– DRAFT –
Web & AI IG meeting 09/02/206

Check the regluar Web & AI IG meeting time

w3c/webai#4

Fabien: Max had this proposal to have a rotating meeting time

<mala> +1 to rotating time slots

Fabien: leave your feed back in above issue

<Roy_Ruoxi> +1 from Janina

<Paola> for me this time is a bit late, but it could be worse, so for me it depends which way the rotation goes. whatever works for everyone

<Paola> how can I help to scribe?

<rtroncy> Rotating time slot is more fair for people around the globe

JenStrickland: his is my first time coming to this, it's great to have global people to join

<mgifford2> +1

<Henny> +1 to rotating

<mgifford2> Rotating makes sense

Fabian: anyone against rotating
… hear none

<rtroncy> Link to the issue: w3c/webai#4

<mgifford2> I'm going to have to jump off in about 15 minutes. Sorry.

AI ethical principles

AI and Accessibility

<matatk> I'll recap my AC2025 talk, the video and transcript you can find here: https://www.w3.org/2025/04/AC/talk/atkinson#talk

<matatk> The slides are here: https://www.w3.org/2025/Talks/AC/lt-accessibility-matthew-atkinson/

<matatk> After this I'll talk about a few developments that have happened since, which have relevance for accessibility, and we can move into a discussion about them. These 3 areas are: discussions around accessibility metadata and AI; further exploration of the agentic paradigm; and national standards work around AI and accessibility.

Slideset: https://www.w3.org/2025/Talks/AC/lt-accessibility-matthew-atkinson/

[Slide 2]

matatk: https://www.w3.org/2025/Talks/AC/lt-accessibility-matthew-atkinson/

matatk: All that stuff that we're calling Ai is a tool.

[Slide 4]

<JenStrickland> Please mute if you're not speaking. There's a random noise that keeps happening. Thanks!

[Slide 5]

matatk: We have horizontal reviews—there are groups that review upcoming proposals for accessibility, internationalization, security, Privacy and general architectural, fit for the web platform.

[Slide 6]

[Slide 8]

[Slide 10]

[Slide 12]

[Slide 13]

<JenStrickland> Thank you so much, Roy!

[Slide 17]

[Slide 21]

matatk: The 1st one is. Discussions around accessibility, metadata and AI
… we realized that machines could really benefit from this sort of information, as well sort of same posting. but we do need to be very careful that we don't misuse these rich set of accessibility attributes that we have for content.
… 2nd generative AI

<matatk> https://ourtake.bakerbotts.com/post/102lx09/canada-releases-accessibility-standard-for-artificial-intelligence

matatk: everything generative, but we need careful about accessibility and hallucinate
… 3rd, at the national level of around standards relating to AI

<fabien_gandon> ack

anssik: talked about webMCP
… also need talk this together about accessibility

janina: when I think about AI, I think about patterns.

<anssik> WebMCP accessibility considerations

<Zakim> fabien_gandon, you wanted to ask How to make the provenance accessible? Is there work on how to know at anytime and across changing contexts the provenance of every element of context and in particular if it was generated?

janina: AI is essentially a powerful pattern matcher that rapidly processes large amounts of information and generates new content based on those learned patterns and our instructions. In a private, on-device, agentic context, it could act as a personalized assistant，observing how a user interacts with a system and offering timely, opt-in guidance to

improve efficiency, especially when adapting to new tools or environments

<JenStrickland> When I speak I will have a response to fabien_gandon's question.

fabien: is there anyone working on the accessibility of the provenance？

<Zakim> tzviya, you wanted to respond to fab

tzviya: don't have particular knowledge of a group working on this, but I I'm sure about the the work happening in the ietf with ai preferences and things like that.
… this is just metadata at this point. I think it's actually a great suggestion for this group. this is something that should be done by browsers.

JenStrickland: I worked a lot on AI assurance, Whatever is being presented so incorporating principles of AI assurance in the work would be a way of helping to evolve this and be able to surface the accessibility.

<matatk> Concerned about flagging things as AI-generated - is it solving the right problem? And would it actually work? For example...

<matatk> Do we really want to know if it's AI-generated, or whether it's accurate? E.g. human may fail to fill in alt text for an image, or do it badly. AI may do it well or badly.

<matatk> If we have an attribute to indicate something is AI-generated it may give false impression that something that _doesn't_ have the attribute is _not_ AI-generated, but it could be.

AI ethical principles

anssik: presentation about Ethical Principles for Web Machine Learning
… background: WebML launched at 2021
… publication from WebML WG
… web machine learning allows model run in client or cloud
… use cases: person detection, facial recognition etc,
… nowadays, generative AI

<JenStrickland> MITRE's publicly available AI Assurance materials: [a] https://www.mitre.org/focus-areas/artificial-intelligence/ai-assurance [b] https://www.mitre.org/news-insights/publication/ai-assurance-repeatable-process-assuring-ai-enabled-systems [c] https://www.dmi-ida.org/knowledge-base-detail/A-Framework-for-the-Assurance-of-AI-Enabled-Systems [d]

<JenStrickland> https://wbjournal.com/article/umass-chan-to-open-ai-assurance-lab-to-validate-use-of-the-technology-in-health-care/ [e] https://www.mitre.org/news-insights/news-release/mitre-opens-new-ai-assurance-and-discovery-lab [f] I can't locate the AI Assurance Knowledge Base. It may have been cut under the current administration, or just become not publicly

<JenStrickland> noted.

anssik: running ML workflows on the client enables large-scale deployment without costly cloud infrastructure, benefiting small and long-tail developers while aligning with decentralized web principles, reducing single points of failure, and lowering latency for use cases like immersive experiences and accessibility.

It also improves privacy by keeping data local in the browser, though it raises ethical considerations that the working group aims to address.
… Ethics is about what is right and wrong, and technology is never neutral,it always carries social and ethical implications. Failing to consider those implications can cause real harm, so even if we can’t solve everything, we have a shared responsibility in the web ecosystem to act ethically and improve the situation.
… current ethical issues address in our document, and we borrowed principals from UNESCO
… we now have the deliverable the ethical framework doc as W3C note
… We had established this doc as a joint deliver between the WebML WG and this IG. more feedback and review needed.

<Zakim> janina, you wanted to disagree --It's not the tech that has ethical value, it's the use humans make of the tech. Example: fire can warm you and cook your food, or be used to terrorize by burning a city like Los Angeles

Janina: talked that technology itself isn’t inherently ethical or unethical, it’s how humans design and use it, as shown clearly by accessibility being a result of design choices that can always be changed.

<Zakim> fabien_gandon, you wanted to ask how you weave the principles with the daily technical work of the WG ?

anssik: The ethical principles is a separate doc that is referenced.
… Ethical principles are defined in a separate document referenced by technical specs like the Web Neural Network API, and while the group has recently focused on technical work, there’s strong interest in revitalizing and maintaining this ethics deliverable with help from experts in the Interest Group.

anssik: you can find other resource in our github repo

<janina> I need to leave, but look forward to reading and making suggestions to the ethical docs. Will also reference them in our work. Thank you!

anssik: We have Demos and and a lot of sample code and all that available for developers to use.

<anssik> webmachinelearning/webmachinelearning-ethics#21

mihai: could you share the the guidebook?

anssik: linked share above