Session Date/Time: 03 Mar 2026 21:00

AIPREF

Summary

The AIPREF interim meeting focused on reviewing the working group's GitHub issues list to assess their status, identify potential dispositions, and encourage participants to propose solutions or refinements ahead of the April interim meeting in Toronto. Key areas of discussion included terminology, definitions related to AI and foundation models, distinctions between access and use, and the need for addressing post-training AI activities like RAG (Retrieval Augmented Generation).

Key Discussion Points

Administrivia and Logistics:
- The meeting adhered to IETF Note Well policies, including recording.
- The agenda centered on progressing the issues list for the upcoming April interim meeting in Toronto.
- Registration for the Toronto interim is nearing capacity, with remote participation available. Participants were encouraged to register promptly.
- Eric Stallman volunteered to take notes, with others offering to assist. Note-taking via Data Tracker login was highlighted.
Terminology Issues:
- Issue 179: Use of "fine-tune" in foundation AI model production category:
  - Krishna's concern regarding "fine-tune" as unnecessary, imprecise, and overly broad was discussed.
  - Martin suggested this is a scope discussion tied to the overall definition of AI training, best addressed in April.
  - A participant noted the common understanding of fine-tuning for specializing general-purpose models.
  - Chris Needham felt "fine-tuning" fits within "foundation model production" and suggested defining the boundary of "when a model stops being produced."
  - Pedro proposed a more technical definition focused on "are the weights of the model changing?"
- Issue 163: Replacing "digital assets" with "digital content":
  - Joe proposed "digital content" to avoid "asset" implying ownership (e.g., related to IRS cryptocurrency definitions or copyright).
  - Martin, Leila, and Faris agreed, arguing "asset" is too specific and "asset-level signaling" could lead to IETF enforcing copyright, which is out of scope.
  - A sense of those present indicated support for the change.
- Issue 162: Moving away from "respect" and "ignore":
  - Joe suggested removing anthropomorphic language. This was deemed an editorial change.
- Issue 152: Use of "machine learning" in AI training definition:
  - Ecker suggested removing "machine learning" from the definition of AI training.
  - Timid Robot and Nate Hick argued "machine learning" is specific and meaningful, whereas "AI" is often a general marketing term.
- Issue: Definition of AI (related to Ecker's comments on breadth and "sufficient complexity"):
  - Ecker found the current AI definition too broad, potentially sweeping in any statistical technique, and "sufficient complexity" subjective.
  - Martin explained the current definition is a synthesis of legal and standards organizations' definitions, with some adaptations to avoid overly broad interpretations. He suggested that explicitly not defining AI, or identifying specific use cases to exclude, could be options.
  - Eric Stallman and Chris Needham supported having a definition, but suggested providing concrete examples of what the definition should exclude to refine it.
  - Meredith emphasized starting with use cases to evaluate definitions, offering to help draft them. The Chairs supported creating a GitHub wiki page for use cases.
  - Elaine inquired about the group's stance on existing definitions from OECD, NIST, or ISO. Martin clarified that none were fully fit for purpose, and the current draft's definition draws from OECD but refines it.
Training Issues:
- Issue 191: "How large is large" in foundation model production:
  - Alyssa questioned the ambiguous terms "large" and "very large" in the definition.
  - Martin highlighted the diverse and rapidly changing thresholds for model size in various definitions (e.g., billions of weights).
  - Pedro noted that research is trending towards smaller, more capable models, making size thresholds quickly obsolete.
  - Chris Needham suggested a use-case driven approach rather than quantification.
  - Alyssa clarified she was not suggesting quantification but ambiguity; proposed removing "large" and "very large" and focusing on the model's applicability to "a wide range of use cases." A poll of the room indicated agreement with removing the quantifying words.
- Issue 165: "AI training" and "generative AI training" are too broad:
  - Ecker expressed concern that "AI training" is too broad and "generative AI training" is also too wide.
  - Martin suggested this issue might be "overtaken by events" given the focus on "foundation model training," but the Chair noted Ecker still considered it relevant. Discussion remains on defining "AI training" as a fundamental component.
Search Issues:
- Issue 188: Search category may lead to unintended effects:
  - Caleb expressed concern that the current definition does not reflect modern search engines. He offered to flesh out a proposal with examples to address the perceived "false dichotomy" between traditional listings and generative AI.
  - Pedro cautioned against defining "traditional search" as not using AI, as modern search engines leverage AI models similarly to generative AI, with the distinction often in presentation (indexing vs. generated text).
- Issue 177: Verbatim Match:
  - Krishna's issue on the "verbatim match" language in the draft was discussed.
  - Alyssa suggested re-evaluating Krishna's proposed display-based categories as alternatives for the search definition. Martin recalled previous discussions on Krishna's draft in Montreal that did not yield actionable consensus, partly due to the broad applicability of the framework (robots.txt, content attachments).
  - Meredith raised a structural question: is it neutral or negative if IETF definitions force technology changes (e.g., bifurcating products)?
  - Faris and Paul believed Krishna's proposal merits further thorough discussion at the Toronto meeting, emphasizing the need for use cases to evaluate proposals.
- Issue 160: Search is too broad:
  - Chris Needham reiterated concern that the current definition of search (generating text/links) includes current AI outputs, making it hard to distinguish from generative AI chatbots. The Chair asked if control is needed if AI and non-AI search results are indistinguishable; Chris needed more thought.
- Issue 176: Search should be removed:
  - This issue from Krishna was considered largely duplicative of other search-related discussions.
Use Issues: Missing category for RAG and grounding (Issue 190, Chris Needham, and Brad Silver's proposal):
- Participants expressed strong agreement that a vocabulary category for post-training AI activities (like RAG, grounding, summarization) is fundamental and missing. Previous proposals had not gained consensus.
- Nate Hick emphasized the critical importance of this, offering to work on a proposal. He highlighted the "future of the web is at stake" due to implications for monopolistic search engines and content creators.
- Martin requested proposals be submitted ideally two weeks (minimum one week) before a meeting to allow adequate review time.
- The Chair offered to capture previous use-related proposals and their identified issues on the wiki. Chris Needham offered to help with this.
Overall Model Issues:
- Issue 166: Hierarchical structure is problematic:
  - Martin stated that discussion on hierarchical structure is deferred until top-level categories are more clearly defined.
- Issue 149 (Ecker) / Krishna's display-based preferences: Focusing on purpose rather than use:
  - This issue, overlapping with previous discussions on Krishna's draft, highlighted the need for concrete proposals and discussions about their identified issues.
Defaults Issues:
- Issues 154 (Ecker) & 153: "Unknown" outcome and difference between unknown/allowed:
  - Ecker questioned the "unknown" outcome, preferring a self-contained system (yes/no).
  - Martin viewed Ecker's proposal as code refactoring, cautioning it might foreclose options for handling unknown preferences. He suggested deferring this discussion until categories and structure are clearer.
Context/Conformance Issues:
- Issue 192: Relation to existing laws (Max):
  - Max emphasized the need to clarify that the framework operates "without prejudice to legal norms" (e.g., terms of use, privacy) and cannot be used in isolation. He suggested significant time be allocated for this discussion in April.
- Issue 167: Distinguishing access and use (Martin):
  - Martin clarified the intent to distinguish between access (crawling governed by robots.txt) and downstream use (where AI preferences apply). Editors will sharpen the language. He requested more asynchronous engagement on the mailing list.
  - Timid Robot and Alyssa highlighted the complexity of AI preferences "traveling" with content for downstream users, especially with conflicting preferences or non-owners, and the difficulty of record-keeping. Martin added an issue to track this.
  - Victoria stressed that ambiguity in this area could chill downstream uses and innovation. Joe mentioned ARDC standards that address careful data collection for downstream users, acknowledging diverse purposes.
- Issue 164: 3.1 Conformance (Joe):
  - Joe questioned the necessity of the "Conformance" section, arguing that if implementation is optional, "non-conformant" is hard to define meaningfully, and it risks implying enforcement.
  - Ecker (interpreted by Chair) distinguished between conformance to specification (e.g., processing preferences as per spec) and conformance to preferences (what one does with the results).
  - Paul clarified that conformance refers to implementations of attachment mechanisms with the vocabulary, not enforcement. Elaine noted the few "musts" and "must-nots" in the draft make traditional conformance hard to assess.
- Issue 193: 3.2 Respecting preferences (Alyssa):
  - Continued refinement of section 3.2 is desired. Participants with specific proposals or concerns were asked to record them.
- Issue 158: Bots collected for multiple purposes (Ecker):
  - This issue touched on the distinction between collection time and use time. The Chair felt the existing text might be sufficient now given current discussions, but encouraged proposals for further progress.

