Meeting minutes
Accessibility of Machine Learning and Generative AI.
jasonjgw: understand where the work is at the moment and coordinated at the moment. Janina has been doing work...can you share the update?
janina: i've been working on it. not ready to show yet. Trying things differently. We have a lot of what we need here - ereally great work here. I still think the direction we talked about recently, calling out perspectives...needs to go in there. But haven't figured out how to balance that beyond TOC and use the material we have. But i do have a
question. We talk about ML and gen AI. We dont' talk about agentic. Is any value of showing off in the abstract - there are categories of it. Question 2 - definitions. Is the gen Ai definition correct? Contradict ourselves? And question 3 - are we delineating these enough or back off a bit? Wondering if trying too hard to delineate what we're
trying to take advantage of?
Scott: we do touch agentic reference search tool, it's briefly but not defined.
Janina: if wer'e alling out ML and generative in abstract we might as well call out agentic - be consistent
jasonjgw: first point - I think all of the AI that people are concerned with for the last 14 years or so based on a partciular kind of machine learning that employs artificial machine networks. But that's not the whole of AI. There are rule based and other systems, historically important, but in this paper we're not talking about those. Important
to clarify ML and particulalrly neural network AI. Within that, you can say that there are some applications that produce applications of some kind...how important or how strong a clarification that is, I'm not sure... You might say there is a distinction to be drawn but not sure how important it is. Likewise with agents, not sure where the
boundaries are drawn, able to exectue some functions that have side effects of some kind other than what it generates. I'd say all of the distinctions are relevant but not sure of the importantce level. But ML and neural network is an important clarification. Otherwise they might think we're interested and talking about other kinds.
Janina: i think ti deoews help in the snese it explains why we havce what we have in there. COnfusion of puposes in the paper - impacts on a11y and users who are end users and also creators. Yet another perspective that might want to add for user - third party? That's a separate question. I think that's our primary focus. Academic deep dive on
dileneations are interesting but not the main focus. Propose...knowing you have a deep background - let's back it out and just talk about AI unless or until one of th euse cases says it's clearly "this kind" and distinctive from other kinds. Somehwere in intro we say we'are aware of architecture and deep history, but in recent years practical level
it's everywhere, where previously it was academic. And there's where we lay out which is which. That would make sense to me, but having throughout the document, I dont' see it...Are we loking at the wrong end?
jasonjgw: here's where I think the differnece is. I think distinction between neural network AI and other kinds important. With neural network - worlking on data and probabilities. Take input and layer through network based on probable interactions and come out with some result. Previously AI systems were rule based, so had deterministic outcome.
Like text to speech (Janina says i think that's where we lose people). If you have text to speech rules...Issues today on inaccuracies and inconsistencies due to systems are machine learning-based, and we don't fully understand how they work, and they don't give the same output with the same input, and it's hard to predict the output. Unique to the
systems. And these characteristics are important for effects on people with disabilities and tech challenges these technologies create. A distinction important applies to the network (what we say), but wouldn't apply to rule-based AI systems working on different technology.
Jpaton: I do agree in that expert systems we are not worried about those. Statically hard coded. Looking at AI in terms of that then you also have ones that look at the internet and find on internet, just not on their own model. Model itself trying to pull in data from elsewhere. I'm more on Janina's side in that it makes sense to keave scope open
to say AI at all point and reduce scope when you need to say "this is agentic AI" or other AI. Defining what your' talking about. Reduces scope and complexity, doesn't add to precision or understanding.
Scott: when I look at, for example, something to be said when we think of ML and early days and programs like eliza (rules based and keywords and feels like an interaction) look at where things are today. Catalyst people with disabilities (feedback i get) what's different now is PWD say OK, AI is big, People see it as being helpful. testiment to
rightly or wrongly, having war, petrol prices, US stock market goes up...reason for that is belief in society that AI is going to be the answer to all of our problems. I think then, we comt back to the question how does AI help me? How does AI hurt me? And how can i as a PWD know what's going to be available to me...and where the advancement takes
place? My view is talk about the legacy of AI and ML, good to dfined and good to have definitions in how talking about AI. Agentic, generative, etc. Work that will evolve. Critically to John's point the doc needs to be open ended. 2 years ago when we started agentic AI to book a flight wasn't readily available. Perspctives- how AI can hlep people
with disabilites, but fundamental flaws...paradox how do we know the information is helpfu? Jason's point - acknowledge the legacy, not new, but the applciation of AI and creating it's own rules and path - where si w3c guidance on how all of this applies to people with disabilities - the good, bad....I think this could be that. We have a lot fo
good information. Still keep it open ended.
