WEBVTT 1 00:00:03.530 --> 00:00:21.179 David Sloan: Welcome to the Stage of Accessibility podcast from Vispero. This is episode 18. I'm David Sloan, Chief Accessibility Officer, and each month I'm joined by a guest or two to discuss a topic of interest and relevance to the practice and profession of digital accessibility. 2 00:00:21.310 --> 00:00:25.650 David Sloan: And for this episode, the April 2026 episode. 3 00:00:25.790 --> 00:00:36.999 David Sloan: We're returning to a topic that we've addressed in several previous podcasts, and really, honestly, which we can't ignore for any length of time, given the rapid pace of change. 4 00:00:37.000 --> 00:00:54.240 David Sloan: Yep, it's artificial intelligence again. And joining me this month to discuss what's changed in AI's impact on digital accessibility efforts is a core member of Vispero's Knowledge Center, Principal Technical Writer, Ricky Onsman. So, welcome, Ricky. 5 00:00:55.210 --> 00:00:57.440 Ricky Onsman: Thank you, David. Lovely to be here. 6 00:00:57.910 --> 00:01:02.949 David Sloan: Yeah, and thank you for getting up early to join us for this conversation. 7 00:01:03.300 --> 00:01:16.750 David Sloan: I know you've been tracking the impact of AI and digital accessibility for a while now, and it sounds like from your very first effort, once you joined TPGI Vispero. 8 00:01:17.220 --> 00:01:35.810 David Sloan: So, talk through what's piqued your interest, how you've been tracking and reporting on progress in AI from an accessibility perspective over recent years, and you could include, kind of, defining the different types of AI that are making a difference to how we build and make sure that digital content's accessible. 9 00:01:36.870 --> 00:01:37.570 Ricky Onsman: Sure. 10 00:01:37.730 --> 00:01:52.229 Ricky Onsman: Just for perspective, so that a lot of people won't know exactly what the Knowledge center is, what we do is we provide… we write the guidance for our TPGI, the Sparrow. 11 00:01:52.490 --> 00:02:07.370 Ricky Onsman: Accessibility engineers, so we work on the rules engine that people use to assess and audit whether, websites and apps are meeting accessibility requirements, and we provide guidance to our clients. 12 00:02:07.400 --> 00:02:12.519 Ricky Onsman: on how to address specific accessibility issues. So… 13 00:02:12.840 --> 00:02:20.329 Ricky Onsman: Artificial intelligence comes very strongly into our, kind of, area of work. 14 00:02:20.570 --> 00:02:34.410 Ricky Onsman: And in fact, I've been working for Vispero for TPGI for five and a half years, and I realized the other day the first thing I actually did for TPGI was write a blog post on artificial intelligence. 15 00:02:34.780 --> 00:02:42.459 Ricky Onsman: I reviewed that the other day, and I actually now, five and a half years later, I would write a very different article. 16 00:02:42.670 --> 00:03:00.249 Ricky Onsman: It was, it was very positive, and I remain positive about artificial intelligence, and I think we should, make that clear from the start. I think it has enormous potential, in society in general, but very specifically in how we approach digital accessibility. 17 00:03:00.470 --> 00:03:06.169 Ricky Onsman: I think it has been a hard slog in the world of digital accessibility. 18 00:03:06.290 --> 00:03:15.750 Ricky Onsman: And there are a lot of ways that artificial intelligence can help us to achieve our aims in making the digital world a more accessible place. 19 00:03:16.310 --> 00:03:17.620 Ricky Onsman: Having said that. 20 00:03:17.900 --> 00:03:29.090 Ricky Onsman: I think one of the things that has probably changed in the way I think about artificial intelligence is, first of all, we've got to stop calling it artificial intelligence. 21 00:03:29.610 --> 00:03:32.520 Ricky Onsman: For one thing, It's not very intelligent. 22 00:03:33.120 --> 00:03:38.019 Ricky Onsman: It's a terrific tool, but a lot of it is just automation. 23 00:03:38.230 --> 00:03:42.630 Ricky Onsman: And it's not… doesn't necessarily show a great deal of insight. 