W3C

– DRAFT –
RQTF meeting

19 February 2025

Attendees

Present
janina, jasonjgw, scott_h, stacey
Regrets
-
Chair
jasonjgw
Scribe
jasonjgw, Joshue108, janina

Meeting minutes

Accessibility of machine learning and generative AI.

<Joshue108> JW: We need to discuss this

<Joshue108> JW: We need a structure on the issues and happy to hear update

https://w3c.github.io/ai-accessibility/

JW: I've suggested structure on list

JW: We are not likely to identify all or even most relevant use cases

We can treat the cases we know about as examples and raise issues

This would help with a conceptual framework

<Joshue108> +1 to Jason

JW: This would help the document age well - without the need for constant updates

Leads to a more enduring approach.

<Joshue108> +1 from Janina and Scott

JS: We can point to things in various domains and aspects

SH: Good point - its too vast.

We can pick on examples and case studies

JS: Can we brainstorm?

Re: Table of contents

JS: We have the person with a disability on the web trying to take advantage of AI enabled contents

We need considerations related to that.

We have content creation being assisted by AI

we need to be aware of that - good for text

Problematic for images - what degree of confidence can we have?

Code generation is also to be included

Then the user interacting with various agents - smart travel.

What will make this helpful?

JS: How can AI bring accomodations?

Then there is the user interaction with their own systems.

I wouldn't mind if it was monitoring what I was doing and helping me be smarter on things that I could do.

Or notify me from time to time?

Can vary from user agent. Rich vein is thinking of this.

Assistive Technologies should be looking at this.

SH: Saw some smart home/car videos are interesting, immersive AI

SH: Very granular - and the a11y conversation is not there yet

JS: The standard agent for folks without a11y needs will not work

It needs to be considered or it will be an inaccessible world all over again.

https://w3c.github.io/ai-accessibility/

JS: We are going to build the TOC shortly and get an a FPWD out shortly

JS: Would like to see this published soon

JS: Thanks Janina - these suggestions fall well into my analysis

We have ML learning in the authoring env, as a provider of app functionality, and then its use to enhance a11y of UI

JOC: Above comments from Jason White btw

JW: Hope this agreeable

JW: Suggests working on these contents

JS: I want to move on with that job - I dont disagree with the categories. Not sure its the top level of the taxonomy

The user needs to be the centre of everything

JS: All the usual cliches need to be removed

JS: The hyperbole is exaggerated

<scott_h> LG AI video https://www.youtube.com/watch?v=gYRM00Oe2BM

JW: Cory doctorow has good article

JW: There will be good writing on AI A11y - Jason is working also on interesting stuff

<scott_h> LG video shows AI integraiton across multiple devices and user agents but accessibility not considered

RK: There is a lot of potential with new ways of using AI - advantage and risk needs to be considered

The ack jan

RK: The errors need to be considered, disability tax e.g.

Disability tax is the increase of work and effort because of a disability in your everyday life

RK: This needs to be authored in partnership

As a deaf person, voice detection and captions can be added.

But there is more work, as it is not in Sign - so there is another layer of translation

Additional work needs to be thought of

<Joshue108> +1 to Raja

SSG: There is also a 'real' tax to have equity - the tools that are needed, medical care etc

SH: We have noticed this in Oz - things that are 'special' or disability supported are more expenses

Joshue108: +1 to Raja, agree on tyhese scenarios

Joshue108: agree also with user first

Joshue108: This is our opportunity to flag this

Joshue108: the overlays are built by nurds who don't understand the userneed

Joshue108: this is important to flag in this doc

JW: Jason, I'm hearing a concensus as from the user perspective as the top level taxonomy

JW: I think that is reasonable to a large extent but there are issues around the authoring content

Error handling for example

<Zakim> Joshue, you wanted to say why cant we do both

Joshue108: Why can't we do both?

Joshue108: Agree with jason's points, very valid

Joshue108: AI now a panacea, the devil is in the details; we're still in the sales stage

Joshue108: We should ask these larger questions

JS: My comments are pragmatic - I want to know the structure

Am concerned about scope/size

JOC: I dont see the suggestions that Jason made are antithetical to putting the user first

JW: I dont think user needs capture it well for this document

We have other areas to cover

JS: What is an example?

JW: Ok, regarding content creation - you are satisfying user needs for abstract users at that point

Under application development that could be a heading

JS: Counter argument - that sounds like a philosophical thing - sounds like user needs to me

SH: This work started on talking about their perspective in AI - there are other things - remedial actions, how to engage on AI to support People with disabilities and to Josh point about how can we trust it and its ability to support users with disabiities

SH: Those elements need to come in

JW: I have to agree with Janina on this

e.g. the authoring process does support people with disabilities -

JW: How?

JS: Difference between quality management

<Joshue108> +1 to Jason - quality is critical here

JW: Yes, this will be error prone tech for a long time

JS: Just to be clear - in the intro we can say this tech will be error prone, some of this will produce content that wont work

I don't want to discuss that

JW: I think this is critical

RK: The question about authoring ..

Creating videos and content - what are we using the term authorship?

JS: I think Jason is thinking very broadly, doctoral work, cool app building etc

JW: Yes

<Beethoven is being invoked>

SH: Who we writing for?

JOC: Great question

SH: Devs and designers that engage with the user?

jasonjgw: Anyone trying to integrate AI into their applications; code,, AT, content, etc

jasonjgw: Highly relevant audiences for this work

jasonjgw: We want to influence those people in W3C and elsewhere

scott_h: To support a user need? or other applications?

scott_h: So errors, about ML to pick up errors in ml itself? Is it its own end?

jasonjgw: Well, that's useful, but in the end that needs to serve greater a11y

Raja_Kushalnagar: Currently no way for AI to catch its own errors

Raja_Kushalnagar: maybe someday, but not yet

jasonjgw: We have error handling as a major problem

scott_h: Yes, ai halucinations ...

Raja_Kushalnagar: Showing to the user and the audience

Raja_Kushalnagar: Sometimes we'll see things -- in captions

Raja_Kushalnagar: Both people involved need to see them

scott_h: If AI generates someone with 6 fingers, a low vision person might not catch

We also have the potential for AI to produce trashy code that could be sold as being accessible remediation

jasonjgw: when ai is generating part or all of the ui ... not scripted -- all ml generated

jasonjgw: what capability makes the user's a11y needs supported

jasonjgw: errors an issue there too, but can it meet my cognitive needs, my at needs

Joshue108: also a11y remediation ...

jasonjgw: Two things we've identified:

jasonjgw: 1 Issues of accuracy in general -- accuracy of what the ML does

jasonjgw: 2. what are the capabilities of the ml when it's producing content and/or ui -- does it comprhend my a11y need and accomodate

Minutes manually created (not a transcript), formatted by scribe.perl version 242 (Fri Dec 20 18:32:17 2024 UTC).

Diagnostics

Succeeded: s/everyting/everything

Succeeded: s/taxonomyu/taxonomy

Succeeded: s/yuet/yet/

Maybe present: JOC, Joshue108, JS, JW, Raja_Kushalnagar, Re, RK, SH, SSG

All speakers: jasonjgw, JOC, Joshue108, JS, JW, Raja_Kushalnagar, Re, RK, scott_h, SH, SSG

Active on IRC: janina, jasonjgw, Joshue108, scott_h, stacey