AI Preferences Working Group Materials

• Session Summary: 20250721-1230
  • Summary
  • Key Discussion Points
  • Decisions and Action Items
  • Next Steps

NOTE: This is a non-normative, AI-generated summary supplied only for convenience; it does not necessarily represent an accurate record of the meeting. See the minutes for the authoriative record. See the source for more information.

Session Summary: 20250721-1230

Summary

The AI Preferences (aipref) working group meeting focused on reviewing the progress made at the recent design team meeting in London, discussing new drafts, and addressing open issues related to AI preferences for content crawling and processing. The group emphasized the importance of staying within the scope of their charter, which is to standardize building blocks for expressing preferences, not to dictate policy on AI usage. Discussions centered around vocabulary definitions, attachment mechanisms, and the applicability statement for the framework.

Key Discussion Points

• Recap of Design Team Meeting: Overview of the hybrid meeting in Brussels and a fully virtual interim meeting. The design team made progress on resolving several issues in the drafts, with broad representation from various stakeholders including countries, AI companies, platforms, civic society, and regulators.
• Taxonomy of Categories of Use: Discussion on the high-level categories for preferences, particularly automated processing, AI training, and AI use.
• Generative AI Definition: Extended discussion on the definition of “generative AI” and whether it includes translation. The general consensus was to focus on use cases to determine when a separate preference signal is needed for translation.
• Overriding Preferences: Scenarios where expressed preferences might be overridden, such as contractual obligations or accessibility requirements.
• AI Definition: Arnoud raised the need to incorporate other SDO’s definitions.
• Contractual Agreements: Contracts between content providers and AI vendors may override preference, but the intention is not to allow any contract to override the framework.
• Accessibility: The accessibility use case can override content preferences in some situations.
• Robots.txt: discussion on the obsolescence of the robots.txt mechanism
• Search and AI Use: Clarification on the relationship between the “search” category and the new “AI use” category. Summarization of content with a link to the source can be considered search and therefore be permitted when AI use is disallowed.
• Applicability Statement: The group realized that an applicability statement is required to explain how it works in the bigger picture including for policy makers.
• Attachment Mechanisms: Description of the robots.txt and HTTP header attachment mechanisms. The potential for other attachment mechanisms such as embedding in content, or having registries, was acknowledged for the future.
• Multi-hop scenarios: Capture and transfer of preferences in multi hop scenarios such as the common crawl use case.

Decisions and Action Items

• Action Item: Martin and editors to refine the draft with new names and labels.
• Action Item: Get input from Arnoud regarding definitions that can be sent to the mailing list
• Action Item: Document accessibility concerns with content preferences and propose additions to the current text.
• Action Item: Develop and finalize the applicability statement document to address misunderstandings about the framework’s function.
• Action Item: Come up with new, alternative proposals for preference management to improve the productivity of discussions.

Next Steps

• Editors to incorporate feedback and release new drafts soon.
• The group will monitor progress and may schedule another online interim or hybrid face-to-face meeting (late September or early October) if needed.
• Working group last call to be initiated when editors/chairs think all open issues are addressed, otherwise individual consensus calls will be scheduled.