AI Preferences Working Group Materials
AIPREF Working Group Interim Meeting - March 2026
Terminology Issues
Use of âfine-tuneâ in âFoundation AI Model Productionâ category
Phase in question: âFine-tuning can specialise a general-purpose foundation model for a narrower set of use cases.â
Martin: Overly broad is objectionable as overly emotive language. However, this is about the scope of the definition of what we call âAI trainingâ.
Timid Robot: Maybe some folks at AI companies can help us come up with a definition. It should be clear what we are trying to get at, so maybe people can suggest ways to phrase this.
Chris Needham: We need to all aspects of what we have currently called foundational model production. This fits within this definition. I donât think that it is helpful to include all the aspects of production. To the extent that fine tuning is part of the creation of models, it should be included.
Alissa: Something to think about that might help is âwhen does a model stop being producedâ, which might be the boundary condition. Asking what defines production might be answered by asking when the production process ends.
Pedro: There is something that can be done from a technical research point of view, this can mean several different things. But if we worry about just production, it might be easier to talk about whether the model weights are changing. That might be a better framing. There is no single technical term, but weights changing might be more specific.
Replacing: âdigital assetsâ with âdigital contentâ
Martin: Asset often means someone owns something (agreeing with issue description). Maybe too tied to the notion of copyright.
Lila: Agreeing (avoid âassetâ). âAssetâ seems too specific. âContentâ seems better.
Farzaneh: Agreeing (avoid âassetâ). Could lead to complications with asset-level signalling.
Sarah: Agreeing (avoid âassetâ).
Removing normative language: respect, ignore
No comments
Use of âMachine Learningâ
Timid Robot: I have heard this term being used constructively that carries far more meaning than an unadorned âAIâ.
Nate: Also agree. AI is marketing that doesnât mean anything. Without the term machine learning it is too amorphous.
Definition of AI
Martin: Hard to find good definitions. This extrapolates from a bunch of other definitions, with an eye to our goals.
Erik: Until we have a better alternative, then we have to stick with what we have. It would be helpful to have people identify specific failings in the definition, so if it includes things you donât want to include, maybe specific exclusions could be used.
Chris: I think we should have a definition, rather than leave it to interpretation. This is a reasonable attempt to use existing definitions to create our own. This is not far off what a reasonable definition is. But Iâd worry about omission.
Meredith: Having use cases might be useful in moving forward on these. You have to start with use cases and evaluate the definition against them. Things get circular without concrete tests that w can judge against.
Chris: Would be helpful, but willing to help in a limited fashion.
Elaine: Has a decision been made that OECD definitions are not good.
Suresh: If you have proposals please make them; e.g., ITU definitions were discussed and not felt to be to purpose.
Martin: Happy to look at other definitions. What we have is close to the OECD definition; however a few parts of that are weird. Thereâs an academic paper about this, itâs a persuasive analysis. This is captured in another issue.
Alissa: Not sure thereâs broad support for the concern raised here.
Erik Stallman: Thereâs a separate NIST definition. My sense was that the OECD definition was considered; is that correct?
MT: yes. It was one of the main sources; it is a pretty good definition, there were a few adjustments necessary because theyâre so broad.
Training Issues
Quantified thresholds in definition of Foundation Model Production are ambiguous
Pedro: thereâs an issue if we say itâs too small or too large. could be easy to confuse with foundation model and classical search
Martin: all definitions appear to share a common setem, but diverge significantly when concrete numbers are used. Agreeing with Pedro, we should use other means to distinguish.
Suresh: EU doesnât talk about specific numbers.
Martin: yes, but American models do.
Pedro: Talking in terms of parameters is extremely dangerous. Current research is about retaining capabilities, but reducing the number of parameters. Smaller models that are more capable.
Chris: Trying to quantify is likely to be counterproductive. This needs to be usable by creators and publishers. They arenât going to be experts on model creation. They just need to understand the scope.
Alissa: asnât Maybe take the âlargeâ part out. Focus on the operative part: capabilities, etcâŚ
AI Training and Generative AI Training are Too Broad
Martin: overtaken by events. Current draft has âFoundation Modelâ. Donât need to remove âAI trainingâ from 5.2. We do still need to address previously discussed issues on terminology.
