AI and its impact on digital accessibility

Oliver:
Hello everyone, and welcome to today’s talk on AI and its impact on digital accessibility.

I am your host Oliver Emberton, and my main claim to fame and reason for speaking with you today is because I have two decades of deep commercial experience working with both artificial intelligence and web accessibility, which is an unusual mix. I even did my degree in AI back in 1997, if you can believe that anyone can be that old.

My stated mission right now is to make web accessibility as commonplace as supporting mobile, and as CEO of Silktide, I’m generally the product visionary futurist person who unfortunately does tend to talk a lot, so strap in, because I’m gonna be talking a lot. Right, now. For those of you who have been outdoors or who’ve been connected to the internet at any point in the last two years, you may well be asking the question, what the heck is happening, specifically with AI? Now we only have

about half an hour today and there will be questions and answers at the end. So one thing is to note, if you do have a question, please leave it in chat. We’ll get to those for the last five minutes or so. But I wanna take five minutes and just, high level, for anyone who’s not deeply embedded in the AI universe, what on earth is going on and why does any of that apply to accessibility?

So we’re gonna do this in four bullet points. Four bullet points designed to impress your friends and family at cocktail parties. Let’s get started. So the first one you are going to probably be familiar with, which is large language models. This was a technology that came to the forefront in November, 2022 with the launch of ChatGPT, and the key takeaway from a large language model, what you need to know about it, is firstly it is

the first technology we ever had that could understand natural language. This was a revolution in computing. Until that point, I said I worked on AI for like 20 odd years, and I remember us dreaming of this moment, but it seemed impossible. And then suddenly we had the ability for computers to essentially understand text as it is written,

without miracle solutions or guesswork, essentially. Now those LLMs have improved dramatically since, but the important thing to take away is think of them as chat bots, I’m sure you’ve all used ChatGPT or DeepSeek, and they have the ability to understand natural language. Now, after that, we had the innovation of what is known as multimodality, which is basically the ability to transcend text and to start looking at

content such as images or video or speech. You don’t need to understand the mechanics. All you need to understand is roughly a year later, we more or less started to conquer the frontier of being able to understand those things with a similar level of efficacy to what we had with LLMs and plain text.

The implications for accessibility you’ll see in a moment, but they’re pretty profound. Next up we had something called chain of thought. Now to explain what chain of thought is I need to give you, oddly enough, a thought experiment. So imagine that you are given a math problem, right? I’m sure you’ve all had a math exam and, alright, I like math exams ’cause I’m weird.

But for a normal person, if you get a math question off the top of your head, you are forced to answer it off the top of your head. You’re gonna struggle, right? If I just said to you, and now I’ve read out a formula, solve it, you probably wouldn’t do too well. But if I gave you a pen and paper, you have a pretty good chance, hopefully of actually solving.

Why is that? Well, the ability to reflect on your own knowledge, to reflect on your own thinking, to consider, you know, things step by step, that is a tremendous uplift to your effective intelligence and it’s exactly the same for AI and that, essentially, I’m simplifying, but that’s essentially chain of thought.

It was an innovation we had around September, 2024 with the 0.1 models and so forth. What that gives us is it gives us AI that can reason and reason at more or less a PhD level across a wide range of fields. And our last innovation, which is our most recent, is approximately December, 2024 is agentic, and agentic is essentially the ability for an AI to pursue a goal independently.

As an example, you might say, I need to plan my anniversary next weekend. I’m thinking of going to this city. Can you research, you know, possible hotels, areas, activities to pursue, here’s my budget, here’s my personal interests, and then the AI can just run away, and it can literally take control of a computer.

It can take control of a machine, it can operate a web browser, it can search the internet, it can research and explore different avenues, hypotheses, and then ultimately it can achieve a goal or fail to achieve a goal and it can come back to you with a result. The key things to take away of all these four. These are our high level, like I said, I’m somewhat simplifying, so the AI purists on here may roll their eyes a little at the simplification, but high level, all of these innovations are approximately two years apart.

So we went from the first ChatGPT around November, 2022 in the mainstream to agentic around two years and one month later. And there’s even more besides, again, this is a simplification, but you would not be blamed for feeling a little bit overwhelmed by the resulting momentum. Right now, we’ve seen so much innovation in the last two years that we could literally take the next 10 or 20 just implementing it and we’d still be catching up.

So this diagram shows human progress through time and the absolute stratospheric takeoff that we’ve seen in the last two years. I myself read about five AI papers a week, and I barely feel like I’m scratching the surface.

