How is AI actually changing the way people use assistive technology, and what happens when AI lets developers move faster than they can realistically review their own work?
In this episode of Accessibility Craft, Chris is joined by DevOps engineer and accessibility consultant Alex Stine, along with returning guest and Equalize Digital team member William, for a wide-ranging conversation about AI, accessibility, screen readers, coding agents, and the growing tension between speed and reliability. Alex shares firsthand examples of using AI to make inaccessible PDFs, charts, weather data, and other visual information more usable in his own day-to-day, while the group digs into the accessibility problems that still exist in many AI interfaces.
They also compare how they use ChatGPT, Claude, Gemini, GitHub Copilot, and other AI tools professionally and personally, including AI-assisted coding, research, automation, and building custom software. They don’t shy away from the messy bits: excessive context switching, enormous AI-generated pull requests, production outages, unreliable output, and why “the AI did it” is not a very convincing excuse.
For the beverage tasting, Chris and Alex try Fentimans Victorian Lemonade, while William attempts a homemade substitute. Reviews are… mixed.
Episode Outline
- How Alex, William, and Chris are using AI at work and at home
- AI as an accessibility ‘equalizer’ for inaccessible PDFs, charts, and visual information
- The current screen reader experience in ChatGPT, Gemini, GitHub Copilot, and terminal-based AI tools and why browser-based AI interfaces may still offer the best experience
- AI productivity, context switching, information overload, and the temptation to run too many agents at once
- Accessibility, stability, and security risks when developers move too quickly, why AI-generated code still requires human validation
- The importance of taking responsibility and the shifting sands of accountability in an AI-forward workplace
Links & Resources Mentioned
Download Accessibility Checker and use coupon code AccessibilityCraft to save 10% on any plan.
Subscribe for more practical conversations about digital accessibility, WordPress, emerging legal developments, and beverages that inspire unusually serious debate.
Accessibility Craft is hosted by Amber Hinds, Chris Hinds, and Steve Jones. They are experts in digital accessibility and creators of software, courses, and specialized services that have made millions of websites more accessible through their work.
To learn more about us, you can visit our website.
Listen
Watch
Transcript
Chris Hinds: Welcome to Accessibility Craft, where we explore the complex challenges and emerging trends that are shaping digital accessibility, while sipping on unique craft beverages. This show is proudly produced by Equalize Digital, The most trusted name in WordPress accessibility. Join us every week as we break down accessibility news and share the expert strategies we’ve used to help make millions of websites more accessible.
Grab a drink, the show starts now!
Chris: Hello. Hello everybody and welcome to the
Accessibility Craft Podcast. My name is Chris and I am here today with William.
William: Hey.
Welcoming Special Guest Alex Stine, and Also William!
Chris: And we have a special guest today as well which is Alex. Alex, would you like to introduce yourself for the listeners and the viewers?
Alex: So my name is Alex Stine. I am a DevOps Engineer and accessibility consultant based here in Dallas, Texas.
Chris: Thank you so much for being here, Alex. And I probably should have called out William as a special guest too, because he’s on the live streams and in other areas, but he hasn’t been on the podcast in some time. So welcome again to you.
William: I, yeah, I don’t remember the last time I was here, but last year maybe.
Chris: Well it’s been a minute. It’s been a minute. Well,
for anyone who would like to grab the show notes or a full transcript of this episode, you can do that by going to accessibilitycraft.com/176. And we’re trying this new thing where we ask people to, whether they’re listening on their podcast app or they’re watching this on YouTube hit that like button, hit that subscribe button if you see one. Post a review, post a comment, let us know how we’re doing. Let us know if you like it. Let us know if you don’t like it. We just wanna see you engage. We wanna see your comments. We wanna see your questions. And thank you so much in advance for doing that. And I’m going to now introduce the beverage.
Today’s Beverage
Chris: So today’s beverage it’s kinda funny because the beverage we’re having today got its start in the UK, and originally it was gonna be a bunch of US people on here having this, and then due to schedules and vacations and stuff, now William, who’s based out of Scotland area, is here, and we’re having Fentimans Sparkling Victoria Lemonade, which I think has been around since, like, the 1800s. But much as I was able to get this, and we got it to Alex, but William, you couldn’t find it so you had to improvise.
William: I couldn’t find it. Yep, I had to make my own lemonade, and I won’t be doing that again.
Chris: Did you have an explosion when you added the bubbles to it?
William: No.
