Markdown Version | Recording 1 | Recording 2

Session Date/Time: 01 Oct 2025 10:30

AIPREF

Summary

The AIPREF session focused on intense discussions around proposed vocabulary terms for expressing preferences for AI usage of content. A "hybrid proposal" emerged from lunchtime discussions, aiming to consolidate previous display-focused and non-linked use drafts into an "AI Output" category with granular controls. This was contrasted with a "Cloudflare proposal" that advocated for simpler, distinct search, AI input, and training categories. Key debates revolved around the scope of "AI Output" (especially its interaction with traditional search results), the distinction between controlling "input" versus "output" of AI models, and the implications for AI training and open-source models. A separate, brief discussion introduced a proposal for a "Terms of Service" pointer in robots.txt. No definitive decisions were made, with calls for further refinement and convergence of proposals.

Key Discussion Points

Hybrid Proposal for AI Output Preferences

• Proposal Overview: A hybrid proposal, combining elements of the "Ladawan display-focused draft" and "Simon me non-linked use draft," was presented. This proposal introduces an "AI Output" category to address scenarios where an AI model fetches content to inform a user query response.
  • It allows content creators to set conditions on AI output, such as requiring a link back to the original content or requiring output to be an exact quote.
  • Combining link and quote conditions was noted as approximating traditional search snippets.
  • Initial thoughts on including max-text-length and max-image-size parameters were deferred for later.
• "Blue Link Results" and Search Exclusion: A significant point of contention was whether traditional "blue link" search results (list of links, titles, snippets) should fall under the "AI Output" category.
  • As drafted, stating "AI Output: no" would likely exclude content from traditional search results, which several participants indicated would contradict content owners' desires to appear in search while disallowing generative AI summaries.
  • The need to clarify whether traditional search remains outside or inside this category, and how to allow content in traditional search without enabling broader AI outputs, was emphasized.
• Focus on Output vs. Input: A participant raised concerns that focusing on "output" makes it difficult for content creators to trace how their content is used and transformed internally. They expressed a preference for controlling "AI Input" to prevent ingestion or retention for generating downstream content.
  • Response: It was clarified that robots.txt governs content acquisition. The AIPREF work focuses on expressing preferences for what happens after content is obtained, influencing a crawler's decision to fetch content based on its intended use. The output focus allows content to be ingested for relevance scoring or ranking without necessarily permitting summarization.
• "User Queries" and Agentic AI: Questions were raised about the phrase "following user queries" in the definition and its implications for "agentic AI" applications (where an AI agent makes queries, potentially not directly for a human end-user).
  • The intent was stated that agentic applications are in scope if their multi-step process ultimately results in an output delivered to an end-user. However, content supplied by the user in their request would be out of scope.
  • Concern was expressed that including agentic AI overly complicates the definition and that the system boundary for "output" needs clearer articulation (e.g., "system boundary output" rather than "user query").
• Generative Definition: The challenge of defining "generative" in a way that differentiates search ranking models (which generate snippets) from broader AI summaries was discussed.
• Grounding: The proposal was confirmed to intend to cover grounding.
• Oversimplification Concerns: A sense of those present indicated a tension between the desire for simple, easily understandable preferences for webmasters and the need for granularity to address the complex technical realities of AI models and varied publisher intentions. Oversimplification could lead to unintended consequences (e.g., disallowing safety features) or mismatched expectations.

Cloudflare Proposal (Michael and Lea)

• Proposal Overview: This proposal suggested three distinct vocabulary terms: search, AI input, and training, aiming for simplicity and clear definitions.
  • search: Defined for traditional search, including indexing, ranking, and returning blue links. Explicitly states it does not include AI-generated search summaries.
  • AI input: Aims to prevent content from being ingested by any model, addressing concerns about downstream usage.
  • training: For content ingestion for the purpose of training models for arbitrary usage.
• Critiques and Discussion:
  • The AI input term was criticized for potentially being too broad, restricting AI processes for purposes like spam detection or child pornography protection (which often use AI/RAG methods) that are not typically seen as "AI outputs" to a user.
  • Concerns about the clarity of the search definition, particularly whether "traditional search" can truly exist without AI processing being involved in snippet generation, leading to the same ambiguity as the hybrid proposal.
  • The proposal was seen by some as an oversimplification that reintroduces problems that other drafts tried to solve with more nuanced language.
  • A participant noted that webmasters they surveyed consistently express desire for output or display controls, not input controls.
• Training and Open-Source Models: Discussion extended to how these categories interact with AI training, especially for open-source models.
  • If content specifies "no generative AI training," this could mean models trained on such content cannot be released with open weights that permit general generative use, potentially hindering open-source AI development.
  • The technical difficulty of clearly distinguishing "generative" from "non-generative" models was highlighted, as many modern AI architectures possess generative capabilities regardless of their intended primary application.

