Meeting minutes
Relevant themes from TPAC 2025.
Janina: most important thing to know is that our approach and what we're calling discoverable destinations seemt o resonate with people. We had a breakout, participant from Google, very interested and liked the direction. We're not doing anything "new" with markup, but taking markup opps that have been underused and putting/suggesting
data...socialization approach.Key destinations can be machine discovered and presented as simple list. Not predictable. Not where we landed, but made it possible to programmatically discover and link to/discover login section or a11y statement and some can be machine readable and construct interfaces around it. Didn't get a lot of time in meetings,
but good progress because of good particpationint aht breakout. and APA task force is moving forward and symbol suport and bringing symbosl into the web. successful conversation. And most directly of interest to us, AI is a big thing for everyone and factored in TPAC.
… Need to send the pointer to the list, there was a meeting discussing ARIA mpping for protocol for AI agents that we will need to take note of.
Jason: last point is going to be relevant. Will have to see how that affects our work on the document. First question - anyone who wasn't at TPAC, any questions in response to the summary?
… We've taken note of the things we need to take notice of. Any final thoughts before moving on?
Accessibility of Machine Learning and Generative AI.
Jason: ongoing work on this. As soon as Janina is back to full strength we'll continue
Janina: no competing priorities, so this is number one
Jason: going back to..relevant to this document, I've been reading some articles published recently. Mostly critical regarding use of agents to perform tasks on behalf of the user. Interesting security implications regarding prompt injections. Idea is that suppose you ask an agent to go onto the web and perform and action in email. Could be some
text maliciouslycrafted to interfere with the agent. This should be frightening.
Scott: term is agentic AI?
Janina: Yes. There is a protocol called MCP that's been brought to W3C for possible standardization that facilitates the creation of agents to do tasks.
Scott: not all tasks in teh agents are doing the greater good
Janina: susceptible to prompt injections
Jason: some agent (artticle) had gone and destropyed an entire file system because it had permission to access the data
Raja: They typically have safety nets, but you don't want AI to do that.AI detriments
Jason: users have to be careful of how they use and what access they give it
janina: need to be careful of what to expect. If we say it's early days...we don't ask a baby to go do adult things. Get unpredictable results. A reasonable comparison there. Likely going to be a thing, but kinds of issues that need to be worked out. How to fence an agent to keep from misbehaving. Need to give it limits and certainly an issue
around that to precent prompt injections. In W3C, there is a whole area of people who look at security and privacy concerns. That will be a factor.
Jason: using in the right context might be cognitively taxing. Could introducse demands due to unreliability and susceptibility to security issues and challenging to use who are not skilled in it
Janina: aspect of creation of an agent to do tasks (similar to software development task), yes. But once an agent is owrking reasonably well, it would work as expected and more like a tailor made suit and not be particularly hard to use. It's a finished product that does a task.
Jason: that's the hope. People betting large amounts of money that this is where it's going
Janina: distinguishing between an agent creation and the use. Distinction becomes the real difference. Could be helpful in accessibility. Any number of us might find agents really helpful. But the creation has the limitations you outlined at this point
Jason: use of them has the limitations as well, as the prompts they accept from user are failing open ended. Not confined to a script, as it were. Result is that even thought hey might have access to certain info and applications and might be confined, the input from user could be broad. And unpredictable when getting to the user. Challenge there
at the moment. Unfortunately can't say problems are all or most in creation step and once they get to the user there's a stronger assurance in reliablility. People would like that to be true, but can't assume that at this point.
Janina: can't. Reliability is an open question. But ease of use might not be
jason: ease of use is significant, and one capability they already have worth noting, to some extent can generate user interfaces. I was using one with a series of buttons with all titles i used previously. They must have been generated by the model as I didn't create them. Sometimes it asks if you want to follow up and sometimes it generates
itself. It refines the user interface and engaging the natural language dialogue.
Janina: reliability is going to be a concern. We struggle with it and will work on it because there's a benefit and ease of use and comprehensibility of work effort
jason: how easy will it be and how the UI will work? Vary by agents...Monitor what it's doing and intervene when it's not doing what's intended? Exactly how all of that works isn't clear. Might want to say a bit about that.
… because the problem is the user has to chase after it to find out what it's doing. if it's not, they have to correct that. This gets interesting. More topics to think about as the document progresses
Janina: need to talk about agentic AI
Jason: and the issues at hand
Scott: concur. huge benefits with a11y .But need to know the risks and reliability
Miscellaneous topics.
Jason: anything we haven't covered we should know about?Scott: internation people with disabilities day - my day is just finishing. Lots of presentations and lots of stuff out there. Announcements with Google and Microsoft. Lots added. Lots of stuff worth checking out
jason: thank you. I'll go look and start reviewing
Janina: anything of particular interest would be welcome to list
Scott: Google added some a11y things...everything seems to be plus AI. Auto generated alt text. A whole lot of announcements as a result
Jason: involved in some AI related activity looking at the role of ML and decision making that's in collaboration with university with Uda's group. Drawn into tactile displays resulting from a presentation I gave earlier this year. Won't know more until that happens. These areas are delivering some standards and activity at the moment.
… apart from other work I'm doing not related to a11y
janina: glad to hear you're involved wtih OCAD and UDA. They wanted us to be in touch with UDA before going to a working draft. Glad we have that connection.
Jason: daisy consortium user doc got published. final version did come out
… anyone who wants to make use of that should go in and read it
janina: we did have a meeting with daisy and epub as part of TPAC APA. Minutes I can forward. several items of interest that popped up with that
Jason: let interpreters know that it will be on an as-needed basis moving forward. Will formalize in an email.