Meeting minutes
meeting RQTF meeting
Accessibility of Machine Learning and Generative AI draft.
janina: No news yet from me
jasonjgw: OK, that's the first part, but there's more ...
jasonjgw: We should integrate discussions of upcoming mtgs, incl TPAC which Matthew has joined to discuss
<matatk> Proposed issue for TPAC: w3c/
jasonjgw: Recalls attending a telecon of Web AI CG ...
jasonjgw: Useful to coord ...
matatk: posted APA issue -- though that's general
<matatk> This is the next Web & AI IG meeting, to which we were invited: https://
matatk: Believe the invite can be extended to RQTF
matatk: May be a conversaion away from TPAc time, as WebAI will not be physically at TPAC
<matatk> Web & AI IG discussion threads that explicitly mention accessibility: https://
matatk: mtg will be Jul 13 at 1500 UtC
matatk: a few threads open:
jasonjgw: reviews our interact with WebAI ethical doc
matatk: Areas of identified interest are listed in WAI issue github
matatk: ATAG also interested in conversation with us, MA will be joining to coord
janina: AI represents a profound change in human consciousness - like the A-bomb before it. We need to get back to what the W3C context is. AI is creating a lot of content.
… We are interested in ethical considerations. We develop two kinds of deliverables... is there something about AI where a normative standard would be valuable? We would be interested in accessibility considerations in any normative spec.
… the biggest one I know for sure is just a warrant for what ATAG is about to do. We can tell the industry: we know it's going to generate code; it should all be accessible (to the best we can define that).
… Education use cases are very strong ones.
… There are a few areas in best practice (Note) docs that we could work on. I am curious as to what that list is, about where accessibility really matters in specs.
matatk: Interesting issue
My team found this issue this morning: anthropics/
matatk: presumably a function of the content it trained on that it's biased against accessible code
matatk: Asks nature of our AI doc? AUR or high level principles?
matatk: If the latter how do we get more views? Pub in their space?
matatk: Other approach is to consider from their perspective -- asking these questions and learning their view
matatk: Notes we share Roy as Staff Contact with them
jasonjgw: Agree approach MA gives is correct
jasonjgw: our doc is a kind of issues paper -- the central challenges and issues between web a11y and AI
josh: huge subject -- not inherently good or evil, but we should set expectations of what "good" looks like
josh: Most people using AI won't have an understanding of a11y and what concerns there might be
janina: Agree mostly with Josh. I would give it a slightly different spin in that it's for informaing future W3C accessibility work. We want to name the stuff that's coming at us, and frame it in a way that gives us a basis, with some group agreement, as to how to think about it and move forward with it.
… I wouldn't rule out an AUR, though this initially is not that.
… If a group forms to talk about how Agentic AI could make tailored solutions, that must have a strong accessibility considerations.
… Thinking like Adapt TF's work... instead of being declarative, we could have tailored solutions for the user. Could generate more potent AT than we have had to date. That's where we'd need an AUR. (Here's what people who are blind, or deaf, or ... would need to ensure it's accessible.)
… We're building stuff for people and we're leaving people out.
josh: Agree. We don't want these systems creating token accessibility efforts that are functionally useless.
janina: Would like to let them know on the 13th that this is where our thinking is going.
… about how to inform W3C thinking in different groups/levels
josh: Asks how this relates to TAG generated docs ...
matatk: Three docs so far ...
matatk: web platform design principles
matatk: Ethical Principles
matatk: This one endorsed by U.N.
matatk: Privacy Principles --
matatk: Then Web ML WG has been working on Ethical Principles -- and that's where we commented
matatk: Concept of "burden shifting" came up there
matatk: Addresses various concerns incl content ml trained on
matatk: Think we have limited time and we're just one item on the agenda on the 13th
matatk: Offers to synthesize for best effect
matatk: Phps Roy can add RQTF to mtg Calendar
jasonjgw: Agree. Approach appropriately pragmatic!
jasonjgw: Happy to respond on drafts
jasonjgw: Interesting scope q in ML identifying security bugs -- some extant for decades
jasonjgw: Phps an approach that identifies a11y?
Re authoring tools, here is our TPAC session proposal: w3c/
janina: I like your challenge about using AI to identify accessibility issues. I think this should be important to WCAG 3...
josh: Will there be an ATAG 3?
janina: Looks like it!
… Jutte is still involved, which is great.
josh: The point about AI being biased against making accessible content - good point, how can we expect it to make it accessible based on what we've seen. (ref matatk re training data, etc)
janina: agree
jasonjgw: Yes, it will reproduce the patterns from the large corpus it's been given
… Not aware there's an easy solution for this problem, as it will inevitably reflect the training data.
… A large proportion of what's out there will have accessibility issues. Do they have ways to tune models or overcome issues?
janina: What about 'take the data you're trained on and evaluate it according to WCAG 2'
jasonjgw: Given my understanding of neural networks, that's a hard problem.
… need an ML expert to say what strategies are available to mitigate those issues.
janina: Excellent reason for W3C to be in this space.
jasonjgw: There are some ways to tune the output but I'm not sure to what extent this is possible.
matatk: Disclaims ML expertise, but has one on staff!