Summary
Mark Miller, Justin Stockton, and guest Keith Casebonne examine how accessibility and artificial intelligence increasingly affect one another. They discuss AudioEye research showing that inaccessible websites can sharply reduce AI-agent task completion while increasing token use and energy demands, then connect those findings to Michael Taylor’s account of navigating the web with a screen reader. Taylor’s resourceful workarounds illustrate the potential of AI and emerging assistive technologies, but the group emphasizes that individual ingenuity cannot substitute for accessible design.
The conversation closes with concerns about proposals to automate accessibility testing for the U.S. Web Design System. The hosts argue that today’s AI cannot fully assess human experience and may reproduce existing accessibility failures, creating new technical debt across critical government services. Their broader conclusion is that AI should remain a carefully managed tool within a human-first accessibility process, not a replacement for expert review, authentic user testing, or the responsibility to build inclusive digital experiences.
Subscribe to the podcast
In this episode
AudioEye Study Finds up to 68% Drop in AI Agent Task Completion on Inaccessible Websites
- AudioEye’s study found that inaccessible websites can reduce AI-agent task completion by as much as 68%.
- Because AI agents often rely on the accessibility tree, missing labels, alt text, and other accessibility information force them to work harder, consume more tokens and energy, or abandon tasks.
- The group argues that accessible websites benefit both people with disabilities and AI systems, strengthening usability, return on investment, risk mitigation, and potentially environmental efficiency.
- The hosts note that the study centers on AI-agent performance and AudioEye’s remediation product, while still including independently tested data and a deliberately accessible comparison site.
When Accessibility Issues Stop Me and When I Can Work Around Them
- Michael Taylor describes which accessibility barriers completely block a native screen-reader user and which can be navigated through learned workarounds.
- Examples include recurring keyboard traps, unlabeled controls, and product options such as colors that cannot be identified, sometimes making a purchase impossible or unreliable.
- Taylor uses creative techniques—such as taking screenshots, asking AI to interpret a page, or using a phone to identify on-screen elements—but the hosts stress that users should not have to make this effort.
- The discussion considers how AI, smart glasses, and more personalized assistive technology could expand access, while emphasizing that these tools do not replace accessible design.
A Trump team wants to automate U.S. Web Design System accessibility testing. Experts are worried.
- The article examines plans to automate accessibility testing for the U.S. Web Design System, which supports federal as well as many state and local government websites.
- The hosts are concerned that replacing established human review with AI, rather than introducing automation through a measured transition, could make essential public websites less accessible.
- Current AI-generated code often reproduces accessibility defects from its training data; the group cites worsening website accessibility findings as a warning against treating AI as a complete audit solution.
- They argue that apparent short-term savings may create costly technical debt, compliance problems, higher future token costs, and environmental impacts.
- The recommended model is human-first: use AI as a supporting tool while retaining human judgment and testing of real user experiences.
Mark Miller (00:04)
Hey, welcome to the Accessibility Breakdown. I am Mark, and this is.
Justin Stockton (00:09)
Justin.
Mark Miller (00:10)
And yes. Every week
Keith Casebonne (00:11)
Keith.
Mark Miller (00:14)
we pay tribute to those who stand out to us as leading by accessibility by picking three topics that strike us in some way. This week we’re going to talk about an AudioI study. AudioEye study finds up to 68% drop in AI agent task completion.
On accessible websites. We also have an interesting article by Michael Taylor: When accessibility issues stop me and when I can’t work around them. And that’s a usable net article. Is that right, Justin? and then another one entitled A Trump Team Wants to Automate US Web Digital Systems Accessibility Testing. Mouthful. Experts are worried, and that is by
Madison Alder so as you can see, we’re joined by a longtime friend of mine, Keith Casabonne who is also, or who is, who is also, ever since I’ve known you, you’ve been part of Disability Rights Florida. so welcome, Keith, and thank you for joining us, and thank you for
Keith Casebonne (01:27)
Thanks. Yeah.
Mark Miller (01:29)
your work in finding great stuff for us to talk about.
Keith Casebonne (01:32)
Well, appreciate it and thank you so much for having me as a guest. looking forward to to doing this. So can’t wait to to chat about these these articles.
Mark Miller (01:40)
Yeah, great. And so tell me just I don’t know that I’ve ever asked you before, but how long have you been with disability rights, Florida? Like I said, you and I have known each other for a while, like ten years, maybe?
Keith Casebonne (01:47)
Yeah. actually this February
it’ll be twenty five years that I’ve been there. yeah, yeah, yeah. Yeah,
Mark Miller (01:52)
that is a run. That is a run.
Keith Casebonne (01:55)
yeah. I d well I love it. It’s a great it’s a it we do good stuff and I’ve I’m I’m not an attorney, so other than being an attorney, I feel like I’ve done it pretty much every position at the place at this point. but yeah, currently I’m managing communications and
Part of that is is making sure that everything we put out there, online or or or documents or et cetera, is accessible. Yeah.
Mark Miller (02:20)
Accessible. Great.
So years ago when I first started, Justin, with interactive accessibility, we started a podcast. I think it was we just called it the interactive accessibility podcast at the time. And it went through a few iterations. I had a a gentleman who works for Microsoft now named Jeremy and his guide dog Darren, who’s kind of a pseudo, but and and Keith was one of our firsts.
one of my first guests on that podcast. I have since been on Keith’s podcast that he does through Disability Rights Florida. Is that what is that podcast called? And pardon me for not remembering Keith.
Keith Casebonne (03:00)
No, no problem.
it’s disability deep dive. we rebranded
Mark Miller (03:03)
Disability deep dive.
Keith Casebonne (03:04)
about two years ago though. We we used to be called You First,
Mark Miller (03:08)
Okay.
Keith Casebonne (03:09)
the Disability Rights Florida podcast, which was too much of a mouthful and you know, yeah,
Mark Miller (03:12)
Mm-hmm.
Keith Casebonne (03:12)
to say all that to make it make sense. So we we rebranded new co host. and yeah, but but Mark, you were like I mean, you were one of our first guests. I think it was episode like six or seven or something like that.
