AIPREF

Summary

The AIPREF working group held a virtual interim meeting on June 9, 2026, to review the latest updates to draft-ietf-aipref-vocab (version 06), address outstanding issues from the Toronto interim, and plan the path forward for the upcoming IETF meeting in Vienna.

The primary focus of the meeting was to address lack of convergence on key parts of the draft. Key discussions centered on the public interest exceptions and conformance safeguards in Section 3.2, refining the definitions for training (Section 4.1) and search (Section 4.2), and exploring a potential new category for inference-time/grounding usage (Issue 172) in light of recent regulatory developments.

The presentation slides for the session can be found here:

Chair Slides

Key Discussion Points

1. Status of draft-ietf-aipref-vocab

Suresh Krishnan opened the meeting and welcomed participants, noting that while mailing list discussion has been active, the group has not yet converged on several issues.
Paul Keller outlined the main updates in draft-ietf-aipref-vocab (version 06). Notably, the editors added explicit disclaimers to sections lacking working group consensus, particularly Section 3 and the vocabulary definitions.
Paul Keller noted that the definitions for AI model training (Section 4.1) and search (Section 4.2) represent the consensus-seeking outcomes of the Toronto interim. They have received very little negative feedback since publication, though some minor adjustments may still be needed.

2. Section 3.2: Public Interest Exceptions and Conformance Safeguards

Lila Bailey reported on the progress of the small group that volunteered in Toronto to work on Section 3.2. The group agreed internally on the necessity of including public interest safeguards but struggled to find a compromise with participants who prefer to delete the section entirely. Lila invited publisher and web ecosystem participants to join the subgroup to help forge a middle ground.
Kevin Kelley suggested adopting a common IETF approach: utilizing non-normative, illustrative examples rather than attempting to draft an exhaustive, normative list of exemptions. Alissa Cooper and Lila Bailey agreed that a succinct, illustrative approach based on Lila’s original draft could serve as a viable middle ground.
Farzaneh expressed reservations about any list of exceptions, arguing that they are rarely comprehensive, can be interpreted subjectively, and risk creating barriers for end-users and developers. She advocated instead for focusing energy on refining the core technical category definitions.
Lila Bailey proposed an order of operations: first, reach a clear consensus on the technical definitions and their scope, and then assess whether public interest safeguards are necessary to mitigate unintended consequences. She also raised concerns about external policy bodies pointing to the IETF to resolve policy questions, and asked if Chairs or ADs could make a formal scope determination.
Suresh Krishnan responded that a draft omitting Section 3.2 or equivalent conformance/applicability guidance entirely would likely fail to pass a Working Group Last Call.
Paul Keller agreed with Lila Bailey and Alissa Cooper that technical definitions in Section 4 must remain crisp, and that editorial clarity requires keeping applicability and conformance guidelines in Section 3.
Timid Zehta cautioned that any proposed refinements to the model training definition (Section 4.1) must internalize previous group discussions, noting that overly specific technical definitions in earlier iterations were found to be factually incorrect.
Chris Needham, Nate Hake, and Leonard Rosenthol volunteered to join the Section 3.2 small group to represent a wider range of viewpoints.

3. Section 4.1 (AI Model Training) and Section 4.2 (Search) Categories

Suresh Krishnan noted that no formal pull requests have been submitted against Sections 4.1 or 4.2 since the Toronto meeting.
Erin Simon stated that the current Search category definition feels overly restrictive. She is analyzing how to introduce more operational flexibility to support product deployments, especially in light of recent regulatory developments, and plans to submit a text proposal.
J. Gimbel reported that they are reviewing the text alongside Chris Needham and expect to open issues or submit comments on these sections shortly.

4. Inference Time, Grounding, and RAG (Issue 172)

The working group discussed the high demand for a post-training/inference-time category (e.g., Retrieval-Augmented Generation or "grounding"), which would likely become Section 4.3.
Paul Keller summarized a mailing list thread regarding the UK Competition and Markets Authority (CMA) publisher contract requirements. His analysis of the CMA documents indicated four primary use cases: three are covered under the draft's current training and search categories, but the fourth ("grounding") is not. He proposed introducing a new "grounding" category modeled on this framework.
Chris Needham supported a standardized, interoperable grounding definition that applies broadly across the web ecosystem rather than to a single company. He suggested adding the various use cases proposed on the mailing list to the AIPREF Use Case Wiki to help test future definitions.
Kevin Kelley and Leonard Rosenthol emphasized the need to align on a precise technical definition of "grounding" before drafting text, to ensure all participants are discussing the same concept.
Paul Keller noted that while there is strong support for an inference-time category, the group must systematically resolve its technical scope and boundaries.

Decisions and Action Items

Action Item: Lila Bailey to lead the expanded small group (including Chris Needham, Nate Hake, and Leonard Rosenthol) to draft a compromise proposal for Section 3.2 based on a more succinct, illustrative approach.
Action Item: Alissa Cooper to review the Section 3.2 draft proposal once produced by the small group.
Action Item: Chris Needham to document grounding and inference-time use cases on the AIPREF Use Case Wiki.

Next Steps

The Section 3.2 small group aims to share its text proposal on the mailing list by early July 2026 to allow for review and discussion prior to the Vienna meeting.
J. Gimbel and Erin Simon to submit comments or open GitHub issues regarding proposed flexibility updates to Section 4.1 and Section 4.2.
The working group will continue discussions on Issue 172 (grounding/inference-time categories) on the mailing list to prepare for a deeper technical discussion at the IETF meeting in Vienna (week of July 20, 2026).

Automatic IETF Minutes