Chris: That happened to me the last time.
William: I fizzed the water first and then I put the lemon…
Chris: oh.
William: … in after.
Chris: Yeah. We’re excited to try this, and it’s it’s non-alcoholic, botanically brewed. I think it’s, I think it’s naturally fizzy, so it’s like yeast or something, some kind of fermentation that makes it fizzy, and they don’t actually add bubbles to it, through some kind of industrial process. So we can crack these open. We can give these a taste. And William, y- you can give us your own impressions of your own homemade lemonade.
William: I might skip it for a bit. It was a bit too sour.
Chris: It is a bit too sour? I’m gonna give this a try, and Alex, feel free I mean, this is pretty sour too, I’m not gonna lie. But it does…
Alex: Well…
Chris: … have a hint of sweetness
Alex: That– there’s no hint of sweetness in that. That is sour as can be.
Chris: Oh. I think they call that European sweet. They’re, they tend to be lower on the sugar spectrum depending on where you are. William’s shaking his head no. Is that not true?
William: We pretend, but we like our sugar.
Chris: Yeah. That’s pretty good. I haven’t had one of these in, gosh, probably 10, 12 years. Back when I was in restaurants, we used to sometimes do mixed drinks with these, and they have another one that’s called, like, Dandelion & Burdock or something like that. It’s got, like, a, a very funky flavor to it.
But we used to do mixed drinks with them.
William: Nobody likes Dandelion and Burdock. Why would they make that?
Chris: I don’t know.
Alex: I have no clue.
Chris: So, Alex, I think the last time we had you on here, we had, like, some kind of really sweet grape soda. So we’re skewing more into sour territory here, which I’m gonna use that as an opportunity for a segue here in a minute. But our super sour, slightly bubbly lemonade, I’m curious what your impressions are.
Thumbs up, thumbs in the middle, thumbs down? What do you think?
Alex: Oh definitely a thumbs down. I would never order this. Like, never.
Chris: Never. maybe you can I don’t know, use rest of it to strip some stuff off of some dirty dishes or something. I don’t know.
Alex: May- maybe we’ll turn it into a mixed drink. Who knows?
Chris: Yeah. There you go. There you go. And William, you’re holding off on yours. It’s…
William: I have the same problem. Lemonade is definitely not for us.
Chris: Mm-hmm. Okay. Well, I’m I’m a thumbs up on this, but I like drinks that are less sweet. And I appreciate the sour note, but can see I’m outvoted.
How We’re Actually Using AI Personally and Professionally
Chris: But I am curious, we’re gonna talk about… We’re gonna talk a lot about AI, screen readers, and the real user experience around AI and everything in between those things.
We’re gonna kinda see here in this conversation that’s about to take place, whether the three of us are sweet on AI, or if we’re sour on it, not unlike the lemonades we just experienced.
And so with that said, I think I’d like to do is kind of just go around the table as if we were sitting at a table and just talk about how you’re using AI right now, what model you use both at work and in your personal life. And, Alex, if you don’t mind starting, and then we can go William, and then I’ll go last.
Alex: So I use AI all the time. I mean, why write a script anymore at work when AI can just do it? I’m really good at reading code, so it’s just saves me a lot of time, and then I can start asking it questions like, “Why would you do it this way? Like, it’s obviously better to do it that way.” And then it can say, “Well, no, based on these standards, we should do it this way.”
It’s like having your very own little development partner. It’s pretty insane. And the feedback is just so quick. It’s amazing. So that’s my professional use of it. And for personal use, I’ve done everything from feed inaccessible documents into it to even having it just read inaccessible charts and graphs for like weather forecasts.
So very wide range.
Chris: And William?
William (2): Yeah. So at work I use it much the same way Alex does. I use it to speed up the writing of the code. I’ve written hundreds of thousands of lines of code over time, and don’t really enjoy the typing of the code. I like the solving of the problem, and AI really helps me, nail down getting to the root of the problem quickly. And I do quite often still hand write the business logic, if you will, of solving that problem. But I rarely write the boilerplate code anymore. And at work, mostly using Claude either Sonnet or occasionally in Opus. O-outside of work, I use it for, the weather, keeping notes for me, and also to help me e-edit blog posts that I write.
I write a lot of blog posts. I rarely publish them because the editing process is undesirable for me, so I tend to just leave them in draft forever. Now the AI helps me get them to the point where I’m happy to click the publish button.
