Meeting Minutes
Wendy Reid: All right. Why don't we get started? Hi, everyone. I hope everyone's had a good couple of weeks.
I wanted to start today off with, I think, just like a couple of business items, for those that…
I don't know. Registration for TPAC has opened up.
Wendy Reid: https://www.w3.org/news-events/tpac/2026/registration/
Wendy Reid: Shared the link there. Just as a reminder, TPAC is the last week of October, so October 26th to 30th. It's in Dublin, Ireland.
we will be meeting during TPAC on the latter two days of the conference, so, Thursday and Friday.
remote and in-person. So if you're not able to come in-person, you are absolutely able to join remotely, and we'll have, call information closer to the date once we get that set up.
I highly encourage everyone to register, because registration not only helps.
It helps us know how many people are going to show up, either in person or remote, and just plan accordingly. It also helps the organizers for TPAC.
There is…
I always like to mention this, so there is a…
anyone, so if you're a W3C member or not, you are allowed to come to TPAC. You are able to register for TPAC, so you do not have to be a W3C member.
there is a fee associated with registration, however, there is a fee waiver program. So, if, for whatever reason, you… the fee does not work for you, but you still want to attend, I recommend using the fee waiver program.
I like to mention that, because people don't… it's on the page, but people sometimes miss it, and I think it's really… I don't want cost to be, a barrier to anyone, so, you know, please make use of that if you need it.
If you are an invited expert with W3C, your fees are waived automatically. So, some of you may not see fees. I didn't even think about it when I registered because I'm an invited expert. I was oh, I didn't even get asked for the fees. They just waive that automatically.
but obviously, if you're part of a member company, there are fees.
So please do register. It'd be great if people are there in person, but totally understand if travel is not possible.
And I'm excited to hopefully get to see some people in person. I will be there in person. I should also mention the Inclusion Fund. There is the Inclusion Fund.
I will share the link for that as well, because that can help people as well. There are criteria.
Wendy Reid: https://www.w3.org/get-involved/tpac-inclusion-fund/
Wendy Reid: But it tries to be as inclusive as possible, hence the name.
And it might be able to help with some costs for travel and stuff.
Rémi Bétin: Just a reminder that, so, you are really welcome to attend the ATAG CG meeting, but you will also be able to attend all the meeting.
doing TPAC as observers. So that's also something you might be interested in. there is also all community meetings, there are some breakouts, so, yeah, feel free to take a look at the schedule, and you may find very interesting.
Rémi Bétin: Events,
beyond the ATAG CG meetings.
Wendy Reid: Yes. Thank you for mentioning that. So typically, I think the schedule is a little bit different this year in that there's going to be some breakouts every single day, but Wednesday is still…
breakouts day. And I would say, if you're planning to come, especially in person, and you're not sure if you should just come for Thursday, Friday, or plan to be there longer, I highly recommend including Wednesday, because Wednesday breakout day is really fun.
the breakout sessions are always very interesting. It's also a good chance to meet people, because you'll be jumping from session to session, and there will be different people in every room.
And it's a good chance to get some exposure to the community in a, I think, a lower stakes environment than a working group meeting.
All right.
OK, so that's TPAC. We've got that.
And then…
Why don't we get into our agenda for today, which is very exciting. We've finally completed, I think, or we've gotten all the way through Section A, or Part A.
of our review of ATAG 2, and we are moving on to Part B.
Wendy Reid: https://www.w3.org/TR/ATAG/#part_b
Wendy Reid: so just…
You know, hopefully everyone's done their done their homework, but if not, that is okay. Part B of ATAG 2 focuses on.
supporting the production of accessible content.
So…
Here, we're focusing on the production of content, so when we're thinking about each of the requirements, each of the guidelines.
and…
And, you know, what needs to change, or what needs to be considered when we incorporate AI.
Think about that. Think about that side of it, the production of content.
So…
Starting from principle.
B1, which is fully automatic processes produce accessible content, and 1.1, ensure that automatically specified content is accessible.
and 111 is, authoring tool does not automatically generate web content after the end of an authoring session, or authors can specify the content be accessible web content.
Wendy Reid: The authoring tool does not automatically generate web content after the end of an authoring session, or, authors can specify that the content be accessible web content (WCAG).
Evelyn Wightman: The difficult and scary thing with ensuring that automatically specified content is accessible is, to me, the non-deterministic part.
What if it's accessible 95% of the time? Is that enough?
Wendy Reid: It's a good question.