24 00:03:43.450 --> 00:03:45.860 Ricky Onsman: As an example of that. 25 00:03:46.340 --> 00:03:53.680 Ricky Onsman: When I ask a generative AI chat model, to… 26 00:03:55.380 --> 00:03:59.239 Ricky Onsman: Let's say I have a set of code here. 27 00:03:59.380 --> 00:04:04.949 Ricky Onsman: And, it's got a bunch of divs and spans in it. How do I make them accessible? 28 00:04:05.080 --> 00:04:11.080 Ricky Onsman: And what it will tell me is to use ARIA, Labels to make it accessible. 29 00:04:11.490 --> 00:04:16.450 Ricky Onsman: And… What it doesn't do is question my basic premise. 30 00:04:16.550 --> 00:04:31.700 Ricky Onsman: It doesn't say, why are you using divs and spans? You should be using semantic HTML. It just answers the question that I put to it, but it doesn't make me question myself, and it doesn't make me think about what I'm doing. 31 00:04:32.020 --> 00:04:34.869 Ricky Onsman: That's a real problem, because then we just… 32 00:04:35.200 --> 00:04:42.589 Ricky Onsman: perpetuate the same issues, instead of rethinking things. Now, that would be intelligence. 33 00:04:42.710 --> 00:04:46.660 Ricky Onsman: And I don't see a lot of that evidence in what 34 00:04:47.100 --> 00:04:54.030 Ricky Onsman: Claude, ChatGPT, your other artificial intelligence, chat, models do. 35 00:04:54.210 --> 00:04:58.939 Ricky Onsman: And they are… almost sycophantic. 36 00:04:59.140 --> 00:05:08.519 Ricky Onsman: In that they will respond in a lovely conversational tone. Oh, that's a very good point, you've raised a terrific thing there, but they don't make me think. 37 00:05:09.450 --> 00:05:14.429 Ricky Onsman: So, I think one of the things there is we've just got to find a different name for it. 38 00:05:14.710 --> 00:05:15.740 Ricky Onsman: And, 39 00:05:15.890 --> 00:05:22.280 Ricky Onsman: I guess one of the things that comes out there is that we should stop referring to it altogether. 40 00:05:22.470 --> 00:05:26.370 Ricky Onsman: I think a good model for that is what happened with Google Translate. 41 00:05:26.820 --> 00:05:31.930 Ricky Onsman: In 2016, Google Translate moved to an artificial intelligence model. 42 00:05:32.110 --> 00:05:35.189 Ricky Onsman: Its accuracy rose by 80%. 43 00:05:35.370 --> 00:05:45.860 Ricky Onsman: And nobody thought about it. Nobody questions it. Now, you use Google Translate, and you don't think, I'm using artificial intelligence. You're just using Google Translate. 44 00:05:45.940 --> 00:06:00.040 Ricky Onsman: What drives it, what powers it, is almost irrelevant. The question is, does it do a good job? If it does a good job, fantastic. And that's how we've got to kind of incorporate artificial intelligence into what we do with digital accessibility. 45 00:06:00.140 --> 00:06:11.710 Ricky Onsman: So that when it helps us write code or address issues, we don't think about it being artificial intelligence. It's just a tool that's helping us to do our job. 46 00:06:12.120 --> 00:06:13.390 Ricky Onsman: Does that make sense? 47 00:06:13.700 --> 00:06:16.660 David Sloan: Absolutely, yeah, and it also helps 48 00:06:17.190 --> 00:06:20.120 David Sloan: Move us on from this kind of pressure to be 49 00:06:20.220 --> 00:06:26.380 David Sloan: using AI, quote-unquote, without really un… understanding 50 00:06:27.360 --> 00:06:31.910 David Sloan: what do you mean, and what for? You know, it's just this… it's almost like a… 51 00:06:32.020 --> 00:06:38.290 David Sloan: A bandwagon, or a, you know, something that you feel like you have to be doing, and have to be showing that you're doing without 52 00:06:38.760 --> 00:06:40.829 David Sloan: Really having the context of 53 00:06:41.190 --> 00:06:58.260 David Sloan: where can I… where can I make my work more efficient? Where can I make it quicker and reduce errors, or whatever? So, almost taking the… taking the underlying technology out of the equation and just thinking, what are the tools that I could be using that I'm not using now, will stop people feeling 54 00:06:58.510 --> 00:07:03.760 