Chris: That’s amazing. And thank you for dropping some names of some models. So at… That actually reminded me that was one of the components. Alex, I wanna circle back to you. Are there any, names of AI models that you’d like to drop that you use or that you recommend?
Alex: My company would hate me for saying it, but I use Opus all the time, especially in advanced systems works. It is super expensive, but GPT just can’t compete.
Chris: Mm-hmm.
Alex: Not in that type of advanced logic where you need not only the ability to gather context, but the ability to, you know– When I say cite sources, that’s what I wanna see. Like I want you to prove to me that works on this version of software and that it’s well tested.
Chris: Yeah. Well, compared to these two, it’ll surprise probably no one listening that I’m… I skew a bit more towards vanilla in my use of AI. But I am much further along this year than I was in prior years. So up until the end of 2025, I was definitely still only engaging with it through a chat interface and having it do what it could do for me within the context of a chat window.
But I… The reality was I had barely scratched the surface, which is what I learned this year in 2026. This year, I’ve shifted from Gemini in 2025 to using ChatGPT and, well, Codex, but now I think it’s just called ChatGPT on macOS running that. And I am now using GPT 5.4 and 5.5.
Having it build and execute skills for me for any kind of repetitive process, whether that’s on our ops side or on sales or things of that nature. And it’s taken, multiple tasks that would take between 45 minutes and 90 minutes and cut them down to a five-minute or less prompt and review. So it’s saved me, I’m just gonna say it, a shitload of time. It saved me so much time this year using it in this way that I have had more time for the more important parts of my job, at least in my opinion, which is having the sales conversations and being out there more publicly versus down in the trenches doing kinda in the weeds ops kind of work.
On the personal side probably the craziest thing I’ve done is I used ChatGPT 5.5 with the GSD framework to build a personal progressive web app for myself that’s on my phone, and it’s it’s like a, a food logging and weight tracking app. ‘Cause I don’t wanna pay for one, but I wanna be able to do that and journal and log food and weight.
So I built one, and it’s just mine, and it’s just my data, and no other company has it. And that was kind of gratifying to do over a weekend. So, that’s just one example of what I’ve done on the personal side for it.
William (2): But there’s a really good use of AI to build a thing that is a commercially available offering, but gives either too much of your data away or has too many features you’re never gonna use. You can build your own exactly as you need.
Chris: But it’s kind of addictive in a way ’cause, like, I’m using it now and I’m like, “Ooh, I wanna do this change, and I wanna make this improvement, and I want this to work like this now.” And, it’s like, it… You burn tokens real fast if you let yourself. I’m resisting because technically those are work tokens, so I shouldn’t use them for personal reasons unless I literally have no other use for them.
William (2): Tokens have to be burned.
Alex: Many tokens have to be burned. I mean, I had a team member, I’ve never done it, but they managed to burn so many tokens that we hit a license limit, and I’m like, “That is impressive at enterprise scale. Like, good job.”
Chris: Yeah. Yeah. Well, I’ve seen memes and articles of companies that have run, like, contests for which engineers can use the most AI tokens and, incentives around that, which I’m like, “This is kinda weird,” but maybe there is a reason to do it that way. I don’t know. ‘Cause they’re worried they’re wasting them if they’re not using them, I guess.
Is AI Changing Things for Assistive Technology Users?
Chris: So my, my next question is for you, Alex, and I’m curious, in that intersection of you’re a site reliability engineer, you’ve been doing that for years, you’re super talented but you also use assistive technology. How has AI changed the way that you interact with technology over the past, say, two to three years?
Alex: Oh, incredibly. But admittedly, I didn’t get involved in it three years ago. I am– I’ve always been the one that’s been a late adopter of technology ’cause I, I have a very small amount of time that I like to put towards new things, and if others put forward that time first, I don’t have to waste mine.
Chris: Mm-hmm.
Alex: So very recently, I started really getting into AI probably last year, and it is just incredible what it’s been able to do for my life.
Chris: Give like two or three examples if you don’t mind of some things, if you don’t mind sharing.
Alex: So one of the– probably one of the most impressive accomplishments was the ability to use AI to start tracking barometric pressure trends. I have been having some fairly severe headaches in the last couple of years, and the ability to be able to model the data from these charts and actually track the trends when none of this information is accessible.
None of it. It’s all just one big visual chart, and AI is a great equalizer because it can create… It can take visual data and create very accessible rep-representations.
Chris: Very cool.