But anyway, let’s take a look at that, but apply it to accessibility. So we’re gonna do a bit of a deep dive. I’m gonna do a deep dive into alternative text. Now, the reason for this is it’s gonna serve as an indicator of a broader theme, which is how these innovations can be applied to pretty much any aspect of accessibility.

So I’m gonna assume that everyone on this call understands what alternative text is, essentially a text alternative to multimedia that not everyone could otherwise experience. But before we get into this, I want you to ask yourself a quick question. On a scale of one to 10, how good do you think current day AI is at describing images?

So if you ask AI to, for example, write alternative text for an image, how good do you think it is? Rate it from like one, one is terrible, 10 is perfect, out of 10, I’ve got my own number. I encourage you to hold your own in your head. And what do you think it gets wrong? A lot of people have some idea of this already, but just keep that in mind.

So here’s an example image. Now this gurning chimp this, this is me, unfortunately apologies for that, I hope no one’s too traumatized by this photo. This is what I look like. Sorry. But I’m gonna ask some AI to caption, or actually specifically to write alternative text for this image. Now.

I’m gonna start off and I’m gonna use a $3 trillion corporation, that’s Microsoft, and I’m gonna use their latest version of Microsoft Office, which automatically writes alternative texts for you. And we’re gonna see what they come up with. So Microsoft Office gives us the alt text of “a person with a beard and a scarf” which…

It’s not great, but this is an important point that you’re going to come across a few times today, which is you should not be misled by mainstream AI. Many of the AIs that many of us experience in day-to-day life, such as Alexa or Siri or Microsoft Office and so forth, are not remotely representative of the actual AI you need to be concerned with.

In fact, they’re all generally pre-generated or they’re extremely early forms of generative technology, which essentially means they’re not even in the same universe as the AI you need to be thinking about. So ignore those. Let’s take a look at something a bit more modern. We’re gonna look at ChatGPT, which I’m sure you’ve heard of, and we’re also gonna look at Claude.

So ChatGPT 4.5, their latest, says “A man with a neatly trimmed reddish brown beard wearing a scarf and blazer stands in front of a sign featuring the classic Atari game Pong, which reads ‘Pong 1972 Atari, California.’ Below it is another sign that says, this room is sponsored by Silktide, accompanied by the Silktide logo.

Now this is, in my opinion and I’ve worked with alt text for a long time, this is pretty damn good, to be honest. If a human wrote this, I would probably say, well done. This is, you know, good job. But there’s some interesting points here to explore. And one in particular is the fact that it figured out there’s a Silktide logo in there.

If you think about it, there’s no training data or highly unlikely to be training data for that. This is where the AI has inferred what something is, based upon, in this case, the presence of text written stylized form next to an abstract symbol. Now it’s obvious to you and I as humans, that is in fact a logo.

But it’s not necessarily obvious. It certainly wasn’t obvious to AI for the longest time, but that was the case. But now this is a solved problem. Also by the way, managed to capitalize Silktide, figuring out there’s a proper noun despite the fact that of course it wasn’t written as one. So it’s kind of impressive there.

But let’s move on to another example. This is a more abstract image and Microsoft Office will tell you this is a watercolor of a person running, but ChatGPT, more appropriately, will tell you it’s an expressive watercolor illustration of a running figure depicted in dynamic shades of blue, teal and purple.

Now it goes on. But the thing that I think is particularly noteworthy here is the end where it says it’s creating a sense of speed, motion, and energy. So this is something that often catches people out. They think of AI in the sense of maybe a year, two years, three years ago, as looking at an image and drawing a box around a bit and going, “cat” or, you know, “traffic cone.” That is dinosaur technology. The stuff that is widely available. Everything you’re seeing here is, by the way, is available for free. So you can just Google for this and play with it. But this technology can do not just interpretation of the physical contents of an image, but also of the intent of the meaning of the purpose.

And it’s implying that, it’s inferring that, from what it sees, just like a human. Now this is an even trickier example. This is a photo from my phone. I chose it on purpose to be awkward. That’s just who I am. And it’s a rich, saturated purple hued photo of a party venue. And not an easy thing for an AI or frankly a human to describe, but let’s give it a shot. Microsoft Office with its $3 trillion budgets, I should say revenue, came up with “a room with table and chairs” which is technically correct, but not particularly helpful.

ChatGPT correctly establishes that this is an Alice in Wonderland-themed party, which is not stated. There’s no text there. What it’s figured out is that there’s a large screen which has Disney’s Alice in Wonderland in the background, and that there’s things like a Mad Hatter hat on the table, and it’s correctly ascertained from all of that that we in fact are running a Mad Hatter’s Tea Party, which is exactly what we were doing.