David Sloan: I don't know, I don't know where to go, or… or I… or worse, kind of. 55 00:07:04.170 --> 00:07:09.989 David Sloan: Turn the opposite direction and kind of walk away from something that could help them. 56 00:07:10.320 --> 00:07:15.060 David Sloan: do their job more… more effectively. You know, putting aside the 57 00:07:15.210 --> 00:07:18.979 David Sloan: Or now the, you know, the reasonable concerns around 58 00:07:19.630 --> 00:07:24.330 David Sloan: AI in terms of the, sort of, economic and the environmental 59 00:07:24.690 --> 00:07:30.499 David Sloan: aspects, but I do like the idea of just, like, let's get rid of the whole 60 00:07:31.480 --> 00:07:33.900 David Sloan: Raise and talk about the tools that 61 00:07:34.080 --> 00:07:38.660 David Sloan: Absorb it and make use of it, and then maybe it becomes more… 62 00:07:39.000 --> 00:07:44.050 David Sloan: In a way, more accessible to people who just need They're always looking for tools. 63 00:07:44.320 --> 00:07:53.769 David Sloan: to do their job better, and I guess in the accessibility profession, that is a place where we're always looking at tools that can automate things that we trust will 64 00:07:54.000 --> 00:07:59.620 David Sloan: Do a better job than we could do on our own, and cut down on errors, and cut down on time. 65 00:08:00.820 --> 00:08:06.849 Ricky Onsman: Yeah, I think that's right, and I mean, I think there are two main drivers in that context. 66 00:08:07.060 --> 00:08:13.440 Ricky Onsman: One is that it can help us to carry out tasks in an automated way. 67 00:08:13.600 --> 00:08:15.779 Ricky Onsman: That we can rely on. 68 00:08:16.130 --> 00:08:30.199 Ricky Onsman: And the other is that it can replace certain things that we do. So, in the latter context, what I'm thinking of is things like generating alternative text for images. 69 00:08:31.080 --> 00:08:35.139 Ricky Onsman: Even 5 years ago, The job that it did… 70 00:08:35.289 --> 00:08:41.850 Ricky Onsman: was not very good, because what it did was… it did the same thing that inexperienced 71 00:08:42.070 --> 00:08:51.979 Ricky Onsman: accessibility engineers do when assessing the purpose of an image. They look at the image and say, it is this. It's a dog sitting at the base of a tree. 72 00:08:52.080 --> 00:09:00.330 Ricky Onsman: And AI would generate alternative text that says, dog sitting at the base of a tree. It doesn't tell you why that image is there. 73 00:09:00.920 --> 00:09:17.850 Ricky Onsman: Nowadays, they're actually a lot better. Gemini is a lot… has become very, very good at understanding the context of the image, why that is there, and telling you the dog is sitting at the base of the tree because it's hunting truffles. 74 00:09:18.060 --> 00:09:24.190 Ricky Onsman: You know, and that comes from the surrounding text. It understands the context better. 75 00:09:24.650 --> 00:09:29.469 Ricky Onsman: If we can get to the point where we don't have to think a lot about 76 00:09:29.570 --> 00:09:40.350 Ricky Onsman: what we want to put in alternative text. We can rely on artificial intelligence to do it accurately. That's going to make our job a lot easier. And we know, you and I both know, that 77 00:09:40.540 --> 00:09:59.460 Ricky Onsman: Alternative text for images is one of the top issues that exists. We see the WebAIM million report, and every year, and it's not getting better. And the same issues keep coming up that account for 90% of the problems that exist in accessibility on the web and in apps. 78 00:09:59.580 --> 00:10:12.059 Ricky Onsman: And one of those is that images don't have alternative text at all when they should have, or that the alternative text doesn't convey what the image is actually supposed to convey. 79 00:10:12.320 --> 00:10:22.100 Ricky Onsman: So if artificial intelligence can do that for us in an automated way, and we can trust it without having to think about it, then that works really well. 80 00:10:22.600 --> 00:10:35.099 Ricky Onsman: Having said that, at this stage, I would not trust any AI to do that job for me. What I would like trusted to do is make suggestions that I check. 