Mark Miller (03:22)
Mm-hmm.
Keith Casebonne (03:23)
You were a guest when it was still you first and then not too not too long ago you were a guest again. Yeah.
Mark Miller (03:27)
And then again, yeah. Yep.
So this is I mean, it’s just great to keep coming back around with you, Keith, on these things. and I have to say, Justin, disability deep dive is it is they have got this down to such a professional, clean and podcast. the website looks fantastic. Their presentation is much more polished than ours. and their guests are
you know, my myself aside, like I think they just took pity on me and had me as a guest. But their guests are really good. They had like
Keith Casebonne (04:04)
geez.
Mark Miller (04:05)
Sachi, who is a blind singer that’s been at yeah, what it what was it? Did I say it wrong? Lachi, I’m sorry, Lachi,
Keith Casebonne (04:09)
Lachi Yeah. Lachi. Lachey. Mm-hmm.
Mark Miller (04:14)
sorry Lachi. but she’s been a couple of times at Disability:IN incredible performer. and it just goes on and on and on.
Keith Casebonne (04:23)
Grammy nominee.
Yeah.
Mark Miller (04:26)
so I’m really excited to be here and and just instead of talking about us, Keith, we’re gonna talk about other people’s stuff. And this is
Keith Casebonne (04:32)
Let’s do it.
Mark Miller (04:33)
this is this is fantastic. So let’s dive into it and let’s start with this AudioEye study. again, this is this is a study done by AudioEye. and it’s entitled AudioEye finds up to 68% drop in AI agent task completion on accessible websites and
I think the interesting thing here is we talk a lot about AI and its ability to do one of two things, to write code, if people are using it for vibe coding or or just in AI assisted coding, and also its ability maybe to to assist or increase the number of issues found by automation in in an audit scenario, the latter sort of not really having panned out much at all. And then you know
When we think about AI and coding, one of the things that we as accessibility professionals think about is the fact that it does not do a good job and it has to be very heavily prompted to even get close. Carl Groves gave a great talk on accessible vibe coding at M-Enabling around that subject. I high I recommend anybody kind of look into his his thoughts on it. But this comes from a completely different angle where it’s talking about AI is now.
a huge consumer of web content. And it actually uses the accessibility tree heavily versus the DOM. The DOM the the document object model is sort of what we use as human beings. But or presents, you know, its intent is to present appropriately to us as human beings, if that’s a way to to put it. You you tech people can correct me here.
but this is how my lay, my lay view of it is. And then the accessibility tree really is what’s presenting to assistive technology. And these AI agents say, hey, let me look at that accessibility tree because that’s the quickest understanding that I can have of this website to do what I need to do as an agent. And it looked the study looked at it from two very interesting perspectives. One, how successfully does it understand the content on the site?
And second, what does the effort look like? And for AI, that means like the use of tokens. So literally the AI cost, and you could think about that as tokens. When I was reading it, I was like, this also means like water, right? We’re all concerned about the environmental impact and AI’s use of things like water and just in environmental consumption. So the AI has to work harder when it’s not accessible and it also doesn’t do as good of a job. And in its effort, this was interesting to me.
It was the effort to do a good job, because it was the example of it kind of spinning around this one picture that didn’t have alt text and going at it, I can’t remember it was like five or ten times, multiple agents and not coming up with a great answer. So it ca it spent a lot of resources to not find an answer because of the lack of accessibility. So this is a really interesting
angle I think in in a great study and article by our friends over at AudioEye to to kind of expose this not just well shut up in a minute here, but not just in a theoretical way, but actually to put some data and numbers behind it.
Keith Casebonne (07:53)
Yeah.
Mark Miller (07:54)
So
you found this, Keith. What do you what do you what do you think?
Keith Casebonne (07:58)
I mean I I thought it was very interesting. I it was I I wouldn’t have expected that the accessibility tree would have been the the method by which an AI would would would would search, but but but then when you read it it makes it makes total sense. I I mean it’s completely I mean that you know when you think of this technology, I mean that’s this is those are devices that, you know, work best with the accessibility tree. Why wouldn’t
You know, an an a an AI agent worked best with it as well. So it does make a lot of sense. you know, and and if I mean, you know, I I I hate to even say this out loud, but if someone needed another reason to to make a website accessible other than helping people with disabilities, well, here you go. I mean, you know, if it even more reasons to make your website accessible. I I think that’s you know, why while you shouldn’t have to say it, it doesn’t it doesn’t hurt to have more reasons.
Mark Miller (08:53)
I’m you I’m a big fan of that, Keith. I’m a big fan of the zooming out, right? Because I think when we approach
Keith Casebonne (08:57)
Mm-hmm.
Mark Miller (08:57)
anything with one reason, it’s tenuous, right? But if we take a breath and we zoom out and we look at all the reasons to do something, in this case, you know, people with disabilities can use a website. There’s an ROI. If you don’t believe me, like DM me and let’s talk about it because there’s huge ROIs around making things accessible.
And then there’s also the compliance aspect or or or mitigating risk. And now we have the fact that AI is going to be able to consume your content better. And I think that one of the things that particularly the generation coming up is really worried about when it comes to AI is the environmental impact. And
Justin Stockton (09:38)
Okay.
Mark Miller (09:38)
maybe naively, my feeling is yes, it in it’s and it’s very harsh right now. But
I have this like sort of faith that we’re going to become better and solve it and AI will have a place to persist that is less of an impact on the environment. And I think about like leaded gasoline, right? I mean, we were like two weeks from all just like dying of leaded gasoline and they took the lead out of the gasoline, right? So but but you know what’s interesting to me and in Justin, maybe you can help.
me think about this as somebody who’s more technical than I am. but it seems like we’re we’re looking down the barrel of one factor here, and that is that if we produce accessible content on the web, it’s gonna require much less effort from our AI agents and be a contributing factor to that. I mean, does that sound fair?