Shawn: @MJ - Mary Ann Jawili in case you missed it: The ATAG group is meeting at W3C TPAC the end of Oct in Dublin. CG participants are welcome to join in person or remotely — you don't need to be a W3C Member. Most other meetings are also open to non-Member "observers".
Jutta Treviranus: Yeah, I'm thinking it might be good to look at, give an overview of the strategies that were there in ATAG 2.0, and then think about how that relates to the current automated processes, or.
supports or, predictions that are occurring with, certainly with AI tools. I think there's quite a change in what are the strategies available to us back then and what the strategies are now. so the… I mean, they're all still relevant.
But the relationship and the amount of control the author has and how that automation occurs is very different. So back at the time, of course, of the fully automated.
that was just simply like adhering to a specific set of roles, rules regarding HTML regarding. So it was a very different form of automation. It wasn't a predictive automation. And I think the.
The landscape has certainly changed.
Hidde de Vries: Yeah, I think the word automatic has completely changed since that was written, basically, as Jutta said. So I think more than ever, it needs a very good definition of what we mean and what we don't mean by that.
maybe at the time, people already had different, ideas about what they meant by automatic, but I think now that was wildly.
changed as well.
Maybe some of it, if I can take a guess at some of the difference, some of it I think is…
That at the time, some of the automated bits were putting in markup and structure, whereas.
A lot of what we're automating today is also the actual words inside of the of the content. And yeah, there's probably a huge difference there where.
also what that means for accessibility, because often, especially in terms of WCAG, there is very little that words alone can change about accessibility. it's much more in the markup. So yeah, I feel like…
Defining automatic there or maybe even being a lot more specific with the specific type of automation we mean is going to help a lot. So if we talk about generative AI or something like that, or generation of text or whatever.
Maybe I'll avoid the word automation altogether.
Wendy Reid: Charles?
Charles Hall: I agree that we need to look at what the definition of automatic means now, given these considerations, but I also think we need to do the same for the word generate.
because, automatic translations.
I wouldn't call generation.
Because the content already exists, it would be transforming versus generating.
Hidde de Vries: +1 re 'generate'
Wendy Reid: You there?
Jutta Treviranus: Yeah, back when we developed this, what we meant by automation was that the author had no control, that it was the tool that basically did something by default. So I think one of the elements of.
whatever we redefine or replace automation with is the sense of what is the tool doing without asking the author, without giving the author any choice. So it was. It was the default settings that.
occurred.
Wendy Reid: Yeah, so maybe, I guess with that in mind, let's back up a little bit to, so the part B, I think the section you were thinking about.
of clarifying a bit was the conformance applicability notes, because we have all those definitions that kind of start the section off.
And I like looking over them like some of these kind of immediately.
jump out as probably needing some revision.
you know, we have author availability.
Any Part B success criteria that refers to authors only applies during authoring sessions. I think that's okay. Developer control.
Part B success criteria only applies to the authoring tools as provided by the developer. This does not include subsequent modifications by parties other than the authoring tool developer.
third-party plugins, user-defined templates, user modifications of default settings. I wonder if, this one in particular is, there's now a kind of a third… a third…
A second party? Because there's, you know, the developer who's created the AI.
Evelyn Wightman: ‘Generate’ -> also define for page structure. Putting content in a pre-defined structure vs assembling a structure from pre-defined chunks vs generating a new structure wholesale.
Wendy Reid: LLM interface, there's the author, and then there's the LLM itself, and what it can do in between.
That is not fully in control of the developer.
Lisa?
Lisa Liskovoi: Hi, yeah, reading the requirement, I'm also wondering if some authoring tool creators, as is, would make the case that it is met, because authors can specify that the content be accessible, but I think what we want is to add an additional kind of layer where.
The tool may potentially like follows up or asks for the information that it needs in order for that content to be accessible. So there needs to be, if we're going back to something we had discussed, which is that we want people that don't.
aren't necessarily experts in creating accessible content to still be able to create that accessible content. Then the tool needs to support that input of information, or that request.
for the content to be accessible.
Hidde de Vries: That's a great point. I wanted to add that maybe the word authoring session here also needs updating. I'm consulting with a, CMS, at the moment that.
has a lot of Agentic workflows around content generation, so it can do things where you would tell it, oh, here's a list of all the.
products we sell, turn it into content that looks like this or something.
And you don't know when that session starts or ends, it does that independently of…
what we would call the author.
And the boundaries of kind of when something is an authoring session kind of changes. So it's not a person typing words into a text field.
but it is some kind of tool in the background that.
Fills in different spots in a database, basically.
So authoring session may be another one that needs updating.
