Meeting minutes
Accessibility of Machine Learning and Generative AI draft.
jasonjgw: Much of this is on Janina's schedule.
Janina: I have one other item that has a firm deadline of April 16.
jasonjgw: Could you think about it later and we'll try and work out...doesn't make sense to have a detailed conversation until you have something in a branch. If you have a sense of when that will be. Should we postpone (meeting schedule) and come back when it's ready?
Janina: I'd like to stay for next week. Two weeks following, I'm unsure about. I have been putting time in, don't want to drop it and reqork.
jasonjwg: next week, we'll take up whatever you have, even if it's just on the call. And we can discuss the schedule for the following two weeks after that
Scott: I like perspectives (name/titling)
jasonjgw: Would like to look at in context of the table of contents.
… I'm focusing on canadian standards work while Janina is focusing on this. There's an email on COGA interested in this
subtopic coga input
Janina: COGA realized thy were talking about AI. They asked if they should be talking to anyone else? Email I shared, points to two papers.
… Apparently they say somethign about AI in those papers. MAinly GOA wants to know they're on our radar. If anything we can pick up from them. No C tower situation again.
jasonjgw: concerned about the limits of what we can say. On one side, it can be various forms (like LLM) good at making content comprehensible given needs but other hand very good at providing misleading information. IF someone may find it hard to understand original content could make it hard to understand (or confirm) the given info. Gain
some/lose some.
Janina: one of their papers has to do with online safety. Whatever they mean by safety.
jasonjgw: that's an ambiguous term. Many cases, acknowledge trade offs and limitations of the tech and make people aware that this is what we can say
Janina: agree. might be something we can pick up and something we can't
… they might have good points, but it may be for a different paper, not this one
jasonjgw: revistit when we see what Janina writing
Janina: we'll ahve to see. But at the moment seeing it as a separate topic
jasonjgw: anything else on this topic?
scott: tech gremlins for bibliography get updated?
jasonjgw: it got updated. No one went in and changed source files
… I did send a summary after the last meeting on what was going on
Scott: glad it sorted itself out.
jasonjgw: Janina, if you run into the same type of trouble, let us know
Janina: I did bring it up to Roy. He did update Renote...so maybe it was that?
Janina: I will have talked with Roy/in close touch with him. He did take the red note reference out in the last 50 hours.
… my affiliation points to my w3c ID
jasonjgw: there's a specific field that needs to be added beyond that
Janina: thank you. I"ll follow up on that field.
… you have my permission to add it, but I don't know what the field is. I can look at it later
jasonjgw: i can take a look at that later
<janina> Article on end point security concerns with agentic AI: https://
JAnina: URI from Matthew. Kicked around in the tag. Wondering how we're going to handle it - security concerns in agentic with agents
jasonjgw: generally aware, what's the specific issue?
Janina: don't have anythign more than what I said. Didn't get any more from Matthew on what w3c is looking at in this context
jasonjgw: yes, agents can be a security problem. They can just link info, they do whatever the probabiility in neural networks causes them to do. Inherent security risks. Work to compartmentalize them (strongly restrict what they have access to and when), imagine those restructions have consequences for accessibility...
Janina: another complicating wrinkle...
jasonjgw: don't give it access to your file system, things you care about. APIs, the root.
… what API keys, websites does it get access to. But what you DO fgive it access to you need to monitor its activities carefully. Interesting, let's suppose that there's a website that's not accessible to a user. they ask agent to perform actions. Limits abiltiy to check what the agent has done...so ability to go and fix what th agent is going to
be limited by the same accessibility issues that were the problems in the first place
… some work off of images of the web pages, not doc object model. Processing the whole thing. That might avoid some a11y issues but not all of them and introduces opportunities for errors. two sides to that story
janina: become a rule of thumb? Will the evidence show that or not? I know we have a concern about a11y framework to be hijacked to inject meta data for benefit of Ai tooling...
jasonjgw: a11y and security issues are liekly to interact
janina: always have
Miscellaneous topics.
jasonjgw: anything going on?
janina: time change for John next week!
jasonjgw: if something comes out of candian standards project i can comment on publicly, i'll let you know
… we have student project going on with XR accessibility, getting involved with that. Inquiry from Janina. Will see more after that discussion.
Scott: been more discussion recently on some California law and global law on age verivication. Do open source OS start to need to have an age verification in the OS. In Australia first country to ban chilredn from social media. So thought what was interesting move beyond that is the discussion of open source OS to have age verification in the OS.
Haven't personally come across before, but popping up in recent weeks. Not sure if relevant to this group. But reading pros and cons..mostly cons...Think ti's something that might pop up on our radar
jasonjgw: age delcaration and age verification.
scott: yes, it's the does it need a way to prove vs declaration "yes i'm over..."
jasonjgw: if it goes to websites from some mechanism, If it goes beyonf declaration to verification will be accessibility issue for sure
Scott: yes, w3c has verifiable credentials for sure and ethical questions...shoudl it even be part of that if no one behing the monitoring it
jasonjgw: Brazil is more of a concern because their regs are enforced. California and maybe Colorado and EU...which country there now?
Janina: legal news in the last week on meta in new mexico on how the platform treats kids/child safety
jasonjgw: already have personal ID verification over the web. No mechanism in place to line up documents for the machine laerning app on the other end to recognize them. Using machine learning based tools to recognize ID and perform the ID. Sometimes they want the user to be in the vie with the document. and all standards to make it accessible but
it doesn't tend to happen
… i think the whole issue of verifying identity is going to be an ongoing accessibility challenge
… CAPTCHA going to come back. distinguish human users from agents. and the agents will be come increasingly effective with the turing tests as the CAPTCHA challenges
Janina: vicious ...or virtuous...cycle depending on the perspective
jasonjgw: not unrelated problems