Terms of Service Pointer in robots.txt (Max Schindler)

• Proposal Overview: A proposal was presented to add a Terms-of-Service field to robots.txt, pointing to a URL, similar to Sitemap. This is intended as an additive mechanism, agnostic to the vocabulary discussions, allowing publishers to convey preferences not covered by standardized terms.
• Discussion and Concerns:
  • Machine Readability: A primary concern was that Terms of Service are human-readable legal documents, not machine-readable. AI models attempting to interpret them could lead to misinterpretations and misinformation.
  • Utility: If not machine-readable, what is the practical utility of the pointer for bot operators? It was suggested it could be for human review, but its impact at scale was questioned.
  • Multiple ToS: Sites often have multiple ToS documents (e.g., by region, user type, content type). The proposal would need to accommodate multiple pointers, and clarification on which specific ToS applies to crawlers.
  • Conflicts: Potential conflicts between robots.txt preferences and linked Terms of Service would need a clear conflict resolution mechanism.
  • Charter Fit: Concerns were raised about whether pointing to a non-standardized legal document fits the AIPREF working group's charter, which focuses on technical preferences. This could be seen as short-circuiting the vocabulary definition process.
  • A participant noted that this proposal represents an "alternative" path for content owners if the AIPREF working group's standardized preferences do not adequately address their concerns.

Decisions and Action Items

• No firm decisions were reached on adopting any single proposal or specific text due to the complexity and ongoing discussions.
• ACTION ITEM: The ad-hoc design team (Erin, Kristen, Felix, et al.) to continue refining their "hybrid proposal" text, focusing on:
  • Clarifying the interaction between the "AI Output" category and traditional search results.
  • Elaborating on how AI training (especially for search ranking) relates to output controls.
  • Integrating and mocking up specific use cases to demonstrate clarity.
• ACTION ITEM: Michael and Lea are encouraged to integrate feedback from the hybrid proposal into their Cloudflare proposal, aiming for convergence rather than developing parallel tracks.
• ACTION ITEM: Max Schindler is encouraged to refine his "Terms of Service" pointer proposal, addressing feedback regarding machine readability, handling multiple ToS, conflict resolution, and the proposal's fit within the working group's charter.

Next Steps

• Further refinement of the proposals, particularly the "hybrid proposal," is expected.
• Discussions on these proposals will continue on the mailing list and during the next session.
• The goal is to emerge with a crisp, self-contained, and contextualized proposal for changes to the working group draft.
• The group aims to review progress on these items in the next session to potentially move towards concrete draft revisions.

Session Date/Time: 01 Oct 2025 07:15

AIPREF

Summary

The session began with a continuation of the discussion on whether to include a "top-level preference" term in the AIPREF vocabulary. Due to inconclusive prior polls and persistent disagreements, the group decided to temporarily defer this discussion to focus on more granular, specific preference categories. The latter part of the session concentrated on evaluating existing proposals for categories like "display text" and "substitutive use," particularly in light of new AI-enhanced search features presented as a test case. Key challenges in defining these categories, such as line-drawing difficulties, technical feasibility, and alignment with user and content owner expectations, were discussed.

Key Discussion Points

Top-Level Preference Discussion

The discussion on a top-level preference term highlighted a range of arguments:

Arguments For a Top-Level Preference:

• Opt-out Expression: Provides a valid mechanism for individuals to express a desire to opt out of uses.
• Future-Proofing: Accommodates unknown or rapidly changing uses, preventing a "whack-a-mole" scenario where new preferences are constantly needed.
• Standardization & Completeness: Ensures the vocabulary is complete and provides a standardized way for people to express existing preferences.
• Regulatory Alignment: Potentially aligns with existing regulations (e.g., in the EU) that may require rights holders to express broad reservations.
• Ease of Use: Simplifies expressing preferences for both opt-in and opt-out regimes.
• Marketplace Promotion: Could facilitate the emergence of a marketplace for content creators and AI consumers through bilateral negotiations for content use.