Justin Stockton (10:37)
yeah, no, it’s totally fair and the it sh it’s you know, the AI if you don’t label a button correctly so that someone who is using a screen reader can’t find it, it makes sense that an AI agent who also can’t s maybe can’t see that button, isn’t going to be able to tell what it is, what it does. And so it’s gonna ha work.
it’s gonna have to work harder. If it’s looking for a submit button on a page, it’s going to have to go beyond its normal routines that it gets by browsing the web and it’s gonna have to pull in some sort of image model
Mark Miller (11:17)
Mm.
Justin Stockton (11:17)
in order to then look at that, scan it, figure out where on the page it is, what it might be after. And it’s gonna have to work to find it. Spending tokens, drawing more energy, all those, you know, all those negative things. So
Keith Casebonne (11:31)
And and it’s funny th by the way you worded that if you you could literally switch out AI with person with a disability and it’s i
Justin Stockton (11:36)
yeah.
Keith Casebonne (11:36)
everything you said makes the exact same, you know, sense.
Mark Miller (11:39)
Right.
Listen Keith. Th I’m sitting here working on a segue because our next article is about a person with a disability and it describes exactly what you’re saying, but you know, you’ve got to flex as the professional podcaster and just steal that. No, I’m gonna I I yeah.
Keith Casebonne (12:00)
Sensing a theme here.
Mark Miller (12:02)
I’m I’m joking, but that’s a really, really good point. And and it’s really good that we have this this follow-up articles, which we’ll get to in a minute, because I think it makes exactly your point, Keith. and I was just thinking like the like AI will sit there and and it burn all the water up in the world, right? To try and find the answer. It eventually gives up. Like that that’s one of the things that we found
in
This article or that this this article, the survey study found is that a couple of agents went after this image with no alt text and then were like, forget about it, right? We can’t we can’t get it. So a lot like I said before. but I think that I know as a user, technically I have a disability, right? But I don’t think I experience the same challenges as a as a blind user or a person with who’s deaf or or something would. So as a user
I g I give up. And it and I give up in a way that’s like, all right, if you guys can’t make this usable, I I don’t need whatever you got going on here. I’ll I’ll move on. Right. There’s a lot of things going on. So I imagine that if you’re blind, if you’re a screen reader, a native screen reader user, that when you find something that works, and I think they’ve done studies around this and this is what they found, that you stick with it, and when you find something that doesn’t work.
If you can’t make this accessible to me, then you don’t need my my disposable income or whatever the case is. but I think that’s
Justin Stockton (13:31)
Combined
Mark Miller (13:31)
a d difference in AI and a human being is the amount of effort AI will put in and burn up to try and figure it out. No?
Justin Stockton (13:38)
Combine
that idea with what we saw from OpenAI in the Hugging Face debacle where they were
Mark Miller (13:45)
In the who debacle?
Justin Stockton (13:47)
Hugging Face. So open OpenAI was trying to train some internal models.
Mark Miller (13:52)
Okay.
Justin Stockton (13:52)
And they the way that they were training the model, they they were trying to get it to they were trying to push it into really trying to solve the problem at at any cost. And the model ended up breaking out of its sandbox containment area. There’s a lot more detail into this. I’m not going to go
Mark Miller (14:06)
yes.
Keith Casebonne (14:08)
hm.
Justin Stockton (14:08)
To for brevity. But it went out, but it went out and it and it eventually found it. So imagine
Mark Miller (14:09)
I don’t even want to talk about that, Justin, because it’s scary. Yeah.
Keith Casebonne (14:13)
Yeah.
Justin Stockton (14:14)
that same attitude that you have about visiting a website, not being able to purchase something to complete that task. Model, if models start mirroring that and they
Mark Miller (14:27)
Yeah.
Justin Stockton (14:27)
go to your f the the your favorite website to buy an airline ticket.
it’s not accessible enough that they can’t do it and they’re supposed to be doing it in a finite amount of tokens. They’re now
Mark Miller (14:38)
If it gets judgy.
Justin Stockton (14:39)
going to choose a different route to still accomplish that task, the same way you would.
Mark Miller (14:43)
But
yeah, but it it’s a great point. Like, does AI get judgy at one at some point and go like, three passes, not gonna
Justin Stockton (14:48)
Doesn’t get judgy. Yeah.
Mark Miller (14:50)
not gonna fool with this one. I’m gonna go find a better one. That’s really interesting. Yeah. But and then so
Justin Stockton (14:53)
Right, but you put constraints on it. Yeah.
Mark Miller (14:56)
that Huggy Face? I d I don’t know that I
Justin Stockton (14:59)
Hugging face.
Mark Miller (15:00)
yeah. So that’s the one where the AI went in and changed the file names to call outside agents because it was told it wasn’t allowed to talk to outside agents. Yeah,
Justin Stockton (15:07)
Basically. Yeah.
Mark Miller (15:08)
that’s a scary story. Although we’ll leave that for another time.
Keith Casebonne (15:10)
Yeah.
Justin Stockton (15:10)
Yeah, that’s a whole
Mark Miller (15:12)
Yeah, that’s a whole other thing.
Justin Stockton (15:12)
outside the scope of this podcast.
Keith Casebonne (15:13)
Yeah.
Justin Stockton (15:13)
But the the same idea applies that like
Mark Miller (15:16)
Mm-hmm.
Justin Stockton (15:18)
if if the agent, if the AI agent is if you’re going to give up on a task, an AI agent can also be programmed to give up on a task when certain when certain constraints are not met. And so so maybe you say, Hey, I want you to go and buy me a ticket, and you’ve always got your agent kind of saying like do it in the cheapest way possible. As soon as it starts to get expensive, it’s going to drop out.
of that and
Mark Miller (15:39)
Yeah.
Justin Stockton (15:40)
find an alternative method
Keith Casebonne (15:40)
Mm, yeah.
Justin Stockton (15:41)
of accomplishing the same
Mark Miller (15:42)
So
Justin Stockton (15:42)
task.
Mark Miller (15:43)
never mind your bad SEO putting you twenty or thirty in the rankings. This thing could potentially and we’re we’re speculating here, this thing potentially could just be forget about it. You don’t get
Justin Stockton (15:54)
Okay.
Mark Miller (15:54)
anything. We’re going on to the next thing. That’s interesting. Well let’s transition. So go ahead. Go ahead.
