Meeting minutes
<scott_h> presen
Accessibility of Machine Learning and Generative AI draft.
Accessibility of Machine Learning and Generative AI draft.
JW: There are two tracks - the publications under dev and then the issues Janina is looking at
There is also the publication itself and WebMPC
There were issues raised and things that need to happen there.
Lets start with Janinas list
JS: This has ended up being helpful. I've got a summary of the conversations about our paper. Going back to Jan.
The end result is what I've posted and names the issues we identified
They are chronological - there are 43 points.
Some are obvious and would be good as a TOC
Some may not be appropriate e.g. some stuff on overlays
An argument is that you cant avoid talking about post source
There are conversations around how we define and describe accessibility
I have some questions - so we could go through them?
JW: Responding to detail would be good on the list.
JOC: Have we a URI?
<janina> https://
JS: I will categorise these
9,10,11 could be things we may need to get our head around
16 -19 we need to think about more
21,22 are out of scope
9. Defining the scope of the work is difficult: whether to address neural-network machine learning narrowly or "AI" broadly, and how to signal which claims apply to which kind of system.
10. Rule-based, deterministic AI and neural-network machine learning behave very differently, and the variable, hard-to-predict output of the latter is what matters most for disability.
11. Discussion of AI risks getting lost in background and definitions rather than telling users what the technology does for them and how it affects them.
16. If users interact with the web only through agents, source websites may shrink to API-only endpoints, eroding the large accessible surface of today's web.
17. Control over branding and presentation shifts into a negotiation between content providers and agent providers rather than remaining with the user.
18. It is unclear what is actually being licensed, in copyright and intellectual-property terms, when agents mediate and repackage content.
19. In an agent-mediated web, remediation becomes the preferred accessibility strategy, with human intervention as a costly and limited fallback.
21. The large data centers behind AI carry significant environmental and economic costs.
42. Access to the benefits of AI is not equitably distributed.
JS: 17 I think is just wrong
SH: Are you happy if we just comment on the list?
JS: Please do
JW: Lets do that - best on list
JW: Some of the Out of Scope issues arise from the editors draft
JS: Great, we have made some progress.
secodn point - webMCP proposal
WebMCP Proposal (Continued)
Matthew: APA has a tracking issue for this
accessibility tree should be vehicle?
JW: There are updates to that thread
<matatk> APA tracking issue: w3c/
MA: I'll paste into the tracking issue
<gives overview of tracking process>
MA: The a11y is the one I've linked to comment from the Web Machine Learning repo
I've drafted a potential comment to this question - I will mention it in the APA call and ping RQTF
It needs some editing - but captures what we are saying - captures comments from Leonie and James C
There are things there to make accomodations - but the A11y tree may not be the right place to do this
Leonie and James speak to this
Note, even if this is right place - the a11y tree - this will break existing interoperability
They could create a parallel a11y tree - that wont break existing interactions
JS: Just to suggest for the minutes - as an experiment that makes sense - is there any reason not to keep parallel structures in the future?
MA: There is no suggestion that they will do this.
The question is, is it related?
I've not seen all comments - but Leonie and James thing this is different kind of problem.
MA: I'd like to see more semantics available but this WebMPC is more abstract
Doesn't follow the same rules - separate problem.
They likely have two distinct solutions as they are two different problems.
JW: This MPC info is independent of the UI
JS: We do like declared semantics - not likely to change
JW: Thats right.
There is a question around how LLM will use it when it parses declares semantics on the web, potentially powerful with agentic
JW: Semantics are of course useful - I agree with Matthew
JS: What is the expected APA path forward?
MA: Thank you for that Janina - your point last week. I will point to the fact we discussed it in RQTF
MA: The proposal is should be MPC thing use the a11y tree - we think no. Cloning may not help but a slightly more deterministic proposal may be better.
We will talk in APA today and come up with something end of the week
Will ping RQTF later et
MA: Hearing no objections will ship it shortly after
<DIscuss the logistics>
<matatk> Web and AI IG meeting (APA & RQTF are invited): https://
<matatk> Next Tuesday
<matatk> (or otherwise!)
Media accessibility, coordination of work.
JS: I think this is under control. We had to cut short the convo last week
JW: Opening this up for comments
Game accessibility.
JW: We are narrowing scope on an RQTF contribution
JW: We didn't complete convo last time, John where are we at?
JP: The question is how to identify what is necessary for a game to be accessible
It is too much in some situations - can it be nailed down and separate a problem from an a11y issue?
Needs debate and discussion
This is key need for a new document
See if we can make sense of it, and then publish this along side existing a11y guideance
Thoughts?
SH: The *AUR documents look at best practice in terms of a11y
Doesn't make a call if people should do this or that but states the use cases and things you can do
As well as being clear about what needs to be addressed
Your doc is also great
Getting clear guideance on what best practice is is great - but may not need to be proscriptive
JS: We are jumping a little on this.
JS: This will give us tools for other a11y convos - which I'm looking forward to
JS: For responsive visual games - the speed of sound is not the same as light, so there are challenges but for people with low vision we should be able to push this
JP: I question if you cannot transmit as much information via audio
This convo may bring this out
Challenges can be abstracted to support multiple modalities
MA: This is interesting
Focusing on a gap analysis etc would be a good start - where would they be of use?
What frictions are necessary?
I've seen older games doing interesting in universe AT
If everyone an use it its usability, if only some can use it, its accessibility
<Give example>
Claude dev has a questy aspect to it - fun
Core game mechanics may have wiggle room to present a fun challenge
Story telling has traditional a11y considerations
In my experience, there is a trade off between immersive things and speed
There can be trade offs in games operating on a timescale etc
MA: Could we mention Authoring tool a11y here? In the area of games?
I dont think there is a successfull accessible 3D authoring environment
<Josh still plays Quake *grin>
MA: High level descriptions could be written
MA: < Mentions AI generative environment and gaming tool that isn't accessible>
JW: We will miss Matthew, Janina and Josh next week
SH: Lets postpone and meet in two weeks
JP: I wont be there
JS: Lets leave gaming to the 4th