Arguments Against a Top-Level Preference:

• Unintended Consequences: A broad or poorly defined top-level preference might not function as expected and could have unforeseen negative impacts.
• Inhibits Innovation: A blanket opt-out could inadvertently stifle the development of beneficial new technologies or uses. A robust, granular vocabulary is seen as a mitigation.
• Regulatory Misalignment: A generic "opt-out of TDM" may not align precisely with legal definitions or exceptions (e.g., regarding temporary reproductions), potentially creating legal ambiguity.
• Technology-Centric vs. Purpose-Oriented: Existing proposals for top-level categories tend to be technology-specific, making it difficult to align with the actual purpose content owners intend to control.
• Scope & Charter Concerns: A broad top-level category might extend beyond AI-related uses, potentially exceeding the WG's charter.
• Technical/Definitional Impossibility: Difficult to define a truly universal "no, don't touch my stuff" preference that is technically feasible and consistently understood.
• Defaults Matter (Leading Sharp Tools): Broad default opt-outs, especially if applied by CMS platforms without user understanding, could lead to unintended restrictions and a non-meaningful permissions model.
• Extensibility Challenges: A universal "off switch" makes future vocabulary extensions more complex; this might be better addressed by tooling rather than protocol changes.
• Expectations Misalignment: Average webmasters may misunderstand the scope of a broad opt-out (e.g., thinking it includes search).
• Unknowns: The debate involves distinguishing between "known unknowns" (existing but unlabelled uses) and "unknown unknowns" (future, unanticipated possibilities) and how a top-level preference addresses these.

A fundamental disagreement exists on whether a top-level category should be "bounded" (pre-defined scope) or "unbounded" (catch-all for future uses).

Specific Preference Category Proposals & Test Cases

The discussion shifted to concrete proposals, using an example of AI-organized search results (dynamic categories, short generated text explanations alongside links) as a test case.

General Observations:

• Line-Drawing Difficulties: There are inherent challenges in defining categories like "substitutive use," as intuitions about what constitutes "substitution" are hard to formalize into technical definitions.
• Design Team Concept: The use of an IETF design team was suggested as a tool for developing standard proposals for complex issues like "substituted use."
• Current "Search" Definition: The existing draft defines "search" as directing users to third-party content, explicitly excluding "overviews or summaries." This created ambiguity when applied to the AI-enhanced search example which includes both generated text and links.
• "Substitution" Nuance: It was argued that "substitution" is not merely about reducing clicks but also about improving usability by directing users more efficiently, potentially leading to fewer clicks on irrelevant content. However, concerns remain regarding the integrity, accuracy, and potential misrepresentation of summarized content, especially for sensitive topics.
• Innovation & Opt-outs: The argument that opt-outs "stymie innovation" was challenged, with participants noting that innovation can still occur (e.g., through payment for content, academic datasets) and that preferences do not remove content from existence.
• Local Search: The "substitutive use" category also needs to consider local search scenarios where users have lawfully acquired assets and perform device-based searches, rather than only wide web search.
• AI Model Assumptions: Current definitions might be too tied to specific AI model architectures (e.g., pre-trained + RAG). A preference should ideally address the objectionable outcome regardless of the underlying technical mechanism (e.g., if a model generates content that would otherwise be found by reading a document).
• User vs. Publisher Perspective: The importance of considering the end-user's experience (e.g., quick access to relevant information) alongside the content publisher's concerns (e.g., referral traffic, content control) was raised.

Specific Proposals Discussed:

• Krishna's "Display Text" Proposal: This proposal offers granular control over how content is used for summaries, including limiting text length or requiring exact citation. It was seen as offering publishers a choice to participate in features like AI-organized search.
• Bradley's "Substitutive Use" Proposal: This category aims to cover uses where AI models process assets and produce outputs that "replace, reduce the utility of, or make the source asset redundant." Its broad definition could potentially preclude features like AI-organized search if interpreted to cover the generated summaries. Clarification on its scope (e.g., excluding mere ranking) was sought.

Decisions and Action Items

• Decision: The discussion on a top-level preference term will be put on hold for now, to be revisited after progress on more specific categories.
• Action Item: The group will continue to focus on refining the text of specific preference categories (e.g., "display text," "substitutive use") and testing them against real-world use cases.
• Action Item: Participants were encouraged to engage in informal discussions during the extended lunch break to explore solutions for the complex issues raised.

Next Steps

• Continue refining the definitions and scope of specific preference categories.
• Further evaluate proposals like "display text" and "substitutive use" using real-world examples to identify ambiguities and ensure desired outcomes for content owners and users.
• Consider forming a design team if specific issues, such as the precise definition of "substitutive use," continue to pose significant challenges for consensus-building within the main working group.
• The group will reconvene to continue this discussion.