81 00:10:35.970 --> 00:10:47.140 Ricky Onsman: So, it comes up with, this is an image of… and I… no, you're missing the point here. That's not what the image purpose of this image is. The purpose of this image is this. 82 00:10:47.260 --> 00:10:49.839 Ricky Onsman: And it then carries out that task for me. 83 00:10:49.980 --> 00:10:57.649 Ricky Onsman: Now, that's a step in the right direction. It's not total automation, because we can't rely on it with full confidence. 84 00:10:57.900 --> 00:11:01.479 Ricky Onsman: But it is a step towards making our jobs easier. 85 00:11:01.720 --> 00:11:10.480 Ricky Onsman: And that applies as well to using artificial intelligence in content management systems, to say, you haven't put in alternative text for this image. 86 00:11:10.930 --> 00:11:12.169 Ricky Onsman: What should it be? 87 00:11:12.360 --> 00:11:14.280 Ricky Onsman: To make us question ourselves. 88 00:11:14.800 --> 00:11:21.190 Ricky Onsman: And to act like a checklist of things that you need to do. And artificial intelligence is good at checklists. 89 00:11:21.310 --> 00:11:25.949 Ricky Onsman: It's good at coming up with things like, have you done this? Have you done this? Have you done this? 90 00:11:26.270 --> 00:11:29.740 Ricky Onsman: But it's still up to us to carry out the tasks. 91 00:11:29.860 --> 00:11:34.379 Ricky Onsman: And it requires some human intervention, at least at this stage. 92 00:11:34.480 --> 00:11:38.339 Ricky Onsman: Frankly, I don't see that changing a lot over time. 93 00:11:39.650 --> 00:11:54.420 Ricky Onsman: An example of where we put too much trust in artificial intelligence is the dreaded overlays, which rely on, you know, that you can, for a very low cost, install a widget on your website. 94 00:11:54.540 --> 00:12:00.970 Ricky Onsman: press a button, and it will solve all the accessibility problems, because it is AI-powered. 95 00:12:01.380 --> 00:12:05.809 Ricky Onsman: And that is just, it's basically almost snake oil. 96 00:12:06.040 --> 00:12:13.109 Ricky Onsman: A single push-button response to all accessibility issues on a website or an app 97 00:12:13.350 --> 00:12:15.829 Ricky Onsman: It's a long, long way off. 98 00:12:16.210 --> 00:12:18.950 Ricky Onsman: But people are using them now. 99 00:12:19.150 --> 00:12:24.069 Ricky Onsman: And that's a problem, because it ends up causing more accessibility issues, not less. 100 00:12:24.600 --> 00:12:26.760 David Sloan: Yeah, and that's… this is a… 101 00:12:27.650 --> 00:12:41.989 David Sloan: a theme that… I attended an AI and accessibility webinar last week that Accessibility.com hosted Mark Shapiro and Gerard Cohen talking about advances, and one of the things that I think it was Gerard said. 102 00:12:42.490 --> 00:12:51.679 David Sloan: was that there are companies that look at AI as a way, oh, we can use this for our accessibility strategy, and we can fire all our accessibility team. 103 00:12:51.810 --> 00:12:59.429 David Sloan: And then a more mature way to look at AI is we can empower our team to do more and do better, so there's this kind of line between 104 00:13:00.130 --> 00:13:09.510 David Sloan: You're on one side, it's, like, a less mature approach that thinks that AI's already capable of 105 00:13:10.050 --> 00:13:16.400 David Sloan: Replacing humans to do tasks like, your accessibility, quality assurance. 106 00:13:16.920 --> 00:13:24.209 David Sloan: And, like you say, it isn't. And that leaves companies with more problems to solve. 107 00:13:24.610 --> 00:13:28.950 David Sloan: But there's no longer any staff to… to address them, so… 108 00:13:29.300 --> 00:13:31.410 David Sloan: How… how do you see that? 109 00:13:31.730 --> 00:13:34.969 David Sloan: Challenge being addressed over time, where 110 00:13:35.160 --> 00:13:41.450 David Sloan: You have this kind of less mature perspective that Thinks that the technology's already 111 00:13:42.370 --> 00:13:47.300 David Sloan: Already can replace humans, and they make decisions based on that, but then create problems. 