Lisa Liskovoi: Yeah, I think specifically defining the end of an authoring session, because, and Yusuf, I'm actually wondering if you can maybe speak to the definition, the original definition of the end of the session, because I think now we'll have to look at it as anything that happens between.
when the author initiate… the beginning of the session, I guess, is a little bit easier to define, but anything that happens…
until and up to the content being created, even if the… so it more so becomes…
I think rather than the end of the session, the parts of the session where the author is not actively involved in the conversation.
Excuse me, in the creation of the content?
Jutta Treviranus: Yeah, I think this gets back to our question about the whole process being much more dynamic and the fuzzy boundary between authoring and the user interface and the user agent, et cetera.
Which, yeah, is going to be a difficult thing to tackle.
And the opportunity perhaps for post-authoring interventions to make it more accessible by the user themselves, et cetera.
Wendy Reid: I wanted to, go back a bit to something that Evelyn said, because I think, it's really important, is, the deterministic.
Because I think, one challenge we have now is…
Pre pre LLMs and those kind of interfaces.
even if accessibility tools were limited, or, accessibility checking was limited, it was deterministic, right? You could check if something was… had appropriate contrast, yes or no. You could check if…
there's appropriate semantics, yes or no. but now, you can say…
to your LLM, make this accessible.
But it may or may not do that because there's no deterministic. Now you can check the output using deterministic means.
If you don't know, like for instance, if you're not an expert and you just say make it accessible.
that doesn't…
guarantee anything.
And if anything, it might almost be deceptive because, you know, most of these tools desperately want to please, and so you may say, hey, you know, make this accessible for me, and it will say, yes, I've done that.
And you have no, aside from like maybe a deterministic checker, like a separate step in the process.
That you need to know how to do.
You have to trust.
Hidde de Vries: Yeah, I think that's a really great point.
I think the… reading the, sort of, last sentence of the success criterion, where, owners can specify that the content be accessible, just reading that with…
and like agentic tools in mind.
they… they will say something is successful, while it's not, like you were saying.
So I think that's…
maybe before, when ATAG was written, that was something that you would hope would work, but now that's even more something that you would really hope that would work, and where it's really hard to actually know whether that happened.
so,
Yeah, I think…
we probably wanna make that part of the criterion better, where it says owners can specify that the content be accessible content. That seems, too easy for… it's easy to tell a tool to do that, um.
a non-deterministic tool, it's really hard to know whether it's on it. So maybe that needs to be more focused on some outcome, like that the content meets WCAG or whatever, for lack of a better way of describing that.
but being accessible is, I don't know, something I would feel scary to tell an LLM to do, because you… you literally don't know if you're gonna get that back or not.
Wendy Reid: I do think this is where I actually really like the usage of terminology like deterministic, because, I mean, that's… it's recognized within, it's well known within, the AI
sphere around deterministic.
a non-deterministic, results. And so, it might be adding something to something like 111 that is
Authors can specify that content be accessible, and the process must prove that deterministically, or must have a means to prove that deterministically.
Wayne D: Great. I,
I haven't. Actually, when I use AI.
And I want to have an accessible output. I have a whole…
project, which is called Accessible, and I'll create a document, and I've given it, you know, I've had it read WCAG very carefully.
And…
And then I…
Prepare a document, pass it to it, and…
You know, the LLMs seem to be intelligent enough to recognise most of the structures that we're looking for. And I tell it to be deterministic. I don't tell it to, you know, I mean, and it…
and to look for certain things. So I'm not sure if,
AI is or if accessibility is really outside the range of determinism, except for certain things that, of course, you have to understand in a much deeper way than AI can.
Wendy Reid: So…
I think, yeah, this is interesting because,
There's been some, I think, really… It's tricky with, I think, LLMs, because…
Charles Hall: Despite what is prompted, it cannot actually make a determination
Wendy Reid: There are cases where it does perform incredibly well, but I think, Wayne, in your case, it's because you've given it.
a working example that meets what you're looking for.
to work from, because, I'm trying to find the link. I think we've shared it here before, but, the,
Microsoft.
LLM evaluation or AI evaluation project. Michael Fairchild did a kind of in-depth test.
of the differences in performance of different models when you gave it kind of essentially three different levels of input. And so models perform incredibly poorly on accessibility tasks when you just say, make it accessible.
they perform slightly… I would say slightly, but significantly better when you give them kind of, a medium level of instruction that includes some details, and they perform.
quite well when you give an incredibly detailed set of, you know, these are the behaviors and the expectations that I have for you.
