Meeting minutes
Accessibility of Machine Learning and Generative AI.
jasonjgw: Anything we should say about document revisions? Anything usefully say or coordinate right now?
Janina: there isn't. Need more time to deliver.
AI Agents and Open WEb Ecosystem.
jasonjgw: janina, can you tell us what this is about?
Janina: will give context before turning over to Roy. All kinds of activity in W3C related to AI - what should say or leave to others. Our activity is welcome and important and looked forward to. One of the things - and Roy is deeply involved in breadth of activity in AI and the organization of the workshop - it's an instance of wider activities.
<Roy_Ruoxi> https://
<Roy_Ruoxi> w3c/
<Roy_Ruoxi> w3c/
roy: China member, some discusion on AI agent and web ecosystem.Related with our with in RQTF becauase of research of AI and a11y. Lots of discussion on AI agent and also AI Agent browser use, from human use and paradigm from browser to user use to computer use with AI empowered agent. I think there are something related with a11y, one is about the
generative AI. Was a topic from Hwawei, talked about the use of AI and interactiviy with human use. EX: AI automatically generate UI with some personal features, if you like the color, what kind of degrees of contrast, and also some UI without advertisement, you can customize the UI and gather info from other web pages. Also some MCP tool use and
use of browser like web mCP, machine learning community group (W3C). Maybe we can take something into consideration with regard to our document with AI and a11y things. Besides this, I think W3C has prepared something, and I think I shared this already AI agent workshop - it's not prepared yet. But if there is anything you want to bring into this
in the workshop, welcome to leave your comments. And interest group related with AI is planned to (already reviewed charter) and RQTF liaison for this working group. Talk about AI related things. Include a11y and user experience things. That's ongoing things or if you want to work with this interest group welcome to comment and join.
roy: that's a brief introduction. And we may set up some connections.
jasonjgw: first round of comments/questions from others?
jasonjgw: make sure we have something fiarly substnatial to say about agents in the draft we're curently developing. I know we started discussing recently, but to be ahead of what's in the public realm, need to say something about the a11y implications of those which can be quite significant
Janina: Need to go through the materials and do some reading and figure out our role. Roy- is there something specific RQTF shoudl think about?
Roy: ai agent could be a good topic for a11y. Exploration from big company like Microsoft and others. Recall some news they tried to have ai agent to do a summary of web page and have people to fill in some table forms. that's something I read, need to find. Believe AI agents- people ahve started to look into these and a11y. Another end goal,
because not all websites...some ai agents try to use a11y feature to get to the content of the website, so that's also in that way, the ai agent can get rid of restrictions...a lot of topic of potential things to take into consideration in our document.
jason: listening to a podcast recently, and they discussed protocols to enable these agents to access various web applications. Example was use it to take action on issues in github. To do that, you need to use one of the protocols thats emerging to allo agent to access issue tracking capabilities in a repository. One of many examples... If you
have protocols set up, agent can be asked to perfrorm actions using relevant web API and agent can do so to the extend that it's successful (don't know the error rate). Could provide a11y benefits, agent has advantages in a11ya nd provides automation benefits, indirectly beneficial to people with disabilities. Reduces amount of interaction with web
interface because it's doing it for them.Connect agents with LLMs with carious web applications could have potential. User needs to keep track of what they're doing.
raja: want to jump in quick on LLMs. Example, trying to creat summaries and comments, people with read them fast/quickly and apparently people don't like...get the sense that they're unahppy with current LLM summaries. They're stillmissing out. Not a perfect fit. Still a challenge. Especially for people who have accessibility needs.
jason: yes, putting having to evaluate the output of the agent creates cognitive demands in addition to performing the task so some cognitive capabilites that will be used potentially to assess because of these demands.
raja: yes, and what is the solution for this?
jason: that's a worry, and here's the issue. Suppose that if you can't get teh agent to do what you need, or a different agent to do what you need, need to look at the underlying application. Still has to be accessible in any case.
raja: one question, suppose an indiviudual is using summaries again and again, to the point...they're acessible to some extent, but they could adapt to meet a person's need...developing a relationship over time. Is that the ultimate goal? Agent becomes familiar with the user's needs?
jason: can you create a prompt to better meet the user needs? A line in that development
raja: yes. Give specfici language needs, context how you cognitively use lanaguage....changes time even with a single user. Context, environment...Identify the speaker, doesn't matter who but needs to identify it.