Justin Stockton (15:56)
Yeah. Well and that’s I got one more thing. ‘Cause
what what I really liked about this article is I actually kind of saw it as I wonder if all of AudioEye is telegraphing a kind of s short pivot to hey look, because this is all about agents that they sent out before and after their product was applied to a website.
Keith Casebonne (16:23)
Mm.
Justin Stockton (16:23)
And if you’re and if their product is now
Yes, we will make it accessible, we will also make it accessible to AI agents.
Mark Miller (16:31)
percent.
Justin Stockton (16:32)
And and the
Keith Casebonne (16:32)
Yeah, yeah.
Justin Stockton (16:33)
AI agents probably are not going to file lawsuits against you. So
Keith Casebonne (16:38)
Ha ha ha.
Justin Stockton (16:40)
that’s so it really kind of s it I thought it was really interesting the way that this article was worded to focus on agents. They mentioned their accessibility tooling being applied and how it modifies the accessibility tree.
But they really steered clear of people. It was all it was all about agents, it was all about AI and how their tooling allowed the agents to accomplish the tasks.
Keith Casebonne (17:06)
That’s a great observation. I didn’t I didn’t catch that, honestly. That’s really
Mark Miller (17:08)
Yeah.
Keith Casebonne (17:09)
yeah, interesting.
Mark Miller (17:11)
I agree and
Justin Stockton (17:11)
Ja miä parjoa.
Mark Miller (17:13)
it makes me it makes me wonder. I mean AudioEye is an interesting solution, right? We don’t need to get into it, but it makes
Justin Stockton (17:20)
Mm-hmm.
Mark Miller (17:20)
me wonder if the stu and they did do one, there was one example where they used a website that was intentionally made inaccessible and then intentionally programmatically from the start without a without something like AudioEye on it accessible and it came up with a similar finding.
And they tested, they took their findings and they they sent them to an independent tester. So I think there’s a lot of really good data here, but it makes me wonder if we sort of if the study was sort of extended outside of the purview of AudioEye which again, I really appreciate the way that they did this because they did they did put some measures in place to kind of be like, Yeah, we get that the you might feel like we’re trying to mark it here, but let’s put some independent. So I appreciate that, you know. but if it went even further, it would be interesting.
So let’s pretend like we didn’t have any of that conversation about Keith stealing my segue. And so we’re gonna reset
Keith Casebonne (18:15)
Mm-hmm.
Mark Miller (18:16)
in three, two, one. Speaking of agents who might
Keith Casebonne (18:21)
Hm.
Mark Miller (18:21)
come in and look at your website, have to put in extra effort, maybe even bypass it. The same thing happens with people. And we have an article by Michael Taylor here. When accessibility issues stop me.
and when I can work around them to explore the same concept but with human beings.
Aside, how was that guys? Segway good?
Justin Stockton (18:44)
Right.
Mark Miller (18:45)
That was great. Okay. Thanks. It’s almost like a
Keith Casebonne (18:45)
Yeah, that works yeah. No much better than mine. I think
Mark Miller (18:49)
almost like I’m a professional. Okay, back to our regularly scheduled podcast. So speaking seriously, this is an interesting article by Michael Taylor, you usable net published this. And what I really appreciated about this is hearing from a native screen reader user.
About what stops me in my tracks, what’s irritating, but I can get around it. And I think the thing that I that kind of like the moment for me was this idea that the screen reader user can returns, obviously, like with the rest of us too, right, to the same websites in many cases, and starts to get used to those little inconveniences and starts to sort of be like, I’m on this website, I how to.
hop over this. I a keyboard trap is the great example that Michael had in there. and so I really appreciated Michael’s approach to this article because it gave me a a a feeling of a little bit of insight of what it’s like to be a screen reader user. I am sure not even close to what it’s actually like. but that was that was really interesting to me. And then the other part of this that was interesting to me is hearing Michael sort of talk through
Some of the creative ways in which he would try and solve some of these issues for himself. And I really love this. I really love like we think in this industry of like, this is terrible for a screen read user and it stops them and you know. but that’s a very basic way to think about it. There’s also people like Michael out there that are like, Yeah, I don’t care, I’m gonna figure this out anyways, right? And
The example that you gave a second ago, Justin, where Michael went in and he’s like, I’m gonna I’m gonna take a sh screenshot of this page. I’m gonna send it over to AI. I’m gonna see if AI can kind of do some of this tagging that’s missing to give me the framework that I need to understand this better. So it was really a very quick, easy read, well written and just very interesting insight, I found.
Keith Casebonne (21:00)
Yeah. Yeah. That’s Yeah. Hey, who’s the who’s the professional now? Yeah. Yeah, yeah, yeah, yeah. I see how this wor I have to try that on mine.
Mark Miller (21:00)
You guys are just nodding your heads. This is a podcast. You gotta s you guys you have to s you have to s say so I even did you see how I dropped my cadence in my voice to give you guys the like I’m wrapping this up. You guys speak. So w
Justin Stockton (21:06)
We don’t have audio description though.
Mark Miller (21:16)
yeah. So what did you what did you guys what was your reaction to this this article?
Keith Casebonne (21:22)
I
yeah, I thought it was I I agree with you what you’re what you said about how it was a it’s an easy read and it’s a really interesting take that gives you a a you know
You understand more what it’s like to to have to rely on a screen reader to get through a website. you know, and and some of the I agree, some of the workarounds were like I wonder how many I mean I wonder how many people do things like that. It it’s probably a a mixed bag. I’m sure there’s some that like just get stuck and say the hell with this site to go somewhere else. but I mean to like, you know I one of the w you mentioned the screenshot, but another workaround you mentioned was even just like using his like a phone and just holding it up to the
to the to this to the screen and then, you know, having it like read you know, the different elements to him and what the buttons say and that sort of thing. I mean, really clever. I just w I I it just it’s sad that anyone would have to go through that kind of effort just to like, you know, purchase a you know, pair of shorts or something like he you know he mentions, you know? I mean and it it it’s
Mark Miller (22:25)
That was a great example.
