Session Date/Time: 02 Oct 2025 07:15

AIPREF

Summary

The AIPREF working group session continued discussions on defining preferences for AI models, focusing on the tension between intuitively understandable categories and technically precise ones. A significant portion of the session was dedicated to a report from an ad-hoc group that met earlier, which proposed refined categories for AI output and training. While the group acknowledged progress in understanding diverse perspectives, fundamental challenges remain regarding the technical feasibility of certain preferences (e.g., opting out of AI training while maintaining search engine visibility) and the interconnectedness of various technical aspects. The session concluded with a plan to develop a pull request reflecting the refined categories for review before the Montreal IETF meeting, along with outlining future interim meetings.

Key Discussion Points

WG Discourse and Conduct: The Chair opened by emphasizing the need for respectful, good-faith participation and charity in discussions, acknowledging the complex, non-technical societal issues often underlying the group's work. Progress in this area was noted, despite initial concerns.
Session Focus: The main agenda was to continue discussions from the previous day regarding proposals for AI output and training, with a particular focus on core vocabulary definitions. The "attachment document" and detailed definitional issues around "AI" were deferred.
Report from Morning Ad-Hoc Group:
- Category Definition Tension: A key tension identified was between creating categories that are intuitively easy to understand (e.g., "Google Search") and those that are technically precise (e.g., output linking, exact quotes). Product-specific categories were seen as potentially brittle due to evolving technology.
- Generative AI Training Challenges: Concerns were raised that a strict "generative AI training" opt-out might be incompatible with the fundamental operation of modern search engines, which increasingly rely on generative AI models. This prompted a re-evaluation of the AI training category's scope.
- Interconnectedness of Issues: Progress was noted in fostering trust, but participants highlighted the difficulty in resolving individual issues (e.g., AI output definitions) without simultaneously addressing their deep interconnections with other aspects (e.g., model training). A suggestion was made to try and isolate pieces for progress, acknowledging that nothing is fully agreed until everything is.
Technical Feasibility vs. Legitimacy of Preferences:
- The discussion explored whether content owners' desired preferences (e.g., "content discoverable in search but not used for model training") are technically possible.
- Participants emphasized the need to clearly inform users about the consequences of expressing certain preferences, drawing an analogy to the Brexit referendum where citizens were not fully informed.
- Design principles of "no more complex than necessary" and avoiding "locking in" current internet architecture were highlighted.
Refined Category Proposal (Post-Ad-Hoc Group Report):
- AI Output: A more narrowly defined category focusing on observable system outputs rather than internal workings.
- "Lederhosen" Category: Proposed as a refinement of "AI Output" with additional stipulations, approximating a "traditional search" experience (e.g., exact text match, snippets with links, not synthetic content).
  - Associated Training: Crucially, this category is proposed to inherently include the training of AI models necessary to produce and operate the search system itself. This addresses the technical impossibility of separating search functionality from underlying AI model training.
  - Implementability: Implementers indicated this category seems plausible, but noted that overly granular preferences might lead to broader content exclusion for product feasibility.
- "Foundational Models" Category: This concept emerged as the primary focus for AI training preferences, replacing "General Purpose AI" (due to legal/jurisdictional connotations) and refining "Generative AI." Foundational models are understood as the large, expensive models critical for many AI applications, offering a clearer scope for content owners' concerns.
- Nesting: The question of whether "Lederhosen" is a nested subset of "AI Output" remained unresolved, with implications for how preferences would cascade.
Draft Development and Future Steps:
- A pull request (PR) currently exists in the repository for "AI output" and "search" (the precursor to "Lederhosen"). This PR does not yet include the "associated training" stipulation or the "foundational models" concept.
- A strong desire was expressed to update the working group draft to reflect these discussions, even if full consensus is not yet reached, to provide a current basis for discussion and feedback.
- The use of non-descriptive, placeholder names (e.g., "Lederhosen") was suggested for categories to avoid premature assumptions and confusion, focusing instead on precise definitions.
- The group needs to continue refining the definitions of these categories and explore permutations of preferences (e.g., "Lederhosen no, AI Output yes") to ensure logical consistency and clarity.

Decisions and Action Items

Decision: The working group will proceed with developing three key concepts for preferences: a narrower AI Output category, a stipulated "Lederhosen" (search-like) category, and a Foundational Models category for AI training. The "Lederhosen" category will implicitly include necessary model training for its operation.
Action Item: Editors to develop a pull request (or update an existing one) to include text for the "AI Output" and "Lederhosen" categories with their definitions and stipulations, including the "associated training" aspect. The "Foundational Models" category will also require clear definition.
Action Item: The PR/updated draft should use placeholder/nonsense names for categories initially to focus discussion on definitions rather than labels.
Action Item: Chairs to consider publishing a new version of the draft reflecting these changes before the Montreal (IETF-118) deadline, contingent on the text being in a reasonable shape and clearly stating that its inclusion does not imply consensus.
Action Item: Participants are encouraged to review the upcoming draft/PR and provide specific feedback, particularly on the clarity of definitions and logical consistency of preference combinations.
Action Item: Develop a clearer enumeration of what various combinations of preferences mean (e.g., a web page with permutations) to aid understanding.

Next Steps

IETF-118 (Montreal): Two sessions are planned. The updated draft/PR will serve as the primary anchor for discussions, aiming to solicit broader IETF community feedback.
Online-Only Interim Meeting: An online interim meeting (approx. 2 hours) will be scheduled in December to maintain momentum.
Hybrid Interim Meeting: A hybrid interim meeting is tentatively planned for late January/early February 2024. The Chairs will solicit input on potential locations (East Coast North America like New York or Toronto, or accessible European cities like London, or Tokyo) and specific dates, balancing participant accessibility with the need for in-person collaboration.
Ongoing Work: Participants are encouraged to continue collaborating on proposals and discussions outside of formal meeting cycles to bring well-formed materials for review.
Communication: Ensure clear communication of preference meanings and their consequences for content owners and consumers.