Evelyn Wightman: https://microsoft.github.io/a11y-llm-eval-report/
Wendy Reid: And I think that is, that is the key, and it would almost be, saying, I don't even know how we'd work this into a requirement. It might be saying to authoring tools, hey.
For your the type of content you author, you might need to have.
accessible models, or accessible, like…
Examples to work from.
And it's for your tool to know exactly what to do. Because if you're just like raw putting in Gemini or ChatGPT.
Any of the models, if you just put it in as is, it will not perform, especially if you just give it.
Wendy Reid: Going back to some of these notes, so we talked about after the end of the authoring session.
Authoring systems, I think, that's broad enough.
Multiple authoring roles. This one hit me as…
I guess worth talking about in the sense of, does the…
You know, we have multiple author roles with different views and content editing permissions. Does this…
Charles Hall: My understanding is that despite a rich context window, a highly specific prompt, and loops of agents checking agents, the outcome is still essentially plausible fiction through expensive guessing.
Wendy Reid: Does an LLM rise to the level of a role?
Does it does it have a role in the.
Hidde de Vries: Probably. I mean, it depends on what you use, but some agents could have a role, I guess, could have content that was generated by it.
an agent. So it kind of depends, but I guess the…
The change is that it is possible to have non-human roles, maybe. Not everyone will have that. Some will, and it's probably good to think about it in there.
Wendy Reid: I was thinking, too, with roles, because, part of the definition of roles is each with different views and content editing permissions.
I was thinking too, different tools, different LLM.
Our agent tools will… you can kind of… I always run mine with, really high…
Barriers.
So, it has to ask me permission before it does anything.
Versus… I know, you could just have it, run in auto mode and do everything, and then you just check it when it's done.
but that's essentially, an authoring permission level. either I want to see everything you're doing and approve it along the way, or you run, and I'll look at it when you're finished.
So it might just fit in without any modification at all.
And then the last one, unrecognizable content. When success criteria requires authoring tools to treat web content according to semantic criteria, the success criteria only apply when these semantics are encoded programmatically.
Hidde de Vries: Can someone explain this one? I don't think I understand what is meant by this at all.
Wendy Reid: Jutta, do you know the background for this one?
Jutta Treviranus: Which part of that? So what we had this notion of programmatic content creation, or it actually…
Wendy Reid: Unrecognizable content: When success criteria require authoring tools to treat web content according to semantic criteria, the success criteria only apply when these semantics are encoded programmatically (e.g. text describing an image can only be considered a text alternatives for non-text content when this role is encoded within markup).
Jutta Treviranus: the term content in and of itself here, it needs to be redefined as well, because, of course, the primary criteria for accessibility at the time was the markup, and so…
all of this has very, very significantly changed. But by programmatic, we mean that by programmatically determined, we meant the developer. So the author has no, determination, or the author has no control over the particular choices.
And so the, which part of, if you can ask further, what, what, what part of this is most confusing? Because all of it, of course, is to some extent different now.
Hidde de Vries: Yeah, I think that's probably where…
my confusion is that a lot of it is different. I mean, I know in WCAG 3 we're still saying programmatically determinable in a lot of places.
So that kind of holds up.
Maybe my main confusion is, what makes it unrecognizable? Is that basically…
The fact that it doesn't have…
Accessibility semantics like that.
It's content that isn't in any way.
Given the semantics that are useful.
Jutta Treviranus: Yeah, that's a component of it. So it's the unrecognizable component.
Hidde de Vries: What's that?
Jutta Treviranus: that you're talking about? So, yeah. I mean, so basically, it's looking at the interaction between how much is under the control of the developer of the authoring tool, and the rules that are available to them, and how much is under the control of the author.
And I think Wayne has some further clarification.
Wayne D: I think one of the reasons that programmatically determined was so important, especially at the time, and still is, is because.
Remember, we're talking about people with disabilities, and the only way they have of accessing the data is through.
The intervention of a program.
So that as you author, as you use the content, as you have the content interact with.
Say a CMS or an LMS.
At every stage, you have to be able to understand that your user is…
Using a program to interpret all the data.
And so if a deterministic program cannot.
Interpret this data and transform it into a format that the user can use.
It's,
It's not usable data.
And we're… we're still at the point where we don't have AI.
assistive technology, so I don't really see…
a huge collection of non-deterministic assistive technology at the time. So I think programmatically determined really means.
Deterministic.
Totally deterministic.
Hidde de Vries: All right. Yeah, I think the developer and author differences is also sort of blended in in the last.