Alex: That’s probably my biggest, but if we had to go for a close second, PDFs. PDFs are still a really big problem.
Most are not accessible. I will stand behind that statement. And you can very simply feed a PDF through an AI chatbot. I use ChatGPT a lot personally, and it can give you a fully accessible Word document. It can just output the text. I mean, it’s just– it saved me so much time. And it’s not always accurate.
I mean, it’s always trust but verify, but, that’s… There are services out there that do that, and they cost a lot of money. And to be able to have, just for less important stuff done for basically free, why not?
Chris: Yeah, totally. And I can see that being a, a great use of it just on a personal level. I wanna get into more of just where AI has made life easier, but maybe also where it’s created some frustrations. But before we do that, we’re gonna take a quick commercial break.
Brought to you by Accessibility Checker
Steve Jones: This episode of Accessibility Craft is sponsored by Equalize Digital Accessibility Checker, the WordPress plugin that helps you find accessibility problems before you hit publish. Thousands of businesses, nonprofits, universities, and government agencies around the world trust Accessibility Checker to help their teams find, fix, and prevent accessibility problems on an ongoing basis.
New to accessibility? Equalize Digital Accessibility Checker is here to teach you every step of the way, whether you’re a content creator or a developer, our detailed documentation guides you through fixing accessibility issues. Never lose track of accessibility again with real time scans each time you save, powerful reports inside the WordPress dashboard, and a front end view to help you track down hard to find issues.
Scan unlimited posts and pages with Accessibility Checker Free. Upgrade to Accessibility Checker Pro to scan your website in bulk, whether it has 10 pages or 10,000. Download Accessibility Checker today at EqualizeDigital.com/Accessibility-Checker. Use coupon code AccessibilityCraft to save 10% on any plan.
AI Interfaces and Assistive Technology, All’s Not Well
Chris: And we’re back. So we already got a couple of good examples around barometric pressure, around getting information out of dense and inaccessible PDF files that you need, where AI’s made life easier. We talked about how we’re using it on the professional side and on the personal side. I’m wondering if we can get into where maybe it’s created some frustrations or some challenges, and I don’t… anyone can chime in.
Alex: Chatbot interfaces are mixed with issues. Many accessibility issues.
William (2): Actually would like to ask, so I think both Codex and Claude have recently im-apparently improved their screen reader capabilities in terminal code agents. First off, is that true, and is it really better? They announced it in their changelogs, but it’s questionable whether it is an improvement.
Alex: I don’t currently have access to Claude in the terminal, so I really couldn’t say. I don’t… We use GitHub Copilot at work. It has some very bad accessibility issues in the terminal. It’s very hard to navigate. It displays it in a visual format where you have to be able to essentially scroll up and down.
And yes, those commands are accessible on the keyboard, but then trying to have to navigate up and down through that output, it’s a miserable experience. And I know these companies can definitely do a lot better than that. Is that something that’s fixed in Claude? I can’t speak to that. I don’t use it.
But if the updates, like you say, in the changelog are what they are, then they probably do offer a better experience at this point.
William (2): Haven’t tested it myself, so couldn’t comment on the quality of those improvements, but that, that might be something to consider checking out for someone. Might be useful for you to be able to, actually be told when the messages are streaming in as opposed to what I assume Codex or I assume Copilot– I mean, I use Copilot and I can see the messages stream through. They– if they don’t get announced, there is basically no way you can go back through all that.
Chris: Yeah. Well, and some of what it’s streaming in, it’s like stream of consciousness as it’s working. I mean, some of it’s just gobbledygook, so I don’t even know if you would want it spouting all that off in your ear as it’s trying to work or if you’d want the option, right? That’s an interesting conundrum. Alex, are you… Like, when you’re using it on mobile, is the experience okay in app form, or is it kinda the same situation where it’s a lot of scrolling and a lot of lack of context?
Alex: I have used Google Gemini on mobile, and just– it’s really a nightmare to navigate. And I tried ChatGPT when I was still on my Android phone, and of course, Android has the best support with Gemini. You can use it right through your RCS messaging. It’s great. Couldn’t recommend that any more. It’s probably the most accessible that I know about.
Chris: Oh, OK.
Alex: Ch-ChatGPT on Android, no. It had worse problems. Focus jumped all over the place. It just… It wasn’t usable
Chris: So when you’re using, when you’re using this tech, are you primarily using it through your terminal on your computer, or where are you most often using it? Or are you using it in like a browser interface?