Jasonjgw: Janina, before a facilitator's question, do you want to comment?
Janina: no. Ask
jasonjgw: we have a couple of options. one - confine scope of what we're addressing to ML neural network form of AI, current contemporatry interest and discuss in various implications. Could do that and make changes but still continue. Another approach - talk about AI in general and implication in that case, when arrive at issue that apply to some
kinds of AI, we would need to clarify to what it applies to and what it doesn't apply to. Clarify in the discussion. Could complicate the discussion? Which way we want to go if those are the options?
Janina: yes, this is a restatement of where we started.
Jasonjgw: i think concerned about neural and ML...
Janina: no i think it's how it applies to me. not under the hood. I think end user, content a11y guidelines, they aren't going to care about underlying architecture, but what it does on bwhalg=f of users and gets wrong or gets right. But want to know when talking about neural net based how is that different and what impact on disability does it
have so we can write specs on that. Expert systems is a useful distinction. Maybe rule based statistcal help us build better expert systems...already intoa rules approach could arguably call it a hybrid. It's going to be done in a branch. How about i try it and then get asked to pull the specifics back in so we don't miss it? Pull some of the
specificity out but if it needs to be somewhere, but then based on why this is different for this model? I guess that's the piece I'm missing the impact of different architectural apprashes on people with disabilities...
jasonjgw: all of the contemporary interest is around neural network models..intrinsic characteristics that make them liable to error. Purely rule based appraoch to AI doesn't have those approach to characteristics. what i'm saying, rule things out of scope and say ML neural network applciations we are writing about as a matter of scope definition.
everyone would understand that we're not talking about other kinds of systems.
janina: almost willing to agree with that. what about Innosearch AI?
Janina: so you would rule that in?
Scott: yeah I'd rule that in. provide a conversation.
jasonjgw: conversation all done by a large language model.
Stacey: tacey recommends focusing on the audience, and ensuring that it is clear to all of the readers what the definitions mean and waht the implications for persons with disabilities are.
Stacey: it needs to be clear how the technology affects them.
Stacey: recommends clarifying the goal of what is needed in making use of the information in the paper.
Stacey: recommends ensuring people understand what we are talking about without developing them in great depth, and clarifying what the implications are for the users.
Stacey: important to know the background and definitions...but as end users we need the "the so what" and how it impacts me
Janina: I like that "and now what' and "so what' test useful for audiences. Everybody audience isnt' it but shoudl be able to read this and say this is what i'm experienceing. As Scott - is w3c caring about us? Are they paying attention? Which takes this to architectural crowd - inform what might take us there and interface with them who is happily
including us in w3c. So yes - we want them to see/understand enough of the architecture and history...not making stuff up.
Scott: could separate into SO WHAT (history) for people with disabilties and NOW WHAT for people with disabilities (implications going forward)
Stacey: I volunteer to be tribute (collaborate with Janina)
Jasonjgw: message seems to be draw the distinctions but don't belabor them
Janina: yes, and you can all agree to say yes or now
Miscellaneous topics.
jasonjgw: discussion so far based on idea that John had a paper in development with game a11y. Should APA working group say anything. Do you have a paper that is ready to be shared with this group?
jpaton: yes, paper on more on legilative - what exists at the moment and potential for legislation and what can be done outside of it. Big thing needs to happen - standards for games accessibiltiy. If that could sit with w3c, a huge step to where we think we need to be
Janina: explore where we can write a11y user requirements on gaming
… those docs have been very influential. RTC clients more accessible or not, due to our good work
jasonjgw: John, put to mailing list when you have something to review? (the paper you're publishing)
Janina: and when you might want to make that a topic of the day to consider, proposing to APA which would be the next step.
jasonjgw: standards and guidelines published, so question from w3c might be where does an APA or RQTF fit into this context. Noting that's something to think about
Janina: bet John can guide us through that
Jpaton: guidelines aren't built around the needs of the user. I think games requirements would help with that. Like menu systems needing to be accessible...different formulation. It think it would bebefit.
Jasonjgw: can we put the topic on for next week? Worth the time
… anything else?