But holy cow, AI figured all that out by itself. It’s pretty well written as well. So, just to come back to my earlier question, how could you think AI is at describing images? Maybe update your number. I’m not sure. I’m curious what you think. Anyway, what I wanna do is I wanna take alternative text and what we’ve just seen so far, and I want you to consider it in a broader model that I’m going to describe now, which helps you to think about any progression in any dimension of accessibility.

And we’re gonna divide this into six pieces you’ll see. But we’re gonna start off with inclusive design and within inclusive design, I want you to think of how we start. So we start off, in this case with alternative texts. We start off by manually writing alternative texts. That’s just what we do, right?

And that’s what we’ve done for 20 or so years. Now, that’s great, and it absolutely works, and is a pivotal part of making many websites and applications accessible. What we’re starting to see with AI is what I’m calling Supervised, which is where the AI helps humans write alternative text, but humans then have to work with the AI to approve it or to correct it, or to basically, you know, reject it in some cases, right.

Now, where we’re clearly headed to, and I’m trying to illustrate this with the quality of the alternative texts you’ve just seen here, which by the way, these are not cherry picked examples. I literally, I had like 20 like this and they were all this good, is that we are headed to a world where this is automated, right?

It’s pretty clear. That’s just a matter of time. Honestly, I would argue it’s possible today in the, we can talk nuance in the Q&A, but let’s just take that for granted for now. Assume that AI can write alternative text. This actually creates an interesting challenge, because instead of the people building the design working on this, arguably, this transitions into the assistive tech.

So this is where the user experiencing the content potentially could have a new level, which I am calling Assumed, and Assumed is where AI could essentially guarantee the presence of alternative texts because it’s able to generate it automatically on the fly. So rather than requiring, for example, application builders or website builders to write the alternative texts, it just happens automatically.

Now we can go further than this, and this is where things get really interesting. We don’t have to provide the same alternative text to everyone now because it’s being generated on their device. So we can personalize it. We can have alternative text for me. I might prefer alternative text to be terse and short and direct.

Someone else might prefer rich, evocative language that particularly draws attention to say the sensory characteristics of what they’re missing in an image. You know, what are the tactile senses? For example. Everyone can have their own personalized requirements, and as long as the AI is generating that text for you on your device in your assistive tech, then there’s no reason why it can’t be tailored for each and every person.

But lastly, any completely new paradigm is to introduce something I’m calling Conversational, which is where the user can talk to their media. So let me give you an example. Say I had the alternative text for this image. Now the alternative text we had earlier was rich and evocative. And nonetheless, it lacks detail, right?

It lacks precision because there’s no way to get around the fact that there’s basically infinite information in this image. And you can only describe so much of it in alt text. But what if I were to ask my computer or my assistive tech, what’s on the table? What are the largest items first? Well. AI again has absolutely no problem answering that question to a human level right now.

Right. A black top hat with rabbit ears, a red teapot, several teacups and saucers… I was impressed by this, but I was even more impressed by Claude, which if you haven’t used by the way, Claude’s great, when I asked it the same question, ’cause it gave more or less the same answer, but it, it beat me,

I’m not gonna lie, ’cause its first answer was “A white tablecloth covering the entire table.” I was like, oh shoot. The AI smarter than me. Oh, well, yes. A more practical example might be something like this. So if you’ve ever tried to take an image of this nature and create an accessible version of it, it is very, very challenging.

This is a complex line chart with two separate axes. We have bands here representing periods of time, indicated through color, recessions and pandemics. There’s also footnotes and so that are relevant. This is a complicated chart and honestly, a lot of humans would just struggle to interpret or describe this chart in any way.

If you were tasked with making an accessible version of this, you would probably want to create some alternative text that indicates the overall, you know, a chart indicating these trends. And then you would probably need additional long descriptions or potentially an entire webpage devoted to explaining the data and the nuance and the, you know, interpretation of this chart, which would be a lot of work, but also a poor experience for the person consuming it, because do you want to listen to a page of text in order to try and visualize what the chart is conveying? People just don’t work like that. But what might work, and is now possible, is to just ask that chart questions. Assuming you can’t see it, you could simply ask what happened to fertility rate and AI today finds this kind of level of analysis

absolutely trivial. This is, again, I’m not using this is not cherry picked example. This is not using some like theoretical science fiction tech. This is, in this case, I think I’m using Claude 3.7. Completely free. Completely free. You can just Google for it, go there, type it in, drag and drop the image, done.

So there is no way that this technology does not eventually find its way into assistive technology browsers and so forth. I would say that AI doesn’t just translate visuals, it interprets their meaning, and honestly at this point, arguably it’s better than most people. So something perhaps to update your priors on.