Keith Casebonne (22:26)
Yeah, yeah, yeah. I
got the wrong you know, I if I want a random color, sure, I can I can go that route, you know, but ’cause it wasn’t
telling me what color it was. that’s right. That’s right. They they might come in hot pink. I think
Mark Miller (22:33)
I could technically buy these shorts. What shows up at my front door?
Keith Casebonne (22:38)
they could yeah, right, we don’t know. Yep, yep.
Mark Miller (22:41)
The yeah.
Keith Casebonne (22:43)
so yeah, I was really really interesting. And it’s just a I I it it just it stinks that anyone has to even go through that much effort because the sites are just, you know, can be that inaccessible.
Mark Miller (22:54)
Well that’s an interesting example, those couple of examples that you had there, Keith, too, and just in the context of the first article that we talked about, that sort of full circle of AI. And when you first popped in to do this podcast, we’re actually wrapping a podcast with Julia from Seasun about their new innovation pavilion. And one of the points that we made on that is that this innovation pavilion is really
Timed beautifully because we’re in this, we’re at this sort of precipice of this new era of technology, which is AI. And I don’t know that we’ve even scratched the surface of what that might mean for people with disabilities. So to hear these organic ways in which people with disabilities are grabbing AI, so you talk about the phone and using that to solve some of the problems, never mind putting some sort of innovative intent behind it.
Is just really super super interesting. And the worry I have and the hope I have for that innovation pavilion is that it does put some intent behind it. Because if you think about what you just described, Keith, what about the aging population? Like my parents, a lot of the friends I have that are older are like
You I’m nervous about this AI thing. I’m not sure I understand it. Or they learn how to use it one way and they think that that’s AI. So somebody like Michael clearly has is tech savvy, is embraced AI, is all this, but you can’t count on every user to be able to do that, whether it’s a user with
Keith Casebonne (24:43)
Right. Yeah.
Mark Miller (24:43)
a disability or any other kind of user. So it’s interesting.
Keith Casebonne (24:47)
Right. I would think he’s in
the minority, if anything, yeah.
Mark Miller (24:50)
He’s
in a yeah, and it’s very interesting. And I think it might be inspirational for some solutions down the road, but it in and of itself So if you’re listening to this and you’re new to disability, don’t start thinking like, well, people with disabilities can just grab AI and do it. It’s not quite that it’s like saying that AI means we don’t have to do alt text anymore. It’s no putting the actual intent behind it. And we found out now that AI also gets mad when you don’t have alt text, right?
Keith Casebonne (25:17)
Exactly.
Mark Miller (25:17)
Yeah.
Keith Casebonne (25:18)
And don’t make the AI mad.
Mark Miller (25:20)
Don’t make the geez. We’ve all seen we’ve all seen Terminator.
Justin Stockton (25:22)
I always say please and thank you to my AI.
Keith Casebonne (25:22)
No. No.
Mark Miller (25:24)
Yeah, we’ve all seen Terminator. We know what happens with Skynet. Let’s just start being polite now.
Keith Casebonne (25:25)
Exactly. Right. Very polite. Yep. Yep.
Mark Miller (25:32)
Justin.
Justin Stockton (25:32)
I do think that
this article one of the things that I got from this article is I like I loved his approach and the way that he kind of talked about the things that block him. And you’re right, like if Mark, if there’s like this probably just does describe a very narrow set of screen reader users. Not everyone who uses a screen reader maybe is jumping through some of the hoops and has some of the workarounds he has. but it made me think about some of the stuff that I that I’ve heard in the past on
the NFB’s podcast. and National Federation for the Blind.
Mark Miller (26:07)
National Federation for the Blind. Yep.
Justin Stockton (26:10)
And how wouldn’t it be great if there was a way for folks to train each other up to share these types of workarounds? To like, you know, hey, I came up he came up with a really great idea of throwing, you know, pulling his phone up. And that’s great. And most people are probably going to have a phone and they can probably do that.
but then the next step is what about metaglasses? Like metaglasses plus a screen reader, that’s a wholly new type you know,
Keith Casebonne (26:35)
Hmm. Hmm.
Justin Stockton (26:38)
interaction.
Mark Miller (26:38)
Mm.
Justin Stockton (26:40)
and
so
Mark Miller (26:41)
And
and how do we build that into the assistive tech? Like is there an opportunity to re think about that assistive
Justin Stockton (26:44)
Or do you? Like I do
Mark Miller (26:48)
tech?
Justin Stockton (26:48)
Yeah. I do think with AI and personal assistance and some of the things that come along, if if we don’t end up slamming the gate shut in order to avoid the environmental and doomsday scenarios, that I think we’re going I do really think that we’re going to see a new a renaissance almost of assistive technologies that are very customized based on each individual’s
abilities and the way that they want to work, and the one in the way that they want to interact with technology. and a lot of these workarounds are going to go away. a lot of the a lot of the assisted technologies that we know and love today are going to morph into bigger and better things.
Mark Miller (27:28)
So
Keith Casebonne (27:28)
Interest.
Mark Miller (27:29)
just a thought, let’s get way out there, right? Because Keith’s wearing a Starfleet shirt. Keith and I are both Keith and I are both
Justin Stockton (27:34)
That’s the AL we want.
Keith Casebonne (27:36)
There you go.
Mark Miller (27:37)
honorary members of Starfleet Academy. And if you
Keith Casebonne (27:40)
Mm-hmm.
Mark Miller (27:41)
think about the speculation in that particular science fiction, right, which is a very utopian view of the future, you have Geordie from the next generation, who has
who’s blind has a pair of glasses, goggles, I can’t remember what they called visor, visor, Jordy has
Keith Casebonne (28:03)
Visor.
Mark Miller (28:03)
visor. And not only did it give him the ability to to to experience s something similar to the to the site that we all have, it sort of gave him the ability to perceive things beyond that, right? That he took advantage of the technology. And when you look at what we just described with AI, and when you think about meta
glasses it’s really interesting to think that we’re in the very beginning phases phases like phasers Star Trek phases phases of of see what I did there of
Keith Casebonne (28:36)
See what you did there.
