Meeting minutes
<wendyreid> date: 2026-03-27
Updated charter draft - https://raw.githack.com/w3c-cg/atag/charter-draft/index.html
Brief intro discussion on using zoom's transcript (not AI notes) in place of manual notes. Not possible now, Wendy to look into settings for next time.
<wendyreid> w3c-cg/
wendyreid: : changes mostly minor, in response to feedback. More care around usage of "AI" term, suggestions around liason section.
wendyreid: Please review. If no comments will probably push to approve and conclude the chartering discussion.
Updates to ATAG 2.0
<wendyreid> https://
wendyreid: Not a complete re-write (that's ATAG 3.0). What minor changes/updates could we make?
Miriam7: Define AI, update definition of authoring tools
wendyreid: Can't change normative, can change informative.
nedzimmerman: Lisa and Ned had discussed need to update tech references for 10 yrs of change. Flash, mobile operating systems (Windows phone, Blackberry)
wendyreid: Update references to WCAG (currently stuck at 2.0?)
nedzimmerman: Glossary - authoring tools section to reflect AI and LLMs? (is that too normative?)
wendyreid: Thinks glossary is informative so ok there. W3C process has 4 levels/tiers. 1: editorial. Spelling etc. 2: clarification. Not changing meaning. 3: change that impacts how system would use/conform to spec. 4: new feature, radically changing feature. We need to avoid top 2 tiers.
liskovoi: clarifying lines between author and content. Introduction section could be good for this. Start tackling complexities around when author doesn't have total control over generated content.
wendyreid: Guidelines sections mostly normative, Principles sections mostly normative. "MUST", "MAY", etc. definitely normative. Surrounding/supportive text that doesn't impact conformance can be informative even in those sections.
nedzimmerman: Some 3rd lvl headings (e.g. 1.1.1) has number in page title, some don't. Should strip out extra numbers
wendyreid: has a css trick for this
shiva: Clarified: Anything that's not normative is informative. Non-normative is the same as informative.
liskovoi: had to drop
nedzimmerman: Glossary definition of user agent - need to figure out how to capture using an LLM to interact with web content in a way that's sort of like a user agent. Also LLM-based browsers.
wendyreid: W3C's web and AI interest group might have a better definition to offer here
Miriam7: Talked before about iterative workflow where hooks could be picked up by machines or humans. Could we visualize that flow to help understand how to apply ATAG in that flow?
wendyreid: That's probably more an ATAG 3.0 version, but helpful for us to formulate
Evelyn: The thing to clarify with me for GenAI, that you the author are still responsible for what is output? Is that clear?
nedzimmerman: Room in section 2 part 3 to clarify that
Shiva: Ultimately human is responsible for output?
wendyreid: Yes and/but the tool needs to provide capabilities/functionality to support generating accessible content
nedzimmerman: answer from 2(B).1.1.2.
<nedzimmerman> https://
Morgan Murrah: Interested in guardrails. Generation tools not very refined. What guardrails to put on auto-generation?
wendyreid: Tools themselves must be capable of producing accessible content. Into 3.0 can be more prescriptive about what tools must do. At CSUN lots of talk about instructions/context files.
wendyreid: Wanted to ask: Why do I have to give this? Why isn't it at the agent level?
(lots of positive emoji reacts on ^)
wendyreid: There's been talk on benchmarking for AI. Do we need to consider creating a benchmark for accessibility? (Don't have it right now)
Barry (in chat): "In he near future the GenAI tools will likely be much better than most humans in determining what it accessible and also to enable A11y. We should address that situation."
wendyreid: Adding on: what is much more accessible to the end user? A lot of people looking for future where experience is tailored to their specific needs/desires.
Morgan Murrah: Visualization tool to help understand LLM probability machine: https://
Like requiring debugging information. Requiring information on why decisions were made.
Miriam7: Would be nice to enforce through ATAG, but too prescriptive to enforce showing debug info. Maybe a best practice in WCAG 3?
AOB?
wendyreid: A lot of agents show debugging text by default now, users wanted to see its process
Miriam7: Don't know when ATAG 3 will be ready, don't know where LLMs will go. Could have newer technologies even short term. Should update 2. specs enough to be broad enough to cover future technology.
wendyreid: Should be as future-looking and broad as possible.
Shiva: Admin Q: Do we have editing rights for this draft document?
wendyreid: Yes. In the repo (link Wendy shared) we have control over draft. Can open issue in GitHub or fork it and open PR with suggestions.
Meeting ended