Decade or something like.
you can be an author, but adding things into the HTML by options that you're given, that could be an alt text, but it could be, you know, people go into HTML mode of their editor and, changing things there.
or even people hand-coding their websites, like myself. I'm both the author and the developer, and, writing, HTML.
There is,
Former-looking part of this as well, where…
People within Agway have been suggesting maybe we don't need the programmatic aspect anymore because there will be AI browsers that will automatically do all of this. I know there was an effort of looking at WCAG 3 to maybe remove some of that.
And I don't believe in AI browsers necessarily. I know there are some assistive technologies, or many of them, like at CSUN this year, many of them showed AI capabilities. It's often for, text alternatives, but also to make up.
accessibility semantics like roles and properties.
Hidde de Vries: make them up based on visuals and, finding out what's on the page and putting it in there. but also making up for mistakes that authors or developers, make, which is something.
Lisa Liskovoi: But even AI browsers ask for programmatically determinable information (like definition of roles)
Wendy Reid:💯
Hidde de Vries: Alcohols are not, for instance, they're changing a lot.
into, the developer has made something, and then the browser will still update some of it based on heuristics and…
I think they will likely do more of that too, because it becomes easier now.
Wendy Reid: Interestingly, though, I would say
At least a lot of the guidance rate coming out of LLM companies is also that the semantics and and programmatic.
properties are actually.
they're actually encouraging people to use them more in their code because they help the agents significantly. that whole… the thing about OpenAI recommending the use of ARIA.
and semantic HTML, because it's informative to…
And I mean, it's informative in the sense that, I suppose now you can train a model on, okay, well, this is what a button is. You know, this is the button element, this is what it does, and then should it run into a div that looks, visually looks like a button and behaves like a button.
it can say, oh, that's a button. You know, that's different. because it's learned from other examples.
But it is interesting. I was thinking, too, with unrecognizable content.
There's, still cases in…
content generation where the role isn't…
Automatically easy to determine.
I think a lot of these tools are…
you know, they can tell that, it's fairly easy to say when something is body text versus a heading versus an image, but, you know, there's, nuance, whereas, I think the example they give here is an image can be considered a text alternative if it's encoded properly. But, there's multiple types of text alternatives for images. You can have simple alt text, you can have caption, you can have extended description.
How do you, can you program, can you deterministically determine which.
Which thing something is based on.
without user input of saying, oh, I want to use an extended description for this image.
Versus I just want to use plain alt.
It sounds like an opportunity for us to maybe make this a little bit clearer.
I think going back into B.1.1 and the two requirements.
We talked a little bit already about content auto generation during authoring sessions.
and maybe the use of the word generation, maybe, is we have to adjust.
Lisa Liskovoi: https://help.openai.com/en/articles/12627856-publishers-and-developers-faq
Charles Hall: Never learn a button from other examples. 99% are wrong.
Lisa Liskovoi, Wendy Reid:👍
Charles Hall: Author supplied data that can be extracted never cared about which method. Only a method.
Wendy Reid: is it… if the authoring tool… so B.1.1.2 is, if the authoring tool provides the functionality for automatically generating web content during an authoring session, then at least one of the following is true. It is accessible…
content is accessible without author input. It is, prompting, so during the automatic generation process. I feel like we were looking ahead back then. Authors are prompted for any required accessibility information.
The content is automatically checked.
Or that checking is suggested.
Yeah, we're ahead of the curve with the use of the word prompting.
This is probably one that we need to adjust.
to separate.
Automation…
In the old sense, versus…
LLM…
I don't know. We'll have to come up with maybe a new term.
Hidde de Vries: also tthe word 'prompting' :D
Lisa Liskovoi: Replying to "also tthe word 'prompting' :D":
Was thinking the same, we definitely have to change the definition
Jutta Treviranus: Yeah, I mean, the prompting meant something very, very different at the time. It was the authoring tool prompting the author, or giving guidance to the author to do something accessible, to make a choice that would lead to accessible content. So, I think that's definitely a term that.
Charles Hall: Dumb automation will still exist
Wendy Reid:💯
Jutta Treviranus: We either need to clarify or redefine or find a different term for.
Wendy Reid: This is also, I think, where, this concept of deterministic versus non-deterministic is probably gonna come in, because automatic check… automatic checking, you know, we might want to say that, determine, like…
deterministic checks versus, um.
Non-deterministic checks.
Definitely need a new word for prompting.
Hidde de Vries: something like 'ask' the author
Hidde de Vries: +1 to deterministic / non-determnistic
Wendy Reid: I also think…
Charles called it dumb automation, but, we have to find, some difference between, you know, if this then that automation versus generated automation.