112 00:13:47.550 --> 00:13:50.559 David Sloan: Further down the line that they don't have the capacity to fix. 113 00:13:51.800 --> 00:13:56.130 Ricky Onsman: I think one of the things that has created that is what… 114 00:13:56.260 --> 00:14:02.020 Ricky Onsman: A lot of people see as a big advantage of artificial intelligence, the conversational tone. 115 00:14:02.400 --> 00:14:13.859 Ricky Onsman: That you're… you're able to ask a question in human terms of a machine, and the machine responds in what appear to be human conversational tones. It's very convincing. 116 00:14:14.000 --> 00:14:29.729 Ricky Onsman: And it's convinced of its own correctness. It doesn't question itself. It's only when you come back to it and say, are you sure that's right? I didn't think those two flags were the same thing. I mean, no, of course they're not the same thing. This is the correct answer. 117 00:14:29.900 --> 00:14:40.299 Ricky Onsman: And she goes, no, I'm not sure that's totally the correct answer either. And I go, no, it's not the correct answer. This is the correct answer. And everything is given with total certitude. 118 00:14:40.630 --> 00:14:42.540 Ricky Onsman: And that is… 119 00:14:43.060 --> 00:14:50.320 Ricky Onsman: it's certainly artificial, but it's not intelligent. Yeah. And, you know, we've got to… we've got to get into the habit of… of… 120 00:14:50.640 --> 00:15:03.610 Ricky Onsman: thinking of it as an assistive tool. And this is where I think, you know, if we come in our industry to think of artificial intelligence as another form of assistive technology, that would be a big step forward. 121 00:15:03.920 --> 00:15:12.799 Ricky Onsman: Instead of thinking of it being the answer to everything, and that it can be totally relied on, it's something that helps us to do something. 122 00:15:13.130 --> 00:15:16.540 Ricky Onsman: Right. And in those terms, I think we need to… 123 00:15:17.020 --> 00:15:28.979 Ricky Onsman: Dial back the overconfidence that artificial intelligence exudes, and turn it into something that's more interrogative, and something that questions what's going on. 124 00:15:28.980 --> 00:15:41.039 Ricky Onsman: and tries to think about what possible answers it could be. If… if I use ChatGPT, and I ask it a question, and it says, well, the answer could be this, or this, or this. 125 00:15:41.050 --> 00:15:48.079 Ricky Onsman: That would be a lot more useful than it just coming out with one blatant answer that says, oh, I know the answer perfectly. 126 00:15:48.800 --> 00:16:06.350 David Sloan: And then you… and then you have to have the persistence to say, are you sure? And then it might… well, actually, no. So yeah, I like that move towards thinking about it as a kind of position support tool, and a, you know, paired programming, almost, relationship where you're kind of working together and… 127 00:16:07.350 --> 00:16:07.970 Ricky Onsman: Yeah, yeah. 128 00:16:07.970 --> 00:16:11.180 David Sloan: the best way forward. So, when we think about. 129 00:16:11.180 --> 00:16:26.060 Ricky Onsman: In that, you know, you need the expertise to be able to doubt artificial intelligence. In which case, if you have the expertise, why do you need the artificial intelligence in the first place? 130 00:16:26.760 --> 00:16:27.310 David Sloan: Yeah. 131 00:16:27.440 --> 00:16:28.880 David Sloan: That's a good point. 132 00:16:29.280 --> 00:16:36.030 David Sloan: So, you would… you mentioned, you started off an exit with an example of where AI doesn't… 133 00:16:36.260 --> 00:16:38.540 David Sloan: sufficiently interrogate 134 00:16:38.740 --> 00:16:46.259 David Sloan: A coding practice that's suboptimal, let's say, and then tries to kind of remediate the code to improve accessibility. 