Now, that was alternative text. I’m now gonna do a faster take on just a couple of other areas. I wanna show you captions. Now, again, I’m assuming everyone here knows what captions are, but specifically I want you to think, how good do you think AI is at present, at transcribing speech, and what does it get wrong?

Chances are you’re experiencing some version of that right now, because I believe this video is being auto transcribed and my guess would be not too great, but who knows? I can’t see the transcription right now, but just keep a number in your head, 1 to 10, how good you think AI is? So we’re gonna start off with one example.

I’m gonna call this difficulty level British. So this is my British coworker. Her name is Anna Bell, that is actually her first name, surname, Anna Bell. And, she often leaves me voice memos which are transcribed by AI, and she starts her videos generally with the words “Hello, Oliver,” which is pretty straightforward. And using the latest in pioneering voice transcription technology, “Hello, Oliver” is reliably transcribed by Slack and others into the words, “Hey lover,” which as I point out to her and my girlfriend is not the same thing. Anyway. This would give you, again, potentially the false impression that transcription is an elusive problem, or at least one that we have a long way to go before we solve.

But I would encourage you not to be misled by mainstream AI in the same way that Siri is terrible at understanding what I’m saying. I routinely use technology on my computer that is virtually perfect at understanding what I’m saying. And the difference is I’m using the latest technology, or something at least within the last six months.

Now again, just to come back to our original model we saw with alternative texts, you can apply the same model here to captions. So once upon a time, humans write captions, then AI has started to help humans write captions. Many of us use those tools today and we’re headed to a world where AI can just write the captions completely.

Like it is a matter of time. There’ll never be a hundred percent efficacy here, but there isn’t a hundred percent for humans either. The current state-of-the-art for AI is around 98 to 99% accuracy. The state-of-the-art best humans for captions is about 99.9, but I do consider it inevitable that we reach that point.

Now with assistive technology, again, AI could guarantee captions, it could write customized captions, again for example, adapting the language or the complexity of the words, or taking out jargon. And users could talk to their media. So you could ask questions like, who said that? Or, why are they laughing?

AI can understand spoken intent, subtext and tone, and at this point it’s basically better at transcription than most humans. I consider it inevitable that it eventually reaches the expert level. And lastly, I just wanna talk about accessible writing. So accessible writing is a particular passion of mine.

There’s a wide range of standards in WCAG which cover writing. And the problem with all of these, they’re treble-A and unfortunately they’re never really met. They’re not really even stride for. These are things about reducing jargon or explaining abbreviations, but also, if text requires reading ability beyond around a ninth grade level, providing a simplified version, and the problem with this is W3C who wrote the standards freely accepts that the concept of easy-to-read cannot be universal. In quoting them, it will not be possible to write a text that will suit the ability of all people with literacy and comprehension problems.

They acknowledge that whilst this is well intentioned, it is simply not possible. Except now with AI, it could be. We forget all the inclusive design for a moment and just skip straight to assistive tech. AI could guarantee accessible text. It could rewrite what you are reading on the fly. It could personalize it for you and it could make it conversational.

What would that look like? Well, imagine you’ve got texts like this and English is your second language, and you are reading some sentence like “Before commencing any physical exertion, it is imperative to conduct an assessment of one’s physical readiness.” Could rewrite that on the fly. Again, AI can do this today. “Before starting exercise, it’s important to check if your body is ready.” Honestly, they should probably have written that in the first place, but if you didn’t, why not adapt it for someone? Likewise, if they have ADHD, you can take complicated prose and you can just, you know, convert that immediately into bullet points. And if you had something like this, this is of particular interest to me, and I think often surprising to people.

Consider you’ve got some emotionally challenging text. Something like “The diagnostic examination indicated a malignant neoplasm necessitating immediate intervention.” And imagine, and this is not a cherry-picked example again, AI does this today, rewriting that. “I know hearing the word cancer can feel overwhelming and frightening.

Your test results indicate that treatment should begin soon.” Now that is doable today. Nothing here is not already doable today. It is nearly a matter of time before this kind of technology is distributed more broadly. I would argue that the most accessible content won’t be universally designed. It will dynamically reshape itself to uniquely fit every user.

And if we do this, we might actually achieve some of those long elusive treble-A guidelines. So just wrapping up the impact of AI on accessibility. Firstly, we’re gonna dramatically expect that the burden of making things accessible is going to be reduced. A lot of routine tasks will gradually fade away, whether they happen to concern, say, the writing of captions or alternative text or code compliance, and so on and so on.