Charles Hall: Old automation
Sambhavi Chandrashekar: So we use the TinyMC editor as an HTML authoring tool in the LMS, and it has an accessibility checker built into it.
Lisa Liskovoi: Artificial incompetence?
Sambhavi Chandrashekar: Is that the kind of checking that's intended?
in the current scenario, I'm saying, I'm not talking about the AI-generated one.
Wendy Reid: Yeah, that's like the old, the old way.
Sambhavi Chandrashekar: Yeah, yeah, yeah. Yeah, yeah. I'm thinking the old way.
And it's because we are trying to interpret the 2.0, right? And then trying to make it into 3.0. Because, I mean, I just wanted to bring up in this context that one of our customers, a government customer, wants us today to give the ATAG to them, which it's not regulatory, but they want us, so I'm trying to go through all this. You remember earlier I had shared.
what we did 8 years ago, and then nobody asked for it, so we didn't do ATAG reports at all. But now somebody wants it, so I'm trying to go through the 2.0 and try to create, and then I'm trying to connect with what we have speaking here, but also looking at what is required today.
Without the AI part.
Charles Hall: ‘Automatic checking’ is pretty limited and will still be limited in wag 3.
So we also need to clarify that any automation is limited.
Shawn:❗
Evelyn Wightman:👆
Sambhavi Chandrashekar: So automated checking, does it mean we check the content? To what extent are we required to ensure that the entire content is accessible?
Building in features into the tool to ensure that the content is accessible is one thing. Making sure that the content that's generated, that's created is accessible is another thing. It's like content audit. So I wanted to know, especially from Utah.
Because we are doing it today, now, going back to 2.0, what was the intention?
Jutta Treviranus: So as I said, it's very much dependent on who actually has control. And there was a recognition that authors could make choices that were not in support. The other contextual piece to know.
There was also an evaluation and repair standard at the time, so there was the… both the… there was the requirement that anything that was fully under the control of the authoring tool would be…
would be accessible. So the developer of the authoring tool was responsible for making sure that anything where the authoring tool had full control would.
be accessible, then, and and that, of course, means that it it would apply deterministic rules to to things that that it can determine. Then the next stage, where.
the author has control, that it would be easier for the author to make something accessible than to not, and guidance would be provided, and prompting would be provided to encourage them to be accessible. And then at the last stage, there's.
there should be a opportunity for the author to test and check that what they have produced is accessible, and that also should be fully available, and the author should be encouraged to use that. So it's,
this gradation of,
The guidance is for the developer of the authoring tool. And so it's do whatever you can basically to ensure that the output at the end is accessible. But you're right, there's these two stages.
because the… that first stage is, sort of more assumed, there's less, yeah. It's not seen as checking, per se, but it's just, what you can do, yeah.
Sambhavi Chandrashekar: provide…
providing it. So that's what I wanted to clarify. Because if the client says, Oh, we want to make sure that it's accessible. Whatever we produce is accessible. That's not within the scope. We do have features if they want to add an image.
The system, the tool won't allow them to add it unless they add an alt text or they mark it as decorative. That's one of the things you said. And then at the end of it, when the HTML file is created, it'll check using the automated checker, which is of course an automated checker, which has a rule set, which is not even so good as the axe.
score, but still there is something there. So that's the limit to which we need to check for our authoring tool to be, I mean, to check the conformance of our authoring tool with ATAG. We don't need to worry about whether the content is fully accessible or not.
That's not within the scope.
Is what I'm trying to… I think I got the answer, but…
Jutta Treviranus: Well…
Right. So you do need to encourage the author to actually go through that checking. And yeah, but you're right. What is not under the control of the authoring tool, I mean, you can't require of the authoring tool.
Sambhavi Chandrashekar: Yeah, so only whatever can be automatically checked, and that too by the checking tool in the tool, in the authoring tool, is within our scope. We can't be responsible for the entire content, for the content to be completely accessible.
Jutta Treviranus: Yeah.
No, that's a WCAG requirement, yeah.
Sambhavi Chandrashekar: We cannot. Yeah. Thank you, and sorry for distracting and diverting, but this is some real issue we are going through today.
And I just also wanted to say that ATAG 2 is also being asked for today by customers, which is something very encouraging. Yeah, that's right.
Shawn: @Sambhavi Chandrashekar There are several resources to support ATAG available from https://www.w3.org/WAI/standards-guidelines/atag/ including an ATAG Report Tool. (lead by Hidde) We also encourage you to submit the tool to https://www.w3.org/WAI/tools-list/authoring/
Hidde de Vries: I feel so encouraged and excited that this was asked for. Just seeing if it happened in the wild.