135 00:16:46.580 --> 00:16:49.870 David Sloan: Have you… You know, stepping back from that, have you seen 136 00:16:50.370 --> 00:16:54.400 David Sloan: You know, genuine advances in the accessibility quality of 137 00:16:54.870 --> 00:17:04.049 David Sloan: code that has been created by generative AI over the years? And if so, where have the advances come? 138 00:17:04.200 --> 00:17:06.039 David Sloan: What have been the factors there? 139 00:17:07.560 --> 00:17:21.590 Ricky Onsman: I think the… it kind of pertains to what we've just been talking about. It depends a lot, and we've come to understand it depends a lot, on what we ask artificial intelligence to do, the prompts that we give. 140 00:17:21.890 --> 00:17:24.770 Ricky Onsman: And if we set out 141 00:17:25.010 --> 00:17:29.330 Ricky Onsman: A clear set of instructions of what our expectations are. 142 00:17:29.470 --> 00:17:38.639 Ricky Onsman: artificial intelligence can achieve those outcomes very, very well. Check my code. Have I used everything correctly here? 143 00:17:38.860 --> 00:17:45.050 Ricky Onsman: Is there a way of making this more accessible that I haven't considered? 144 00:17:45.410 --> 00:18:02.070 Ricky Onsman: Those kind of tools that address those kind of issues have come a long way in recent years. The co-pilot for coding that you get in things like Visual Studio Code, and even in Microsoft products, you know. 145 00:18:02.700 --> 00:18:04.130 Ricky Onsman: That kind of stuff. 146 00:18:04.680 --> 00:18:08.829 Ricky Onsman: To give an example, when you're using a program like Microsoft Word. 147 00:18:09.110 --> 00:18:15.579 Ricky Onsman: Which becomes critical for digital accessibility because of where it goes to. 148 00:18:15.580 --> 00:18:15.950 David Sloan: Yep. 149 00:18:15.950 --> 00:18:22.679 Ricky Onsman: If you have, at the moment, you have accessibility checkers in Microsoft Word, but not everybody uses them. 150 00:18:23.140 --> 00:18:28.540 Ricky Onsman: If it's… artificial intelligence is used to say, you must do this. 151 00:18:28.970 --> 00:18:33.830 Ricky Onsman: I will check the accessibility of this document for you, whether you like it or not. 152 00:18:34.300 --> 00:18:43.690 Ricky Onsman: And that has implications for, you know, one of the big issues that I wrestle with is the accessibility of PDF documents on the web. 153 00:18:44.010 --> 00:18:53.430 Ricky Onsman: And… That is an area where, in generating PDF documents to be accessible, AI has become very good. 154 00:18:53.660 --> 00:19:10.640 Ricky Onsman: For a start, it makes the source documents more accessible, and then it helps the process of the conversion into a PDF document accessible. Now, that still leaves a huge problem for us, because that doesn't address all the PDF documents that are already there. 155 00:19:10.830 --> 00:19:23.890 Ricky Onsman: Right. But it does help in creating new PDF documents that are accessible. And on that front, I think artificial intelligence is very good. And I think the way that companies like Adobe and Microsoft are using it 156 00:19:24.000 --> 00:19:34.079 Ricky Onsman: to… Assist slash force users to consider accessibility issues is a very good thing. 157 00:19:34.230 --> 00:19:37.239 Ricky Onsman: And that's one area where it's been very, very good. 158 00:19:37.570 --> 00:19:40.320 David Sloan: Yeah, I could… I totally agree with you. I think that… 159 00:19:40.710 --> 00:19:47.849 David Sloan: In accessibility, we get… yeah, and there are very valid reasons for doing so, where we get so focused on tools to help us 160 00:19:48.050 --> 00:19:51.070 David Sloan: Test and remediate existing stuff. 161 00:19:51.260 --> 00:19:52.790 David Sloan: We kind of forget about 162 00:19:53.000 --> 00:19:59.010 David Sloan: all the new stuff that's being built, so that ends up also needing tested, and a lot of remediation. 163 00:19:59.050 --> 00:20:14.969 David Sloan: So, it's really encouraging to hear that the innovations are guiding authors of different levels of awareness. You know, let's face it, you know, we talk about accessibility every day, but most people don't. So, any tool that kind of 164 00:20:15.280 --> 00:20:30.050 David Sloan: puts it into somebody's, publishing workflow, whether they… whether they were expecting it or not, and just gives… increases the chances that what they produce is more accessible than it would have… than it would have been had the AI not been there. 