Search Issues
Search Category May Lead to Unintended Effects
Caleb: happy to propose a resolution.
Suresh: Krishna has a similar issue. May be able to resolve multiple with a poposal.
Erik: need clarification on what is/isnât allowed by this category
Pedro: reminder that currently there is no non-AI search. only the output differs between âclassicâ and âAIâ search
Search category definition attempts to redefine established search behavior
Alisa: may be worthwhile to use specific categories in issue instead of single category. not sure if it was considered/discussed
Martin: we did discuss, couldnât make sense
Alisa: they make more sense to me
Martin: may be worth more discussion, maybe room for split into discrete categories
Mark: there was concern about additional attachment mechanisms
Meredith: If there is any difference between search and model building: would it be considered neutral/negative if people change their technology to work around definitions. Is that a likely outcome and would that be good?
Pedro: Maintaining two technologies, especially if one is outdated, is not practical. I donât see people doing that to avoid policy. I would see people not indexing content that they couldnât use with their preferred technology.
Paul: It is fair to say that we didnât discuss this in full detail. It is valuable to discuss them. Two points to make: we need clarity for the context: vocab or specific attachment. Some items in Microsoftâs draft are very attachment-mechanism-specific.
Search is too Broad
Mark: already covered in previous discussions
Chris: no further discussion needed (pending proposal from Caleb)
Search should be removed and replaced with display-based preferences
Mark: also previous covered, with more discussion in Toronto
Use Issues
Mark: encourage group to add/update proposals to move this forward
Nate: there must be something to ground RAGâto address summaries
Martin: please provide proposals 1+ weeks before meeting (ideally 2+). this is the central issue, very important for the future of the web
Mark: also see https://github.com/ietf-wg-aipref/drafts/wiki/Use-Proposals
Chris: agree with Nateâthis is fundamental. vocabulary must address RAG, post-training, etc. Misattribution. Can cause reputational damage unless we have some way to cover post-training usage. Not sure on why things stalled and happy to work on something.
Model Issues
Hierarchical structure is problematic and introduces unnecessary complexity ¡ Issue #170
Martin: in limbo pending other category discussions
Focus on purpose of use rather than time of ingestion ¡ Issue #159
no comments
Defaults, Difference between âunknownâ and âallowedâ
Martin: EKR is trying to ârefactor the codeâ; Iâm struggling to identify the concrete reason to change here. Whatâs proposed is reasonable but Iâm not sure thereâs value in doing so. Not substantive. Prefer we defer this.
Context Issues
Introduction clarity re: relation to existing laws
Max: Want to spend a lot of time at April meeting on this. Two perspectives: 1) work we do is isolated 2) work is more contextual. We need more time to get into nuances.
Distinguishing access and use
Martin: we havenât been clear in our intent. there are two actions involves with robots.txt: 1) existing crawl directives 2) AI Pref only apply to downstream uses of âstuffâ
Timid Robot: much of our conversations have focused on access. we may need more documentation about how preferences travel with âstuffâ so they are available to downstream users
Martin: will oppen an issue to track that
Lila: concern about bundling preferences with âstuffâ (or concern about it not be adequately bundled). Thank you Martin.
Victoria: echo what Lila said. the importance of clarity is hard to overstate for users, if not done well will have chilling effects
Jo: what we did in ARDC standard, was to specify it should be carried with.
3.1 Conformance
EKR: is non-conformance possible with volunteer document
Suresh: itâs not about compliance
Mark: if an entity claims conformanceâseparation of conformance to processing and conformance to preferences
Jo: my concern is that it pertains to enforcement. if there isnât an enforcement mechanism built in, it should be obvious
Paul: there might be some misunderstanding about what layer this applies to. this is about conformance of attachment layers to the vocabulary. it is not about enforcement. (compliance vs conformance)
Elaine: i found it confusing because there are very few MUSTs and MUST NOTs.
Mark: ask for proposals that clarify
3.2 Respecting Preferences
no comments
Bots Collect Data for Multiple Purposes
Mark: appears to be related to use
EKR: may be fine, now, the way it is, given the direction it is moving