That's great. I wondered if it's, useful for us, when we talk about automated checking, maybe this criterion, but maybe throughout the next version of ATAG is to distinguish between different levels of automated checking.
Because especially if we talk a lot about AI, people will maybe think about, oh, so automatic means you have some very clever LLM based setup that knows everything about the world and how to do accessibility.
But it's also old fashioned, old automation.
and I think it's… it's helpful to actually distinguish all of those, and they can all be helpful. I know of checks that are, one line of JavaScript, and that can…
find out things about your content that are useful to know. And there are things that are like a huge LLM-based setup that is really complicated and really different. I think it's helpful to distinguish between those. Also.
keeping the ethical web principles in mind, they talk about sustainability, and there are also sustainability differences between checks that are one line of JavaScript versus checks that are.
a very clever Agentic, setup that, runs all day and, does lots of fancy things and requires lots of energy to to train and to use and stuff. So that may be useful to distinguish between those and encourage people to use all of them.
And explain when which type of thing is most useful and appropriate.
Hidde de Vries: https://www.w3.org/TR/ethical-web-principles/#sustainable
Wayne D: I think one other thing also that was at the time, and it's still pretty valid, is that.
The content management system or learning management system itself is an author and is is an authoring tool. So.
The, the, the fact that you, you develop and then in, in that time it was mainly websites that draws information from a database and fills in information. So it's up to date with whatever institution you're supposed to work with.
But that system in itself is an author.
And that that has to be constructed in such a way that when you feed up information to the final document, which may be very dynamic, the entire output needs to be.
accessible according to the accessibility guidelines. And I think that's one of the…
aspects of automate, and I think that's, in a sense, what automatically generated meant at the time, or at least that's how I interpreted it when I was guiding people.
Wendy Reid: Yeah.
So we have the…
Yeah, we'll have to come think about like separating out.
Sambhavi Chandrashekar: It would be nice if we had an accessibility Grammarly that would check in real time, highlight what is inaccessible and prompt accessible options.
Wendy Reid: So that I I feel like that's like.
An entirely possible future.
Accessibility Grammarly.
Hidde de Vries: Yeah, we actually have a project and a team that I work with that does that for documents. So you upload a document and it goes, you know, change this, make this better. That should be adding. That should be a link or whatever. And I think that's that's a great.
thing to suggest in supplemental guidance or something.
Charles Hall: Replying to "It would be nice if we had an accessibility Gramma...":
In the voice of Samuel L Jackson
Hidde de Vries: In the voice of Samuel L. Jackson, yes.
Shawn: Are there tools that are already that are doing some of that accessibility linters?
Yeah. So there already are yeah there are some tools and I think it's important to figure out how that fits within our scope.
Evelyn Wightman: Okay.
This may actually be more of a…
B. 2 thing, but.
I think there's — if we're having non-deterministic generation, there's always going to be some risk that what comes out is not fully accessible.
And having the ability to edit the output is…
going to be important.
And that should not be assumed.
I've encountered a thing where it's.
You have an LLM that has made some website.
The human author has some editing capabilities, but they're not in the code of it.
So making sure that the controls that are actually needed to fix whatever issues can be created are there.
complicated pairing.
Sambhavi Chandrashekar: Interesting. So we provide an HTML code view as well in the editor. the authoring tool. So that means that's what you're referring to, Evelyn, right? That it should be able to change something that the…
Final product must be editable at the code level as well.
Hidde de Vries: here's a a11y linter I wrote when I worked for the Sanity CMS https://github.com/hidde/porta11y (it has mostly one liner / super simple checks, nothing fancy)
Evelyn Wightman: I don't know if it has to be at the code level necessarily.
But my…
Example here is you've got a bunch of different text blocks, and the ones that are obviously heading-ish, you can adjust the…
The code heading level, as well as the…
what it looks like. But the ones that are just content, like just body text, you can't adjust that. So that relies on whatever the LLM has generated.
that actually works for that scenario. there's never a case where it is using.
body texts.
as something that you want to make a heading.
Hidde de Vries: (example: it found fake lists and told people to use real lists: https://github.com/hidde/porta11y/blob/main/src/validations/noFakeLists.js#L25)
Lisa Liskovoi: Yeah, I just wanted to give us another example of this, but in a very different context is with image generation. So often when you ask like Firefly and Adobe to generate an image, it will give it to you as a flat PNG. So if I want to tweak the contrast or if I want to change the wording or if I want to move elements around.