We’re gonna see increased empowerment where users of assistive tech gain unprecedented autonomy. And we’re gonna see raising standards, which I think is what excites me the most. Like we can rethink how good accessibility can be. Why do we have to be constrained to a second rate alternative version? Why can’t I literally talk to and experience my media in a way that’s, frankly, is so empowering

I suspect many people who don’t even use assistive tech would choose to use this. So this is where we are. This is 2025, at the start of, and by the end, goodness knows where we’ll be. But I think it’s incumbent on us both as individuals and as an accessibility industry to really take stock of these profound and unprecedented opportunities and think, what is the kind of world we wanna build with this kind of technology?

And I personally am very excited to be in that world. Right. That was my talk. Thank you for listening. I’m gonna turn over now Jess, if you are still around, to questions. We’ve got about five minutes to cherry pick.

Jessica:
Chat’s been awesome actually.

Oliver:
Okay. Wow.

Jessica:
So someone had a question, I think is really good.

Can I ask if AI can recognize if alt text’s actually needed? Sometimes the images should be marked as decorative and therefore an explanation’s not needed. Sometimes this is relevant due to the accompanying text, which explains what the image is trying to convey. Do we have a way for them to actually use context?

Oliver:
So a general rule I like to give people is everyone seems to think of AI as software, and that’s the way it used to work, right? And I would encourage you to think of it, to a first approximate, as more like a human. So, if you rethink the question for a second, I know it sounds crazy, but bear with me, rethink the question as, could you get a human to determine if text is decorative or not?

And of course, of course you could, obviously, but to do so, what would the human need? So the human would probably not just need the image, right? Like if I just gave you an image, do you know if it’s decorative? This image on my Q&A, is that decorative? I would argue it is in this case, but maybe if I was showing you a photo of a human cell or something, that’s incredibly important.

So the context matters, and in the same way, the AI needs to be given that context one way or another. So the answer to your question is basically, yes, I believe it can already do it, but the bit that’s usually missing is connecting the AI to not just the image, but also the necessary context for it to be able to apply itself and do that.

And that’s the kind of challenge people building software products of the future are going to have to solve.

Jessica:
Oh, you’ll like this. Crossing the language barrier. Could websites be automatically made for many languages by providing automated translations?

Oliver:
I mean, so many of you probably use something like Google Translate in your browser where you’ve gone to a webpage in a different language and it’s kind of attempted to translate it in real time.

It’s exactly the same thing. We’re just getting better at the translation. There are challenges with that, like it can break the layout, for example, you, again, you’ve probably seen that happen, but it’s absolutely possible today. I mean, imagine, I could literally, and I don’t have the time, but I could take a screenshot of any webpage right now, take the screenshot, put it into Claude or ChatGPT right now and say, translate this into French or something. It will just do it. If you haven’t tried that, you should do it. And you can say, do that, but also make it simple French or, assume that I’m six years old or that I am impatient or easily upset, or whatever you wanna do.

You can do anything you like, and that already exists. What doesn’t currently exist is transferring these capabilities into mainstream products that everyone expects to use, but that is probably months to certainly single years away, and many other people on the call may in fact be constructing those things today.

Jessica:
Do you believe that accessibility tests could be run with a hundred percent accuracy by AI? That might be related to Axe-con.

Oliver:
Oh, oh oh, that’s a fun one. Okay, so, the simple answer is I don’t believe that accessibility checks could be run at a hundred percent by humans. So the simple thought experiment, take amazon.com, imagine that you were to put a thousand people with a range of different experiences you know, disabilities, et cetera.

Thousand people to test amazon.com and leave them alone for a year to test amazon.com, they’re still not gonna get a hundred percent coverage, because it’s a website with, I don’t know, 500 million pages and essentially infinite things you could do with it. So there is no such thing as a hundred percent accessibility testing.

That’s a myth, although it’s a very compelling one. I think AI will close the gap eventually where it gets to human level which could be whatever you wanna call that, 95% or 90% or whatever. But it’s all just invented numbers. It doesn’t really mean anything, and it’s subject to the specifics of the website.

Think we’ve got time for literally one more question. It’s gonna have to be about 40 seconds.

Jessica:
Okay. It sounds like for you, AI can do everything or will be able to do everything. What’s your vision of how things will be in the future when it comes to accessibility?

Oliver:
I have literally written an essay on that topic on our website.

If you go to our careers area, there’s mission, and essentially I believe AI will empower us to be able to make accessibility the default. So literally as expected as making a website work on mobile, I think we’ll get accessibility to be expected in the same way, but unfortunately, real answer is a lot longer than that. So, encourage you to look it up.

Okay. Thanks a lot guys. That’s been great. I hope you found it useful. Thanks a lot, guys. Bye.