Alex: I actually find the browser to be the most accessible currently.
Chris: Okay. Okay. Thank you for satisfying my curiosity there. So inconsistent UX for screen reader users that’s clear enough. Other frustrations. I’ve got one if nothing’s leaping to mind, and maybe that’ll inspire something. Okay, I’ll go.
Context Switching, Overwhelm, and Consequences of Letting AI Run Wild
Chris: So one, one I posted about not too long ago on LinkedIn, and this is more just– isn’t necessarily accessibility related, but maybe more like overwhelm and anxiety related, is found as it has allowed me to move faster I am context switching more and having to absorb information and react to it at a greater cadence than I used to.
So I’m getting more done, but I feel like I’m– my brain is like having to work even harder to keep up with the AI agents running through because it can parse information so quickly, and then it needs feedback from me, especially if I’m running it in multiple different threads. Curious if that’s been other people’s experience or maybe I’m just doing it wrong a-and that’s why I’m having the experience I’m having.
William (2): Definitely having an e- a very similar experience. So I use AI almost exclusively in the terminal to write code or to parse numbers, and whenever one of them is working, the urge to open a second one is always there. And by the end of the day, that first one becomes the second.
Maybe there’s four open, going back and forth between them, and sometimes I just put them onto auto mode and let them do what they want because the– build up so much context shift in one moment that you just have to choose to ignore some of the, the noise.
And I feel that exact same problem. Any wait time where one of them is doing something else is– brings the urge for me to just start another chat.
Alex: I used to have this problem, and then I just quit moving so fast. I unfortunately, in the reliability world, have seen AI do some pretty terrible things. And with that being said, sometimes faster is not always better. Sometimes you need to slow down and re– resist that urge to open four different terminal tabs or four different browser tabs.
Like, it will finish eventually, and when that happens, it’s fine. But I do notice the same pattern, that you start to become more naturally impatient, and that’s not a good thing for me because I’m not really patient as it is.
Chris: Yeah. Yeah, totally
William (2): Distractions are a problem for me while I wait. If it’s taken longer than a few minutes, I might open up X or make a sandwich. I eat a lot of sandwiches. That’s how it…
Alex: You know what? I second that. I’ll be waiting on a response, and I’ll just have to go do something else. It sometimes… It’s gotten so bad now where I won’t even do that. I’ll just stand up and go walk away, and it’s like, eh, I might as well get some exercise out of it.
William (2): I had to implement a rule where I can only have the sandwich if I hit the five-hour window.
Alex: Fair enough.
Chris: I mean, yeah, and now there’s not even five-hour windows. It’s like a weekly limit, so is that like a, is that like a diet?
William (2): That, that…
Chris: One sandwich a week?
William (2): ( Cross talk )
Chris: Oh, man, that’s funny. That’s too funny. Yeah, I mean, the other thing that, that comes up for me occasionally, and I don’t know if I should admit this publicly or, but I’m going to anyway, is like I’ll come across something that I need to do that’s just so lengthy with so many steps and so overwhelmingly complex that I have found more and more this temptation to have AI, versus me put in the thought and the time to do it right.
And so far I’ve been able to stave off that urge, or only give the parts that are more rote and predictable to an agent to assist me and not give it the whole thing. But I feel like this is just gonna become a bigger and bigger problem where eventually the laziness might win as, as the AI gets more and more competent.
But like right now, one of my cardinal rules is any output has to have a human review step. And as the outputs increase I find myself more and more wanting to just trust that it’s gonna do it right. And I have so far managed to not give in to that temptation, but I’m curious if you two have run into that.
Alex: Yes, especially when you watch Claude or any of these other AIs write thousands of lines of code, and then you open a pull request and it’s like, “Oh I’ll do a quick review on that. Oh, did I mention it was thirty thousand lines of code divs?” Great. Great. Yeah, I’ll get right around to that.
Chris: Yeah. Yeah. No, thank you.
William (2): Yeah.
Chris: What about you, William?
William (2): Similar problem to that where it is tempting to let it just do all of the things all at once, and it, in the code world, it does work out exactly like Alex said, where before you realize it, it’s changed 3,000 lines and found 6,000 other bugs that it might wanna look into and solve at the same time, and your PRs end up being unreviewable.