165 00:20:30.330 --> 00:20:33.019 David Sloan: I think that's a really encouraging thing. 166 00:20:33.200 --> 00:20:34.789 David Sloan: So, do you see any kind of… 167 00:20:35.110 --> 00:20:40.620 David Sloan: maturity and best practices of using AI in 168 00:20:41.040 --> 00:20:57.340 David Sloan: digital content creation emerging, you know, whether it's tools, or the way that you write prompts, or the way that you feed the tool with examples of what you mean. Are we getting to a point where we're sharing best practices, or are people still 169 00:20:57.580 --> 00:21:01.040 David Sloan: Kind of more figuring it out on their own at this point. 170 00:21:02.220 --> 00:21:11.029 Ricky Onsman: Well, I think it really comes down to your expectations and the limits that you put on artificial intelligence. And by that, I mean 171 00:21:11.470 --> 00:21:14.089 Ricky Onsman: when it… when I first started using 172 00:21:14.320 --> 00:21:24.620 Ricky Onsman: generative AI to ask questions, it became quickly apparent that the data set that AI was using was fundamentally flawed. 173 00:21:24.890 --> 00:21:29.760 Ricky Onsman: Largely, it draws on what's available on the web. 174 00:21:30.280 --> 00:21:42.210 Ricky Onsman: And a lot of the information about digital accessibility on the web is fundamentally flawed, if not downright dangerous. Practices that really don't encourage digital accessibility. 175 00:21:43.210 --> 00:21:59.249 Ricky Onsman: where that has developed, and I think people like agencies like Vispero and other digital accessibility consultancies are coming to realize that you can set a specific data set for an artificial intelligence model to use. 176 00:21:59.250 --> 00:22:03.999 Ricky Onsman: So instead of using a large language model, use a limited language model. 177 00:22:04.680 --> 00:22:22.139 Ricky Onsman: At Vispero, what we've done is we've created a knowledge base assistant that uses AI to interrogate every audit we've ever done, every article we've ever written, every piece of guidance we've provided to a client, and we know we can rely on that. 178 00:22:22.360 --> 00:22:30.050 Ricky Onsman: So, we can rely on the artificial intelligence to produce advice based on sound information. 179 00:22:30.250 --> 00:22:35.109 Ricky Onsman: And I think that becomes a very refined tool 180 00:22:35.270 --> 00:22:39.400 Ricky Onsman: That, is much more useful and practical and reliable. 181 00:22:39.610 --> 00:22:50.520 Ricky Onsman: So it becomes a bit of, you know, like, instead of what was touted as the huge advantage of AI, that it draws on all the information ever in the world. 182 00:22:50.780 --> 00:22:54.640 Ricky Onsman: That it only relies on reliable information. 183 00:22:54.940 --> 00:23:06.300 Ricky Onsman: And, I think all of us are working on tools that do that. And I think that will make a very big difference, not only just in confidence, but in just… just in accuracy. 184 00:23:06.560 --> 00:23:07.389 David Sloan: Right. 185 00:23:07.390 --> 00:23:19.709 Ricky Onsman: You know, and that is happening, so I see that as a very useful thing, and it's not insignificant that it takes humans to shape that. 186 00:23:20.030 --> 00:23:28.460 Ricky Onsman: You know, we have to tell AI what we want it to do, and what it should pay attention to, and what it should not pay attention to. 187 00:23:28.630 --> 00:23:42.649 Ricky Onsman: And that's a… that's a slow process, because we need to figure out, well, how do we… how do we get that information right in the first place, so that AI uses and spreads the right information? 188 00:23:42.880 --> 00:23:57.420 Ricky Onsman: And we're way off that, but it is part of the task that we have at the moment. And digital accessibility is a terrific crucible for that, because not only is there a lot of bad information out there, but it really matters. 