AI Preferences Working Group Materials
AI Preferences - IETF 123 Minutes
Review of Drafts / Issues
- Aim to submit in August
- Still work to do
Reminder of the charter
The AI Preferences Working Group will standardize building blocks that allow for the expression of preferences about how content is collected and processed for Artificial Intelligence (AI) model development, deployment, and use.
Multiple stakeholders and parties involved
First WG meeting was in Bangkok
Some hybrid in-person and virtual meetings since
Some issues have been filled against the Draft
People are coming from multiple verticals: content holders, AI builders, etc.
Try to do something that works for everybody, not looking for something perfect but good enough to be shipped quickly
There are some areas where more discussions are needed but we will have time during this session
Vocabulary Draft
(Martin Thompson)
There are not many issues left
Issues are used by the Editor to track what needs to be done
Closed issues
- Taxonomy: covers all form of content processing
- AI definitions:
- Search and Discovery:
- Hierarchy: defined the top level types of processings
- Eurocentricity: different back and forth on some elements of terminology
- Extensibility: Discussions on ability to expand the taxonomy - have a registry of terms
- Least privileges was selected as a mean to control the preferences references via different mechanisms
- Discussion on ability to ignore expressed preferences (i.e Adobe service using images as input). Adobe wants to enforce ability of the owner of a picture to say no to processing. Lead to discussions of EU Copyright laws
- refined the labels and terms
- Recognition that there are some terms that spelled differently but express the same notions as the terms used inside this work
- Extensions of preferences with the ability to defined new functions need
Open Issues
- Lots of discussions on inference
- Normal inference of Generative inference
- This discussion continues on
- Lots of discussion in content in Dynamic data retrieval for prompts
- one problem is defining higher levle Categories is ok as it cana always be refined in lower categories later. The reverse is not true
The Drafts are not final and are still work in progress
Alisson Cooper: If you cannot sort out Translation when it will have an impact on Generative processing
R: Translation is accessibility and therefore can classified through that
R: We need examples to establish the differences of definition
R2: In trading verticals the point cames with some details on the differences
Q: Search and Generative are not crisp definition in the document, would be great to come on conclusions for some
R: Agree to look for more definitive
Arnaud Taddei: have you look at other SDO what there are doing. OP Can provide some links to those
Erum Welling: We need to have contracts on what has been consented. Those contracts could be used to overturn the AI preferences on the usage of data.
R: Cannot say that all the contracts in the world can override those preferences. There are also Juridictions laws and regulations.
Q: Question about accessibility, you talk about Augmented retrieval. How would those preferences apply?
R: There is a model for preferences, but for accessibility it will be an implementer point. We don’t have anything there yet. OP can join and contribute
R2: The group has for objectives a framework than a code you can runé
R3: Call for contributors for adding more to the Draft
Q: Do you aim to have some source and legitimacy of why a search is made?
R: There were some discussion on Transparent Audit but this is outside of the scope of the document.
Rony Shalit: There are a lot about the questions about the Robot.txt but it is not a robust mechanism. GenAI transforms the internet.
R: the comment you make is more about the Attachment Draft
Q: (Continuing) Opt-in, Opt-out are creating some kind of discrimination. There are good boot and Bad bot. Also lots of GenAI with lots of money will be able able to be accepted to have access to smaller organizations won’t have this ability.
Q: Robots.txt is not updated, poorly administrated
R: You are right. This might be best to happen at other WGs.
R2: opt-in / opt-out have some values but they are not sufficient. They need vocabulary and tools
Q: We need to be careful about everything happening around
R: Best continue the discussion but outside here
D: Alissa Cooper: With the new AI-user and AI=Search categories, what happen is fht consumer AI Preferences set are not respected.
R: This Draft is not for fine grained controlled environments.
R2: yes there is a case where the processing can modify the preferences on behalf of the user. We need to look into that.
Andrew Campling: Question about authenticating agents
R: See the BoF right after
Attachment Draft
Closed issues
Definition of a default preference if you cannot get one.
Precision in between time of acquisition of the AI Preferences and the use of the AI Preferences
Open Issues
Normalizing the different to attach Preferences
Especially for Robots TXT syntax
Jeff Lombardo: How those attachement cover the continuous control of the attachment in multiple hopes
R: Not yet we focus on the first two attachement for the first leg
R2: You are talking about the case where the requestor of the data is not the crawler of the data. We will cover that in the next iteration
Tommy Jenssen: Can we me more stringent in the expression of expectations instead of yes\no
R: Please provide a proposal
Alissa Cooper: What about the fidelity and the referencing of data in the generation of outputs
Franck Brockners: how to propagate through MCP
R: yes it can be done but not in this first release