<Roy_Ruoxi> https://
jason: on agent develoeprs to provide some flexibility in handling requests for specfic kind of information and interaction preferences. So when someone interacts with an agent it's more likely to meet their a11y related needs, because they can specify or unable to do it themselves, have assistance in specificying
roy:
roy: google trying to do a protocol A to A and there's a communtiy group. Do some verification and authorization of thigns. When try to ask an agent to do a delegation, control how many layers or how deep you want the agent to delegate for you. That's security and prvicay and athorization things. Protocols to make sure all of the different agents
coordinate with each other in a safe way. Recently OAM chat GPT or Gemini, some general for people to use. But if you want to use a specific one, there doing some training in specific industries. So in know your customer, some company will train it with RAG or something else with a list of confidential data to make it much more clever than some
general...Different agents might have different capability. Assign a task to a specific agent, or subtask to some agent which they're good at and give more clever feedback. That's something recently from the industry a lot of companies are trying to work on that.
jason: Google latest (article) LLM work - more facilities for storing info from previous interactions from the user and providing to the user. There are privacy implications, esp if there are disability related info in there. Privacy and functionality trade off and accessibility and what the agent has access to.
Raja: when talking about privacya and security - apple isn't allowing/preventing people from accessing their automated prompts. Like if you try to screen shot or use that...unable to. So many security things in place with automated messaging. (in automated captions). Not allowing folks to use automatic captions for future processing. Can't use it,
they locked it. Limits on the output. So for user output Apple put limits on user output which defeats purpose to try to modify meet user needs, adapt to accommodate.
jason: need to develop an extensive section for agents for this draft that we're working on
raja: so far only Apple is doing the limits on captions. Like screen readers, would that be another (asking) that's taken into consideration?
janina: very impiortant conversation but not new. Privacy tradeoff conversations well before AI. Ai has re introduced this concern with a11y. People don't want to disclose disability and they don't want them to know that. Now you can redline me. ANd go across all types of disabilities and the colelctors of data find out about you. What should we do
by way of law? But real tech question is, how much of really good service are you willing to give up by colelcting that data about you. If the latest AI can tailor output to better respond you might get really superiior service from your computer that might not be as good. If it's not as good, means it will take more time. Been there forever. May
be solutions, and may be one that hasn'tbeen discssed much. Intermediate proxy between you and the web. Tokenize and doesn't disclose the WHO, just disclose THAT which particular characteristics. Might be worth something - $4.99 per month to have an intelloigent agent owrking on my behalf?
raja: disability tax needs to be paid there, right?
janina: if can disclose, maybe we can minimize that tax
jason: worried it may not work but it's interesting. Correlation with other information that they could tie it down to a specific person from other web interstions or mobile interactions that aren't proxy protected. So - it has it's limits. Might solve in some but not others. challenge with agents and LLMs is that the only way to have them
customize repsonses to indiviudal needs, specify needs to the system and have it respond. Differtent than every other web based system I can think of. Alawyas had systems that you can set preferences, but they've been generic. In this case you can specify precicely and it can disclose a lot more
raja: yes, it interfers with privacy and that's a challenge
Jason: that's going to be one of our thigns, then. Appears to say somethign about agent interactions - advantages and potential challenges, storing context, noting privacay straight off. Something of a starting point for making intelligble and thoguhtful comments on this. Anything else we've missed that we should be picking up in our work.
janina: any thoughts (to Roy) for next steps. All of this will help inform the paper. or spcifically do
roy: i talked with people from AG working group. will there be any influence to some WCAG 2.3 or 3.0 in the future. Should look at the AI influence on the standards. And also in regard to user requirements docs and tech AI or AI agent team to take into consideration for use cases. Maybe look into how AI agent developed in the future at this stage.
Can we find out something, maybe a paragraph in the doc and gather thoughts on this part and try to draft something into the document. Start to gather thoughts in one part and exploration to reflect in our existing doc to the future.
Janina: is it really still a possiblity that there's a WCAG 2.3? Is that being talked about?
Roy: I'm not sure.
jason: concrete - Janina is working on Scott's changes and revisions. Additional work on agents - taht would be the next round of changes
janina: maybe the term "agentic AI" may need to be in first working draft. Oh, in terms...yes, Scott changes. then anything in way of agents. Community group on developing voice agents (things you talk to). General service turning web resources into smart agents.
<Roy_Ruoxi> https://
Roy: there's a report from Dominic to talk about agentic web - risk and opportunities. And welcomes feedback if you have thought.
jason: anyhting else to say about this before final agenda item?
Miscellaneous topics.
jason: attending the 2025 Ireland science a11y conference. Will report back on anything relevant.
Jason: anything else going on that we should know about?
jason: more predictions about financial crash as revenue behind on corporations from AI. Pattern that these things follow - overestimate impact in short term and underestimate for long term. Article saying a lot of infrastructure not being used and will be available at a lower rate. Might open up the window if large corporate demand drops.
janina: one aspect I don't hear about is quantum computing - is it really so far in future that we can ignore or is it sooner?
jason: if some known problem that's not encryption...something we should be concerned with and if it can perform better, then suddenly it's an issue
Janina: before can deal with it ask "is there a...?"
jason: so far cryptography the only one I've heard about. Narrow where advantages lie. That's not definitive.
Janina: maybe this should be something to think about...