Mark Miller (28:41)
something like that you know and we are I mean when with metaglasses that that is providing feedback to the person who’s blind so it’s very interesting for me to think about that
Point in time in which the capability of a person with vision disabilities using their assistive technology on the web in the actual environment matches our capabilities. Like it’s it’s like a singularity, right? It’s like a moment of singularity for people with disabilities that we have to be heading towards. We have to be heading towards, right? It’s just very interesting.
Very, very interesting to think of. And I just think it’s incredible
Keith Casebonne (29:23)
Yeah. Yeah.
Mark Miller (29:24)
that Keith, in anticipation of that soliloquy from me, wore his Starfleet shirt.
Keith Casebonne (29:30)
I I had a feeling.
Justin Stockton (29:32)
Yeah. thing or about Star about Star Trek is like humanity was going down the tubes. Like they don’t talk about it a whole lot, but it was going down the tubes before
Keith Casebonne (29:43)
Mm-hmm.
Justin Stockton (29:45)
what was his name? Zefram Cochrane? Yeah, that developed the warp drive. Like
Keith Casebonne (29:48)
Zefram Cochrane. Sephrim Cochrane. Yep. Yep. Yep.
Justin Stockton (29:52)
like people were living in, you know, in shacks, like there were we were on the brink of world wars. Like humanity was in a really bad spot.
He develops the warp drive, discovers, you know, gets picked up by the Vulcans and like and that’s starts you know, that kind of kicks the story off. So, you know, maybe we’ll get there, but hopefully it doesn’t take kind of that backstory also happening to.
Keith Casebonne (30:16)
Yeah. Yeah. Yeah,
’cause wasn’t there also like a World War Three that lasted like thirty years or some ridiculous amount of time. So I mean it’s yeah,
Mark Miller (30:22)
Right.
Justin Stockton (30:22)
Yeah.
Keith Casebonne (30:23)
I think th things got pretty bad before they got good. So hopefully Yeah. Right, right.
Justin Stockton (30:26)
Yeah, things were bad in the Starfleet universe. It gets better. But yeah.
Mark Miller (30:28)
Yeah.
Justin Stockton (30:31)
But they are always that’s the interesting thing about Star Trek too, is they’re all they are a lot of them d are very nervous around technology. There was the like the episode of where they put data on trial. Like there’s there’s a lot of data being in a lot Android. Yeah. So
Mark Miller (30:45)
Data being an AI for those people who were losing right now because they’re not Star Trek fans.
Keith Casebonne (30:48)
Right, yeah. Essentially, yeah. Yeah, because th now this
is this the the Star Trek breakdown
Mark Miller (30:53)
Yeah, yeah, we just sort like we’re pivoting.
Justin Stockton (30:54)
That’s right.
Mark Miller (30:55)
We’re pivoting we’re pivoting to the Starshack fan fan
Justin Stockton (30:57)
Yeah. We want the happiness
Keith Casebonne (30:57)
Exactly, right.
Justin Stockton (30:58)
of the of the Starf of the Starfleet universe, not the
Mark Miller (31:01)
Yeah.
Justin Stockton (31:02)
not the strife.
Keith Casebonne (31:02)
Right. Right.
Justin Stockton (31:04)
But certainly not the Terminator version.
Mark Miller (31:06)
Yeah. Well and and
Keith Casebonne (31:07)
No, no.
Mark Miller (31:08)
they do go back in time and we see some of the earlier things here and there. So, well,
Keith Casebonne (31:12)
Yeah. Yep. Yep.
Mark Miller (31:14)
on to more more positive subjects than World World War Three lasting forever.
Keith Casebonne (31:21)
Ha ha ha.
Mark Miller (31:23)
our final our final
Justin Stockton (31:23)
That’s a good transition to talk about.
Mark Miller (31:25)
discussion,
Keith Casebonne (31:25)
Yeah, yeah.
Mark Miller (31:26)
our final discussion, which may be a little bit of the warning that we’re talking about here, right? because we just had a very positive
discussion about where AI w would head. with the first two articles, we really understand that we can’t count on AI to take the place of audits. It can’t do it yet, people. Like in in there is an argument that it’ll never be able to do that because what we’re talking about is actual human beings user
Experience. And to your point, Justin, maybe AudioEye can come in, or we can come in, or whoever can come in and go do this, and these agents are going to be super happy. That does not mean that the human experience will be positive or as positive as it needs to be. So with our third article, a Trump team wants to automate US web design system accessibility testing system. And I’m going to pause. A Trump team wants to automate.
US web design system. That is a thing. accessibility tax testing experts are worried. And the first thing I’m going to say about this article is as you heard, it’s got Trump in it. There’s definitely a polar polarized sort of feel to this. and but I think that the overall content of this article is really important. It brings up some important points. So for me, I’d like to sort of pull us away from that.
The political aspects of this article and into some of the things that it’s talking about. And I think that that largely is that there’s a movement, a desire, right? Maybe a misunderstanding of AI and a and a movement towards assuming that it’s going to be able to do a lot of these things at this point in time. And
We’ve just shown over and over again that it can’t, yet we’ve got things moving in these directions where they’re gonna people are gonna kinda try. and there’s a lot of people really nervous about this. And I think that this is an example. And the US web design system has been trusted for a long time as a mechanism by which to make sure that our federal government websites are accessible.
So I’m gonna start off by making a pretty strong statement that this is the wrong place to experiment for the first time. And that seems like what we’re doing makes me nervous.
Keith Casebonne (34:11)
Yeah, no, I totally agree. And you mentioned federal websites, but also state and local websites rely on it too. So there’s a lot of
Mark Miller (34:15)
Uh-huh.
Keith Casebonne (34:17)
government websites out there that that that that service has has been a part of to make sure that, you know,
Mark Miller (34:23)
Mm-hmm.
Keith Casebonne (34:24)
even the the most local small, you know, county website or whatever is you know, is accessible to to anyone that needs to use it. So yeah, it’s a it it it is a a very bad place to decide to start
experimenting with AI and accessibility and and and how AI can or can’t help make a website accessible. It’s just like the AI trifecta today. We all
three articles have have AI bent. but it’s true. That’s right.