The, the only way to review PRs of the size of AI is with AI, and then you’re placing the trust in both the AI to produce the code, to verify the code, and sometimes you’re using the same models and the same agents to do that. And a human, we have a bias, but I also think the AI has a bias for its own style as well. So if you have one agent write the code and the same agent review the code, it might not find things that fresh eyes would. It’s easy and,
Chris: Mm-hmm.
William (2): …you know, when you’re talking about two developers, two humans, the eyes are always fresh. It’s very difficult to know whether fresh eyes come from having an AI review pass as well.
Closing Thoughts and Advice for Heavy AI Users, and the AI Curious
Chris: Yeah. Well, one thing that got mentioned earlier on was PDFs a- and AI’s ability to compensate for bad PDFs. Are there other areas in the accessibility space, Alex, where you’ve seen AI as a compensator or as an equalizer beyond what we’ve already talked about?
Alex: Because AIs have access to so much data, I have literally planned out and mapped walking routes to places, and it can even describe what I might pass on the way. It’s actually pretty crazy, and the fact that it’s more right than it is wrong is amazing. And y-yeah, of course, it is wrong quite a bit.
Like, it’s given me information that’s just plainly wrong. But the more times that it’s right is really wild to me. Like, I was doing this on a place I already knew how to walk to, and I’m like, “I need directions to the Starbucks,” and it said, “Okay, well, when you make a right, there’s gonna be a giant glass building on your left.”
And I’m sitting here thinking, “Wow, it’s not wrong.”
Chris: Yeah. Interesting. Interesting. And yeah, the more it gets things right, the more we all trust it, right? And that’s what’s been interesting about this for me from the beginning, because I remember what the outputs used to be like when I first like fired it up and was playing around with it personally. And man then versus now, is like night and day in so many ways. And to that end I, I’d be curious for both the developers in the room, AI’s ability to work on code. Are there areas of development where you feel it is most valuable or has the capacity to contribute the most value per unit of input? Is it on the, the review side? Is it on the patching side and fixing specific issues? Or is it building a concept from nothing? Like where do you think it, it is best as a tool, or is it all of those and I’m just, needlessly trying to put it in buckets?
William (2): I actually think it is a bit of all of the above, but it is not an expert in any of them other than perhaps it, it is the fastest typist anyone ever be, it does still require a fair amount of guidance, both at the input side and also verification at the output side.
And the time you save in having the AI do some of the typing for you, get to spend later on in making sure what is done is correct. So I don’t know whether it really saves time or it just shifts where the time gets spent, you can spend the time on the more valuable parts. But I think it is very good at almost everything, but it isn’t master of any of those things.
Alex: Yeah, I would agree with that. I don’t let AI review stuff. I don’t trust it. Like, if an AI writes it, I don’t need an AI to review it. I’ll do that part. And the other area where it really saves time for me is in research, especially when some of the documentation sites only keep, say, the last five major versions.
But the references are still spread all over GitHub, and AI is great about tracking that down, so I can at least validate the work versus just accepting a blind guess.
Chris: Yeah. And to that end of producing code with AI, I’d be curious what your perspective is, both of you, on where you see AI introducing the most barriers or just general quality issues when it’s producing if there’s areas where it seems to be consistently weak. I’ll lead with what I found interesting was I had, I’ve had two different experiences with building, like, personal software with the GSD framework and with I used Gemini once and I used Codex and ChatGPT the, the second time.
In the first round of it it could not build something that worked. And granted, I was not trying to build the same thing both times, so it wasn’t exactly an apples to apples comparison. And then this, this most recent time building my little progressive web app that I have on my phone that tracks data and saves it and can import it and export it and, shows and compares data in different ways and saves it to refer back to later. It’s pretty sophisticated for a personal app. It does quite a bit.
It built it successfully over a few iterations in one go over a two-day period, and one of the things that I thought was really striking is, without me prompting it at all it went ahead and assumed that I wanted everything to meet accessibility standards.
Now, I don’t know if it, like, surreptitiously went and looked me up, who I was or something without my knowing, ’cause it did have, access to do research. But it did… It was like, it’s gonna… We wanna, have proper color contrast. We want this to meet WCAG. I think it said 2.1 AA.
A- and it was gonna try to code to that standard, and it was without me even saying it, and I even gave it a little attaboy. I was like, “All right. Cool. Yes, please do that. I don’t need assistive tech right now on my phone, but it’s probably good to just go ahead and do it in case I use this for, a really long time or something happens.”
