AIPREF

Summary

The AIPREF working group held an interim virtual session to review open issues, evaluate proposed text changes, and align on topics ahead of the upcoming IETF meeting in San Francisco. Chairs Mark Nottingham and Suresh Krishnan facilitated the discussion through the working group's GitHub issues and pull requests associated with draft-ietf-aipref-vocab and draft-ietf-aipref-attach.

Key topics addressed included issue triage, replacing "asset" terminology with "artifact", clarifying the distinction between access control and usage preferences, evaluating the inclusion of the term "synthetic", retaining usage preferences across data transformations and model redistributions, and refining the definition of AI usage in relation to user intent.

Key Discussion Points

Issue Triage & Closed Issues

Issue #208 (Use of "nonsubstantive"): Confirmed as resolved and closed; the term does not appear in the current editors' draft.
Issue #172 (Category for RAG and grounding): Confirmed as resolved and closed, covered by the current usage categories.
Issue #171 (Category for substantive use): Bradley Silver agreed that the issue can be closed as addressed by usage categories.
Issues #153 and #154 (Defaults, unknown vs. allowed) & Krishna Madhavan's display-based preferences proposal: Remain deferred until core vocabulary structures stabilize.

Terminology: Asset vs. Artifact (Issue #163)

Martin Thomson presented a pull request updating terminology across draft-ietf-aipref-vocab and draft-ietf-aipref-attach to replace "asset" with "artifact".
Timid Zehta pointed out a residual occurrence of "asset" in a definition within the PR, which Martin Thomson noted as an oversight to be corrected.
The group supported merging the change once current in-flight pull requests are aligned to avoid merge conflicts.

Distinguishing Data Access from Data Use

Martin Thomson presented a pull request adding Section 3.3 to draft-ietf-aipref-vocab, reinforcing that preferences express conditions on data use rather than restricting network access.
Eric Rescorla suggested focusing the text on contrasting usage preferences against traffic management directives (e.g., robots.txt) rather than listing specific technical controls like encryption. Kevin Kelley supported this distinction. Martin Thomson agreed to adjust the wording accordingly.

Inclusion of the Term "Synthetic" (Issue #217)

The group debated whether the term "synthetic" in the definition of generative AI models in draft-ietf-aipref-vocab provides necessary precision or introduces ambiguity/loopholes.
Chris Needham, Eric Rescorla, Bradley Silver, Nate Hake, Matthew Hervey, and Timid Zehta expressed concern that "synthetic" is either redundant with "generate" or creates potential ambiguity regarding model outputs (such as verbatim extracts or memorized text). Kevin Kelley noted the term was intended to be evocative of generative substitute use.
Eric Rescorla and Martin Thomson emphasized that specification definitions must stand on their own without relying on unwritten context or list discussions.
The chair conducted a session poll to gauge perspective on keeping the term.

Session Poll

Question: Do you feel strongly that we should keep the 'synthetic' term in the draft?
- Results: Yes: 3, No: 15, No Opinion: 1 (Total: 28)
Outcome: Proponents who wish to keep the term are encouraged to propose a concrete definition and specific load-bearing use cases ahead of San Francisco. Otherwise, removing or simplifying the term will be considered.

De Minimis Content Generation (Issue #231)

Erin Simon raised Issue #231 regarding minor/incidental content generation.
Martin Thomson noted that the generative AI model definition already explicitly excludes models used purely for classification, even when generating explanatory rationale for output.
Erin Simon agreed to review the current wording with a view to closing the issue if covered.

Model Redistribution and Preference Retention (Issues #245 and #246)

The group discussed whether and how preferences must travel when artifacts are redistributed, transformed, or compiled into models or intermediate forms (e.g., tokenization, embeddings, or intermediate datasets).
Timid Zehta noted that Section 3.3 in draft 08 of draft-ietf-aipref-vocab satisfies Issue #245 regarding conveying preferences alongside models.
Eric Rescorla pointed out a potential asymmetry if model trainers are subject to preference retention expectations while upstream data scrapers/crawlers are not.
Martin Thomson highlighted PR #252 and PR #253, which add language to address preference retention across system boundaries, scraping, and transformations such as embeddings.
Bradley Silver, Brendan Quinn (via chat), Nate Hake, and Matthew Hervey emphasized that preferences must generically travel with content across transformations to prevent circumvention by intermediate data brokers.

Defining AI Usage & User-Directed Exceptions (Issues #249 and #254)

Kevin Kelley and Chris Needham introduced PR #254, which refines the scope of "AI usage" around specific user intent and specific resources (distinguishing direct user-driven interactions with specific assets from broad automated crawling/scraping/training).
Bradley Silver, Alissa Cooper, Nate Hake, and Eric Rescorla discussed edge cases in determining "user intent" and identifying "specific assets" (e.g., pasting a URL vs. directing a model to fetch and summarize a site).
Nate Hake expressed strong concerns that excluding user-initiated actions from preference scope would undermine creators and artists whose work is replicated or altered by end users using AI tools, arguing that preferences should be expressible across all AI usage categories.
Farzaneh highlighted the need to address accessibility carve-outs (e.g., text-to-speech, translation, screen readers) alongside general usage definitions.
Martin Thomson and Mark Nottingham suggested that further progress requires detailed text proposals containing explicit explanatory material and concrete examples rather than overly concise definitions.

Decisions and Action Items

Decisions

Closed Issues: Issues #208, #172, and #171 are officially closed.
Terminology: The terminology shift from "asset" to "artifact" (PR by Martin Thomson for Issue #163) will be merged once remaining draft text updates align.

Action Items

Martin Thomson: Fix residual "asset" typo in the artifact PR and update Section 3.3 text based on Eric Rescorla's feedback (focusing on robots.txt contrast and replacing "prevent" with "constrain/restrict"). Link PR #252 and PR #253 to Issue #245 and Issue #246.
Erin Simon: Review current generative model definitions in draft-ietf-aipref-vocab to confirm if Issue #231 can be closed.
Proponents of 'Synthetic': Provide a proposed definition and load-bearing rationale if proposing to retain "synthetic" in Issue #217.
Kevin Kelley, Chris Needham, and Collaborators: Fleshes out PR #254 with explanatory text, examples, and edge-case definitions regarding user intent and specific resources for IETF 122.
Farzaneh: Prepare use cases regarding accessibility carve-outs prior to the San Francisco meeting.

Next Steps

Further proposals and draft text updates will be posted to the AIPREF mailing list and GitHub repository prior to IETF 122 in San Francisco.
The chairs will begin formal consensus calls on stable topics following review of updated drafts in San Francisco.

Automatic IETF Minutes