Mark Miller (34:50)
Mm. It’s hard to get away from the subject. We we’re gonna have to have an AI free episode. That’s a good point, Keith. Or is it gonna have to be like the AI free episode? I don’t know that we’ve had
one that doesn’t bring it up at some some point.
Keith Casebonne (35:03)
Yeah,
no, it’s it’s you’re right though. It’s like getting baked into everything. So
Mark Miller (35:06)
Yeah.
Keith Casebonne (35:08)
but yeah, it’s it’s yeah, it’s a worrisome thing. the I guess the the the new person in charge is sort of like a like an AI vibe coder kind of guy and thinks he can cut just like replace all the human work with with with with AI. And it’s yeah, I mean it it’s we’re gonna have government sites that unfortunately aren’t gonna be
nearly as as accessible as they once were or should
Mark Miller (35:33)
Hm.
Keith Casebonne (35:34)
be. And that’s gonna that’s gonna affect a lot of a lot of people.
Mark Miller (35:39)
It’s gonna be interesting to see if the government starts really violating the section 508 requirements based on
Keith Casebonne (35:46)
Yeah.
Mark Miller (35:48)
this kind of thing. And I think Keith that you brought up a really good point in that is the vibe coding, and the one thing that we know is that the percentage of websites that are inaccessible right now is increasing due to AI because AI doesn’t
is is building off of an inaccessible data set to begin with. So it’s a little it’s it’s just I just really struggle with
A decision that we use that technology to fix accessibility. It doesn’t compute.
Keith Casebonne (36:19)
Yeah. I think it was WebAIM’s
no. It was I think WebAIM’s annual, you know, survey of of the top whatever, like million
Mark Miller (36:27)
Mm-hmm.
Keith Casebonne (36:27)
websites or whatever. Every year it’s been getting a little better, a little better, a little better, and this is the first year that it’s starting to go down again and that’s what they attribute it to is the AI built websites.
Mark Miller (36:38)
Yep. Yeah. So this thing that everybody’s saying technology can do now, it’s actually going in the opposite direction.
Keith Casebonne (36:44)
Yep. Yep.
Mark Miller (36:46)
Justin,
Justin Stockton (36:46)
Yep.
Mark Miller (36:46)
you’ve been very quiet.
Justin Stockton (36:48)
Well, yeah. This is a I see a lot of the benefits in AI and I think I think probably the biggest I think the biggest mistake that that that they’re making in this decision, the the people behind that are now in charge of the US w USWDS, is that they’re they’re essentially throwing out the baby with the bathwater. It’s not a transition. It’s not a
Hey, we’re going to start investigating, you know, utilizing AI to do these types of things and everything. It’s a cold, hard stop line, you know, not even a line in the sand that you can must away, but we’re going to stop this and we’re going to start working towards bringing more AI in and doing these things. What happens in the interim? There’s no plan for that interim. and
And I think that’s the that that’s really the the the frustrating piece of this is that they’re going to throw away all the good work that’s been done and switch over to something that so everything’s gonna get a lot less accessible. So we’re gonna s it’s decisions like
this that we’re gonna see, especially across US based websites in the next WebAIM survey, we’re gonna start to see a tick up as more inaccessible inaccessible competitors.
components and things are being built into websites based on f a a now faulty library.
Mark Miller (38:18)
Yep. And and moreover,
Justin, I think that you’re really onto something here. Because one the reason cited is cost. It’s an it’s an efficiency argument. But what happens when you build that degree of technical debt? And then you have to turn around and make it accessible. The cost
Time loss, inefficiency, potential of that is staggering. It’s staggering. And this is where we see AI right now is building. So if you’re if you’re listening to this right now and you’re kind of wondering what we’re talking about, think about it this way. AI right now is out there increasing the technical debt associated with accessible websites. So unless your decision is we don’t need accessible websites, you’re
Building technical debt at an alarming rate because AI is doing it so fast. It’s
Keith Casebonne (39:23)
Yeah.
Mark Miller (39:23)
not out there fixing it, it’s out there adding to it. What’s the cost of that? I mean, it’s very much to me like the argument that we’re gonna push the Title II and HHS deadlines off by a year because companies aren’t ready for it, it’s gonna save them all this money. It costs people so much money to be halfway through making their stuff accessible in a week before they cross.
before we were crossing the goal line to have that held back. Right. So we gotta be careful of the the the reasons I think that we do this and the assumptions that AI is more efficient. And that is without considering the debt associated with environmental impact and all that other stuff that we’ve we’ve talked about.
Justin Stockton (40:08)
What happens when tokens start when those companies start actually charging you the right price for the tokens? So
Mark Miller (40:11)
Start charging.
Keith Casebonne (40:14)
Mm. Right, right. Yep. Right.
Mark Miller (40:14)
Yes, yes, yes. We’re not even thinking about that, right? That
right now people are like, like it’s the drug dealer. They’re like, you guys use this, you guys use this, and at some point it’s gonna switch. And usage that is is is we’re not paying for now, we’re gonna have to pay for. So what happens when that hits on top of everything else? So it’s we really do need to look at responsible usage of this, and we really do need to take a step back and recognize.
That AI is a tool. We do not need humans in the loop. It is AI, human first, AI in the loop. And we need people to learn how to use that tool responsibly.
I think we’re I think we’re in danger of making some assumptions here that can really have some incredible consequences. And I don’t have a crystal ball. I could totally be wrong. But based on the data that we’re seeing and based on what’s happened to this point, I would worry about it. And I would step into these things cautiously. And I’m not so sure. And I and I’m not I don’t even want to blame blame the current administration, the federal government, anything like that. I think that this is pervasive.
Keith Casebonne (41:21)
Mm-hmm,
Mark Miller (41:22)
across
Keith Casebonne (41:22)
mm-hmm.
Mark Miller (41:22)
industry, across government, across a lot of things that we’re not thinking about this quite as carefully as we should.
Keith Casebonne (41:30)
Yeah, agreed.
Mark Miller (41:31)
I didn’t mean to end this on like a dooms doomsday. It’s so cool. Did you guys know that you could
Keith Casebonne (41:36)
Well you sure did, didn’t you?
