Markdown Version | Transcript | Session Recording

Session Date/Time: 14 Apr 2026 16:45

AIPREF

Summary

The AIPREF Working Group held its second interim meeting of 2026 to make progress on the core vocabulary specification, draft-ietf-aipref-vocab. The session focused on refining the core "training" and "search" preference definitions, moving towards a purpose-based vocabulary model, and streamlining the specification by stripping out unnecessary high-level definitions. Meeting materials and remote participation were coordinated via the Webex Link presentation slides.

Key Discussion Points

1. Introduction and Use Case Review

• Meeting Kickoff: Jo Levy opened the meeting, reviewing the IETF Note Well policies, agenda, and the group's timeline. Jo Levy noted the need for concrete progress outside of meetings to ship specifications successfully.
• Granularity of Preferences: The working group reviewed the use cases documented on the wiki (training, use, and presentation). Bradley Silver and Chaerin Lim noted that preference signals should allow for granularity, particularly regarding how private agreements override general preference signals.
• Scope of Use Cases: Kaustubh Phatak and Bradley Silver discussed whether use cases should capture the specific reasons an entity might train on content or if the focus should remain strictly on the preferences expressed by content owners.

2. Purpose-Based vs. Technology-Based Model

• Core Philosophy: Jo Levy and Suresh Krishnan noted that the working group has transitioned away from defining preferences based on specific technical ingestion methods, focusing instead on purpose-based categories.
• Handling Evolving Use Cases: Eric Rescorla, Andrew Sullivan, and Bradley Silver debated how to handle secondary use cases and future technology evolutions. Andrew Sullivan cautioned against trying to specify every possible exception, emphasizing that the protocol should focus on the most common, clear preference expressions.

3. Training Preference Category (draft-ietf-aipref-vocab)

• Removal of "Large": The group discussed Issue #183 regarding the subjective nature of the word "large" when describing models. Chaerin Lim, Suresh Krishnan, and Bradley Silver agreed that quantifying model size thresholds is impractical. The group agreed to remove "large" from the text.
• Generative vs. Classification Models:
  • A major technical discussion occurred regarding whether the training category should encompass simple machine learning models (e.g., spam classifiers, security analysis, or CSAM detection).
  • Eric Rescorla, Andrew Sullivan, and Bradley Silver debated whether a model trained for classification purposes inherently possesses generative capabilities.
  • Bradley Silver and Andrew Sullivan expressed concerns that overly broad definitions could unintentionally sweep in benign, non-generative utilities.
  • A self-organized subgroup (including Bradley Silver, Eric Rescorla, Chaerin Lim, Andrew Sullivan, and Kaustubh Phatak) met over lunch to draft a refined training model definition (Issue #1908). They proposed separating definitions into "general-purpose generative models" and "specialized generative models" while explicitly excluding simple classifiers.

4. Search and Non-Generative Search Preference

• Defining the Search Exception: The group debated how to formulate a search exception (often referred to as "traditional search" or "laterhosen search") to allow indexation and snippet generation while opting out of broader training.
• Proposals and PRs: The group examined two primary text proposals:
  1. PR 1909 (Amended by Chaerin Lim): Focuses on allowing the backend processing of assets to perform indexing and ranking, utilizing models exclusively for the search application, while explicitly excluding generative search summaries.
  2. PR 201 (Proposed by Jo Levy): A single-sentence definition outlining search output capabilities (links, snippets) with detailed sub-bullets specifying exclusions for generative outputs.
• Points of Contention:
  • Verbatim Snippets vs. Adapted Content: Kaustubh Phatak, Bradley Silver, and Chaerin Lim discussed whether translation, transcription, or non-substitutive modifications should be permitted under the search preference. Chaerin Lim argued that search preferences should not regulate translations, while Bradley Silver highlighted how minor text rewrites can assist user navigation.
  • Model Training for Search: Bradley Silver and Andrew Sullivan debated whether it is technically necessary to allow model training to evaluate and rank content for a non-generative search index.

Decisions and Action Items

Decisions

• AI Training Definition: Agreed to remove the subjective quantifier "large" from the definition of target AI models in draft-ietf-aipref-vocab.
• Simplifying Terms: Agreed to remove the high-level definitions for "Artificial Intelligence (AI)" and "Machine Learning" from the draft, as they are not technically required to define the specific vocabulary terms.

Action Items

• Editors of draft-ietf-aipref-vocab: Synthesize the feedback from the interim meeting to generate an amalgamated proposal combining the concepts from PR 1909 and PR 201 regarding the "search" preference category.
• All Participants: Review the proposed text for the training model definitions in Issue #1908 and the search exception proposals, and file GitHub issues highlighting specific objections or required refinements (specifically regarding translation/transcription and snippet boundaries) ahead of the next session.

Next Steps

The working group will resume discussions in the next session to:

1. Address context/attachment-level text issues.
2. Review the editors' amalgamated search and training text proposals.

Related Documents

draft-ietf-aipref-vocab