I can't do that.
Yeah, just so we're thinking about different types of content as well, this is definitely a problem.
Jutta Treviranus: Just a quick comment that I agree that that is likely part of the intent of ensure that accessible content production is possible, because some of the scenarios that we were thinking of there were.
Evelyn Wightman: Images-as-slides is a great example
Lisa Liskovoi:👍
Jutta Treviranus: Where the author has… there is no mechanism whereby, or the author has no control, even if they wanted to, to change something to be more accessible.
Wendy Reid: A good example, I was thinking, I've been running into this a bunch, just with coworkers, is people using, for example.
LLMs to generate slide decks.
And often what is happening is instead of, your typical slide deck, which would be text boxes on top of… on top of a, you know, an image background, it's…
One big image. One big image with lots of text and diagrams, and there's no…
It is just a flat PNG or, you know, whatever, and it…
I've, called people out on this because I'm hey, not only is this,
hard to… hard to maintain and keep… keep, keep in order. It can't be indexed by internal search engines, because there's no text to index. it would have to use OCR, and even that could be, you know…
Lisa Liskovoi: It’s also an image of text, so a WCAG failure
Wendy Reid:💯
Wendy Reid: It's gotten a lot better than it used to be, but it's still not perfect.
It's definitely,
Interesting.
We have just a couple of minutes left.
So I think for like the, so for B.1.1, and possibly we haven't gotten to B.1.2, but I think B.1.2 kind of similarly.
we're gonna have to look into separating out…
deterministic versus, non-deterministic.
I do think there's like opportunity for…
like advanced checking, or more I was thinking, I don't know if you, you know, when you send an email, and a lot of email clients now, if you say something in your email, oh, see attached image, and you hit send, and it says, hey.
There's no attachments on this email. Are you sure you like? Did you forget something?
and I wonder if, that's almost the next step for… for checks, is, we have the deterministic ones of, hey, there's an image here, you don't have alt text on it, do something. That's deterministic, but there's the, in, you know…
You say something in the text that sounds like there should be something there that there is not. Do you want to do something about that?
You know.
Jutta Treviranus: Yeah, I wonder if the next thing we should do is to look at the definitions in that spectrum of control. Here are the elements or the roles that have control, and here is a definition of control.
that or come up with definitions that make it clear what we're talking about in terms of what has control, whether it's the developer of the tool, whether it's the LLM, whether it's the individual author, and also.
I think in the process, we may need to rethink what we mean by author.
Charles Hall: I prefer to never knowingly allow a tool to read my email to get clues to make a decision about what I intend.
Sambhavi Chandrashekar: Great. Just one more point, because there's always this boundary between human intervention and automated stuff, right? And that keeps shifting as well. So somewhere to scope out what we mean by what we intend.
for the authoring tool to do. Will there be a definition of that boundary? beyond this, humans have to take care. I'm going back to my original thing about content being accessible. We have an authoring tool. We are trying to define it in such a way that it will help create accessible content.
Even in terms of…
alt text. Nowadays, there is no image without alt text. It automatically comes with a machine-generated alt text. That's another issue we have these days.
And we put it on the human, the onus to verify whether it's correct or not. But someday machines might be able to verify as well. So the line is shifting between deterministic.
Determination of accessibility and.
what is left for the human to determine further. So where does the authoring tool's responsibility end?
Wendy Reid: Replying to "I prefer to never knowingly allow a tool to read m...":
Normally yes but I do appreciate being asked if I meant to add an attachment, even if I didn’t.
Jutta Treviranus: Yeah, so perhaps one other component of B.2.1 is that there should always be the possibility for a human to verify, right? That that never be removed.
Sambhavi Chandrashekar: I think that's very important.
It could be contextual as well, depending on the seriousness or severity of the or the importance of the content or criticality of the content.
Wendy Reid: All right.
That's a good spot for us to end. Thank you everyone for participating.
Hopefully. And we'll meet again in two weeks to continue our going through Section B.
other thing that will be happening probably closer to the end of August or into September is that.
We will be having a joint meeting with APA, the Accessible Platform Architectures Working Group. They requested one during TPAC, but there's a lot of conflicts, and it's a rough week to try and cram more meetings into, so I reached out to the chairs to.
do it ahead. Either… it'll either happen ahead, but if scheduling it doesn't work, we might do it after, TPAC. but it's a good chance for us to meet them, and they always give really good advice and ask good questions about what we're doing, so…
It will be. So keep an eye out for that in our future agenda.
All right. Have a great day, everyone.