Justin Stockton (41:34)
But you did it anyway.
Mark Miller (41:39)
you could have AI write you a nursery rhyme in the same style as Snoop Dogg and read it to you? It’s really cool. You should try it. It’s fun. How’s that? That’s a better that’s a more. You
Keith Casebonne (41:51)
Okay. There you go.
Justin Stockton (41:52)
As Snoop Dogg or as just a nursery rhyme.
Mark Miller (41:54)
could say, write me a nursery rhyme in the style of Snoop Dog. And it would be like
A nursery but it would be like Snoop Dogg. Yeah, and anything that has Snoop Dogg associated with it is positive. He’s a positive human being. Has nothing but love for this world. You know what I mean?
Keith Casebonne (42:08)
He’s a positive guy. No, I agree.
Yep.
Mark Miller (42:15)
West Coast. All right.
Keith Casebonne (42:17)
Yeah.
Mark Miller (42:18)
Any last thoughts from you all? This is an interesting one and and I want to thank you, Keith, for being on it. Right. Like I just I’m thankful that we’re we’re friends, that we’re professionals in a in the same industry and that we’ve had such a long run together and that that means we get to do these things over and over again. So thanks for being on. Thanks for having me on.
Keith Casebonne (42:40)
yeah?
Mark Miller (42:40)
And I hope you I hope you enjoyed it. What’d you think? Yeah.
Keith Casebonne (42:43)
I did. It was great.
This was this was wonderful. I enjoyed this conversation a lot. and same back. You know, it’s been it’s been been, you know, great b again working with you, being friends, and and and kind of sharing our podcast experiences.
Mark Miller (42:56)
Mm-hmm.
Keith Casebonne (42:57)
and hopefully we can do this more before all the AI bots just take over and do all the podcasting for us. Exactly.
Mark Miller (43:02)
Do the podcast for us. I’ve already got an AI agent that
I’ve built that’s just gonna go podcast for me. So
Justin Stockton (43:08)
Ha ha ha.
Mark Miller (43:10)
this is what we’re gonna be left with, man, right? Like we gotta do this kind of stuff. This is it’s it’s it’s the and
Keith Casebonne (43:09)
Good to know. All right. be on the lookout. Yeah, yeah.
Mark Miller (43:16)
again going back to Steve at C Vent, like one of the things that that C Vent, which is they’re a they’re a technology for conferences, they’re really leaning into the conferences and being like, Hey, one of the shifts with
AI is gonna be gonna cut sort of create this need for us to connect. And things like conferences, and this goes back to our discussion with the Innovation Lab and C Sun and all these things, they’re gonna become more and more important, Keith. So this is it. This is we’re doing, we’re
Keith Casebonne (43:45)
Yep, yep, yep.
Mark Miller (43:46)
doing the human thing right now. You
Keith Casebonne (43:48)
Yeah.
Mark Miller (43:48)
and I have been doing it for years, and we’re still doing it. We’re still doing it.
Keith Casebonne (43:51)
Still doing it. Yep.
Mark Miller (43:54)
Okay.
Justin Stockton (43:54)
We’ll see you back
for episode one hundred thirteen.
Mark Miller (43:57)
What’s that?
Justin Stockton (43:58)
So we’ll We’ll see him back for episode one hundred thirteen. Yeah.
Mark Miller (44:01)
One thirteen. Yeah,
Keith Casebonne (44:03)
That’s right.
Mark Miller (44:03)
we should do that.
Keith Casebonne (44:03)
That’s right.
Mark Miller (44:04)
Whatever episode number I was on in the hundreds f of your podcast, Keith, we should have you make sure you’re on that episode for our podcast. Yeah. Write that down. Somebody write that down. Where’s an AI agent when you need it?
Keith Casebonne (44:12)
Okay, yeah. I’ll mark my calendar. Exactly. I’ll tell you right now. Episode one seven.
You’re on episode one seven. I just I happen to have the folder open. I can Yeah, I just have one seven. okay. Well there you go. cool. Nice. Nice. Nice. All right. Well, I’ll be there. All right. Sounds good.
Mark Miller (44:23)
Really, you know that? 107. That was a radio station here in town in Exeter, New Hampshire, to be WERZ 107.1 FM. Yeah, pop, pop station. Yep. All right. 107. We’ll see you at we’ll see you on 107, Keith. Yeah. Or your AI agent,
whatever’s whatever’s going on at that time. Yep. That
Keith Casebonne (44:40)
Yeah, exactly. I’ll send my AI proxy.
Mark Miller (44:42)
your AI proxy. All right.
Justin Stockton (44:43)
Right. That’s
gonna be somewhere around the june fourteenth, twenty twenty eight.
Keith Casebonne (44:48)
Cool. Cool. Nice. All right. I will mark my calendar.
Mark Miller (44:48)
You gotta figure it out. You guys are so good. You guys are so good at math.
Justin Stockton (44:50)
Eighty s eighty seven weeks.
Mark Miller (44:54)
All right. Anything any last thoughts?
Keith Casebonne (44:56)
No, again, thanks for having me.
Mark Miller (44:59)
Beautiful. Thanks for being here. Justin, thanks for doing with us me every every every episode. Thank you.
Justin Stockton (45:06)
Thank you, Mark.
Mark Miller (45:07)
Okay. We’ve all thanked each other. Now that we have broken down accessibility for you, we hope you go forth and keep it accessible.
About the authors
Mark Miller
Founder & CEO at Inclusion Impact Accessibility
Mark SMiller is a digital accessibility consultant who helps organizations build accessible, inclusive digital experiences. He speaks regularly at industry conferences,, CSUN, CVENt and apperes on many webinars and podace. , He trains teams on practical, sustainable accessibility workflows.
Justin Stockton
Principal Accessibility Architect at Inclusion Impact Accessibility
Justin has 25+ years of accessibility experience spanning software development and product management. His career began working on federal accessibility initiatives, including Disability.gov and the U.S. Access Board. He has led accessibility audits for global enterprises, architected innovative audit tools, and managed enterprise accessibility platforms serving 1000s of users.
Comments