So that’s been my experience. I’ve noticed that it went from not being able to build something that worked at all and not ever mentioning accessibility unless I brought it up to it kind of, brought it up even though I didn’t, and it managed to build something that worked, and that was a progression over about an eight-month period. Curious back to my original question, if you two are seeing kind of barriers or reliability concerns, i- in, in your current use of AI where it’s consistently underperforming?
Alex: I mean, yeah, for sure. Like, I’ve seen AI cause real outages in production systems. Like, oh, well, I planned it to work that way, but when I deployed it and didn’t look at it actually, we caused several SEP 1s in one day. Don’t do that. That’s my best advice. Don’t blindly trust AI. No matter how good you think it is, it’s not actually that good.
You still have to validate it, and I tell people that I work with all the time, just because it uses– just because you’re using an agent that acts with your identity does not mean you can say, “The agent did it.” It’s you at the end of the day, and that’s what drives me absolutely crazy about where the industry’s going right now.
Oh, the AI did it. It’s like saying the dog ate my homework.
Chris: It’s actually a really good point. I want to circle back to that point. But William, do you have do you have anything to add on, where you see it introducing issues?
William (2): I mean, m-moving quickly inherently does come with problems. As a web developer, you can very quickly have AI, build you the new component, create a table, build an entire WordPress theme, scaffold a plugin. You can have it do all of these things, but it can’t– it doesn’t really check what it does when you move that quickly.
It outputs exactly as you ask, and it generally doesn’t have follow-up questions. So unless you’re keeping an eye on it or being verbose with its prompt it almost never is gonna come out with something perfect the first time. We hear about, like, loop engineers, graph engineers right now.
I think this is the reason why first iteration almost never is gonna work. You do have to cycle back on it and make sure that it did what you asked. You also have to verify that what you asked it to do is also correct because misunderstandings happen very quickly as the AI starts to compact its context.
Chris: Yeah. Yeah, no kidding. And I do want to circle back to that whole blame the AI thing, ’cause I think it is interesting, right? And that’s maybe a uniquely human thing is we we can, if we can shift blame off of ourselves for something we’re going to do it. But to Alex’s point, who wrote the prompt? Who decided to trust it?
And I think that’s a really fair point, and I love the dog ate my homework comparison, and I’m even guilty of doing this, I have AI agents and assistants that do things for me operationally inside of our company’s project management systems and inside our email inbox, and it, if it makes a minor error I think before I’ve even really processed the thought, the words out of my mouth are in fact, “Oh, the AI must have messed that up.” When really it’s me, and probably what it means is I need a more careful human review step on my own work, because it is my work.
Alex: Like, AI is a powerful technology. We’ve got lots of dangerous technology. Like, I’m not beating up on AI just to beat up on AI. There’s plenty of ways you can mess something up without AI. But AI has allowed us to move much faster with the existing technology that we do have.
So from a reliability perspective, test it in dev, test it in staging, write the code, and then look at it. Just because you don’t tell your AI you want secure code doesn’t mean you want code full of SQL injection vulnerabilities either. So, you have to create smart prompts, you have to re-prompt, and you have to actually check what it spits out.
And if you don’t do those three things it’s on you at the end of the day.
William (2): Yeah, I will just second that exact thing. I see on the social media platforms, people always say things like, “I one-shot this application,” or, “Look what I managed with this one-shot system.” But I always question that because any time I’ve tried to one-shot anything, it has came out as a complete mess. I don’t think you really can one-shot and trust any piece of software that the AI is gonna produce for you. Always need to make sure that it, does what you’re asking, it, in my experience, it quite often forgets what I asked if the chat runs, several hours or across the space of an entire workday. Isn’t gonna remember what… Humans don’t remember what they had for breakfast. The AI doesn’t ever remember what its first prompt is.
Chris: Well, and those people making those one-shot claims are probably selling a course or something, right? Nine times out of 10. Well, cool. Well, thank you both so much for being here for the impromptu conversation on AI, tech, and assistive technology and all of the above, and how we’re using it.
I think it was really fun to talk through this with you two. And we will see everybody here next time on Accessibility Craft. Thank you so much.
Alex: Thank you.
William (2): Bye.
Thanks for listening to Accessibility Craft. If you found this episode valuable, please help us reach more people by subscribing, reviewing, or liking the show, and sharing this with your colleagues. Accessibility Craft is a production of Equalize Digital Inc. Steve Jones composed our theme music. To learn how Equalize Digital can support you on your accessibility journey, visit us at EqualizeDigital.com.