes links provided via RAG or grounding could be interpreted as "directs users to the location from which the assets were retrieved". The qualifying statements: "Only those parts of applications that direct users to the location of an asset are included in this category of use. This includes the use of titles or excerpts from assets that are used to help users select between multiple candidate options" don't address this concern. The same applies to "Parts of applications that do not direct users to the location of assets, such as summaries, are not covered by this category of use", because search engines that integrate generative AI do produce summaries.

4. Clarify "no preference" vs "unknown"

The "no preference" terminology could be misinterpreted as meaning that a website or content provider does not mind how their content is used (i.e., they don't have a preference either way). This would not be a valid assumption to make. It would be much more accurate to state that the preference is unknown. I recommend that the document aligns on the "allow", "disallow", and "unknown" statuses described in Section 5. This affects various parts of the draft:

* In Section 3, change "This can mean that no preference is stated for a given usage category" to "This can mean that the preference for a given usage category is unknown".

* In Section 3.2, which says "allow, disallow, or no preference"

* In Section 6, I suggest a rephrasing: "For example, the following states a preference to allow AI training (Section 4.2), disallow generative AI training (Section 4.3). The preference for categories other than subsets of these categories is unknown." (This would also fix the repeated "and" and "other" in the existing text.)

* In Section 6.5, "Otherwise, no preference is stated for that category of use" should be "Otherwise, the preference for that category of use is unknown".

* In Section 6.6, "A mapping might then define that no preference is associated with other categories" should be "A mapping might then define the preference associated with other categories is unknown."

5. Remove the list of examples where preferences could be ignored

The list of examples in Section 3.2 of cases where preferences could be ignored, should be removed. The preceding paragraph in this section makes clear the considerations where preferences could be ignored, and the specification should not attempt to define or identify specific examples. This should be left to policymakers, and it may vary by legal jurisdiction.

6. Minor editorial suggestions

Finally, I have some minor suggestions, editorial in nature. I'd be happy to create pull requests for any of these.

In Section 2, for consistency "Declaring party" should be "Declaring Party" (capital P).

In Section 3, "A statement of preference -- or usage preference -- is made about an asset" should be clarified, e.g., "A statement of usage preference is made about an asset".

Section 3.1 says "The process of obtaining a statement of preference has very limited scope for variation between implementations." It's not clear to me what the intent of this statement is. Implementations should follow the spec, certainly.

In Section 4.2, I suggest a minor change to the definition of the AI Training Category, to tie it more closely to terms in the Conventions and Definitions section, i.e., "The act of training Machine Learning (ML) models or Artificial Intelligence (AI)".

In Section 4.3, the definitions of the Generative AI Training Category mentions both "general purpose AI models" and "more specialized AI models", but both are described as "that have the capacity to generate text, images or other forms of synthetic content". It's unclear what distinction is being made here, so I recommend either clarifying, if there is a distinction, or simplifying the definition to "The act of training AI models that have the capacity to generate text, images or other forms of synthetic content."

In Section 5.1, for consistency, the second bullet should follow the same phrasing as the first, i.e., "If any statement of preference indicates that the usage is allowed, the result is that the usage is allowed."

In Section 6, I suggest rewording "Exemplary" (which implies excellence) to "Example" (which I think conveys the intended meaning, i.e., one of many possible). This is not to say the format given isn't excellent, though! This also affects Section 6.6, as well as Section 3.2 in the attachment spec, draft-ietf-aipref-attach-03.

With thanks,

Chris


________________________________
From: Mark Nottingham via Datatracker <[e-mail]>
Sent: 05 September 2025 03:09
To: [e-mail] <[e-mail]>; [e-mail] <[e-mail]>; [e-mail] <[e-mail]>
Subject: [ai-control] WG Last Call: draft-ietf-aipref-vocab-03 (Ends 2025-09-18)

External: Think before clicking

Subject: WG Last Call: draft-ietf-aipref-vocab-03 (Ends 2025-09-18)

This message starts a 2-week WG Last Call for this document.

Abstract:
   This document defines a vocabulary for expressing preferences
   regarding how digital assets are used by automated processing
   systems.  This vocabulary allows for the declaration of restrictions
   or permissions for use of digital assets by such systems.

File can be retrieved from:
https://datatracker.ietf.org/doc/draft-ietf-aipref-vocab/

Please review and indicate your support or objection to proceed with the
publication of this document by replying to this email keeping
[e-mail] in copy. Objections should be motivated and suggestions
to resolve them are highly appreciated.

Authors, and WG participants in general, are reminded again of the
Intellectual Property Rights (IPR) disclosure obligations described in BCP 79
[1]. Appropriate IPR disclosures required for full conformance with the
provisions of BCP 78 [1] and BCP 79 [2] must be filed, if you are aware of
any. Sanctions available for application to violators of IETF IPR Policy can
be found at [3].

Thank you.

[1] https://datatracker.ietf.org/doc/bcp78/
[2] https://datatracker.ietf.org/doc/bcp79/
[3] https://datatracker.ietf.org/doc/rfc6701/



--
ai-control mailing list -- [e-mail]
To unsubscribe send an email to [e-mail]