Decisions and Action Items

Decision: Remove ambiguous quantifying words ("large," "very large") from the definition of "foundation model production" (Issue 191). Editors to implement in the next draft.
Action Item: Chairs (Mark/Suresh) to initiate a GitHub wiki page for collecting use cases to help evaluate definitions, with Meredith and Chris Needham offering to contribute.
Action Item: Martin to link relevant academic papers and discussions regarding AI definitions to Issue 192 (Relation to existing laws).
Action Item: Chairs (Mark/Suresh) to ensure a set of agreed-upon use cases are available before the Toronto meeting, and to re-engage Krishna on his proposals.
Action Item: Chairs (Mark/Suresh) to capture previous "use-related" proposals and their identified issues on the wiki.
Action Item: Editors to sharpen language distinguishing "access" and "use" (Issue 167).
Action Item: Martin to add a new issue to track the complexity of AI preferences "traveling" with content for downstream users.
Action Item: Participants interested in clarifying conformance language (Issue 164) are encouraged to propose specific text.
Action Item: Participants are strongly encouraged to provide proposals (ideally two weeks, bare minimum one week) and engage asynchronously on the mailing list and GitHub issues to move work forward.

Next Steps

Shenzhen IETF Main Meeting (one hour): A status update will be provided to the wider IETF community, summarizing progress.
Toronto Interim Meeting (April): This meeting is scheduled for more in-depth technical discussions and to make sustainable progress on outstanding issues, building on the proposals and use cases developed beforehand.
Ongoing: Continuous asynchronous engagement on the mailing list and GitHub issues is crucial for progress, including refining definitions, addressing concerns, and developing concrete proposals for the Toronto meeting. Nate Hick and Chris Needham committed to working on the missing RAG/grounding category.