AIPREF

Summary

The AIPREF Working Group held an interim meeting on August 24, 2026, to discuss key issues and advance draft-ietf-aipref-vocab. The discussion focused on refining the definitions of the "Search" and "AI Training" categories, addressing the non-normative context in Section 3.2 regarding public interest and legal frameworks, and exploring a new "AI Input" category for Retrieval-Augmented Generation (RAG).

The presentation materials and connectivity links for this session can be found in the Webex Link slide deck.

Key Discussion Points

1. Search Category (PR 201 & PR 202)

The working group reviewed the proposed changes to the "Search" category in draft-ietf-aipref-vocab.

Non-Substantive Changes & Summaries: The participants discussed the inclusion of "non-substantive" modifications. Christopher Flammang raised concerns regarding how to clearly differentiate non-substantive changes from summaries, noting that the relationship between excerpts, translations, and generative summaries needs clearer guidelines. Michael Jones requested further clarification on whether the category covers single-result summaries or set-of-results summaries.
Trained Model Scope: Henk Birkholz highlighted the risk of search providers using publishers' content to train broad AI models under the guise of providing search. He emphasized that any model training permitted under the search category must be strictly constrained to the provision of that specific search service.
Excerpts Wording: To address concerns that the phrase "only excerpts... are displayed" was overly restrictive for UI elements (like links or styling), Christopher Flammang and John Mueller proposed refining the language. The group favored a formulation where displayed excerpts "serve to assist users in evaluating the relevance of the results" rather than strictly limiting the display to excerpts.
Defining Search Applications: Christopher Flammang and Michael Jones debated whether "search applications" must strictly be defined by user queries, pointing out that implicit queries or feed-based discovery (e.g., news feeds) should be accommodated. Henk Birkholz cautioned against making the definition too broad (e.g., inadvertently covering electronic program guides or recommendation systems).

2. AI Training Category (PR 205)

The group discussed renaming the "Foundation Model" category to "AI Training" to better reflect its technical scope.

Definition Refinement: The draft definition was simplified to "the act of using an asset in the production or refinement of an AI model."
Parameters & Weights: The group agreed to remove specific references to "weights" or "parameters" to keep the vocabulary future-proof and avoid technical loopholes.
Explanatory Note: John Mueller suggested adding a non-normative explanatory note clarifying that, in practice, any use of an asset that has the effect of changing or refining a model (including fine-tuning, Low-Rank Adaptation (LoRA), or third-party refinements) is encompassed by this category.

3. Section 3.2: Status of Preferences and Context (PR 206)

The working group debated PR 206, which introduces non-normative text on how preference expressions interact with legal and historical contexts.

Historical Context Paragraph: Christopher Flammang and Henk Birkholz strongly objected to the "historically" paragraph, arguing it was confusing, incomplete, and too closely aligned with US-centric legal concepts (e.g., fair use). They suggested either removing it entirely or significantly simplifying it.
Public Interest Considerations: Michael Jones and Suresh Krishnan noted that highly risk-averse organizations (such as those conducting academic research, web archiving, or accessibility work) need clear guidance that the vocabulary does not override legitimate public interest or statutory legal exceptions.
Neutrality: The sense of those present was that the document must remain strictly neutral, avoiding both the appearance of granting a license and the implication that existing legal exceptions are clawed back.

4. AI Input & RAG (Issue 172)

A preliminary discussion was held on creating a category to cover Retrieval-Augmented Generation (RAG) and model grounding, proposed as "AI Input."

Scope of Input: The proposed category aims to distinguish between automated background retrieval (scraping/RAG) and direct user-initiated uploads of specific assets.
Bypassing Site Controls: Henk Birkholz raised concerns about users bypassing site-level crawlers or robots.txt controls by pasting URLs directly into models, complicating how publishers express usage preferences.
Refinement vs. Inference: John Mueller and Michael Jones pointed out that some RAG implementations feed user interactions back into training loops (backpropagation). The group agreed that if a RAG process results in model training, the "AI Training" preference must also be respected.

Decisions and Action Items

Decision: Agree to merge the "Search" category changes (PR 201/202) and "AI Training" category changes (PR 205) into the editor's draft. These sections will be flagged as lacking final working group consensus to allow for further public tracking and iteration.
Decision: Merge PR 206 (Section 3.2) into the editor's copy to serve as a baseline, with the understanding that the "historical" paragraph will undergo further revision.
Action Item: Christopher Flammang and Henk Birkholz to collaborate on a simplified, neutral proposal for Section 3.2 that addresses public interest overrides (such as archiving and accessibility) without US-centric historical framing.
Action Item: John Mueller and Henk Birkholz to draft concrete use cases and initial text for the "AI Input" / RAG category under Issue 172.

Next Steps

Interim Meetings: The working group plans to hold an online interim meeting in June 2026.
Upcoming Hybrid Interim: The group is targeting the week of August 24, 2026, for a hybrid interim meeting in Europe/UK.
Hosting: Suresh Krishnan requested that any organizations willing to host the August interim submit facilities proposals within the next two weeks. Default fallbacks are available if no external host is secured.

Automatic IETF Minutes