189 00:23:57.640 --> 00:24:11.789 Ricky Onsman: You know, this is not a trivial matter. Making the digital world accessible to people with disabilities is only going to get more and more important, and whatever tools we can use to do that, we should use them. 190 00:24:12.030 --> 00:24:15.990 Ricky Onsman: Artificial intelligence is undoubtedly a part of that scenario. 191 00:24:16.150 --> 00:24:26.449 Ricky Onsman: But what we mustn't lose sight of is why we're doing it. And that's… that's… I mean, we all work in digital accessibility because it really matters, you know? 192 00:24:26.570 --> 00:24:42.120 Ricky Onsman: More and more people are having to rely on the digital world for information, for functionality, for processes to do with every aspect of their lives. And we have to make sure that that is accessible to people with disabilities, so we don't disenfranchise them. 193 00:24:42.610 --> 00:24:48.179 David Sloan: Right, exactly, yeah, yeah, no, I totally agree, and I guess it's a bit of a… 194 00:24:48.780 --> 00:24:53.230 David Sloan: There's a race between the content and code generated by the 195 00:24:53.760 --> 00:24:57.630 David Sloan: General large learning models, which are trained on 196 00:24:58.640 --> 00:25:03.119 David Sloan: Code that may not be what we would want it to be trained on. 197 00:25:03.510 --> 00:25:12.930 David Sloan: And then… The code that's created with the assistance of more specialized models that are more accessibility aware. 198 00:25:13.100 --> 00:25:18.939 David Sloan: We'll give more reliable, best practice and guidance and examples. 199 00:25:19.250 --> 00:25:28.199 David Sloan: So this, this race, you know, where the sort of general world potentially is learning from bad practice and, and 200 00:25:28.430 --> 00:25:35.529 David Sloan: Ecolating or propagating bad practice, and then, separately, the smarter models that you were talking about. 201 00:25:35.990 --> 00:25:38.360 David Sloan: That are assisting people, so… 202 00:25:38.890 --> 00:25:41.229 David Sloan: you know, I don't know… I don't know if it… 203 00:25:41.790 --> 00:25:52.070 David Sloan: where that race will head, and who's gonna win it, but I know the part of it that we want to be on and support. 204 00:25:54.170 --> 00:25:59.079 Ricky Onsman: Yeah, there's an aspect to this that we haven't really spoken about so far. We've focused on 205 00:25:59.330 --> 00:26:03.179 Ricky Onsman: Within the digital accessibility industry, how we use artificial. 206 00:26:03.580 --> 00:26:15.079 Ricky Onsman: But there's another aspect to it that I think is going to… is already becoming more pervasive, and will continue to do so, where artificial intelligence assists the user 207 00:26:15.400 --> 00:26:29.599 Ricky Onsman: And I'm talking here about agentive artificial intelligence that makes a browser act like a screen reader, that lets a browser understand a particular user's needs, and 208 00:26:29.700 --> 00:26:44.440 Ricky Onsman: literally influence the presentation of web and app content to suit that individual user's needs, to customize digital experiences so that they meet every person's individual needs. 209 00:26:44.440 --> 00:26:50.759 Ricky Onsman: That's an area that I think is going to explode, and in the… in the short term. 210 00:26:50.800 --> 00:26:55.960 Ricky Onsman: Browsers, by definition, have been for the average user. 211 00:26:56.380 --> 00:27:11.779 Ricky Onsman: But if you can customize your browser to be something that is incredibly personalized to you, even to the extent of not looking like a browser, it just presents web content to you in the way that is most meaningful and useful to you. 212 00:27:12.290 --> 00:27:16.610 Ricky Onsman: Taking into account things like disabilities, 213 00:27:16.880 --> 00:27:20.889 Ricky Onsman: Language backgrounds, cultural backgrounds, all those kind of things. 214 00:27:20.970 --> 00:27:36.959 Ricky Onsman: If you can have your own personal AI present web content to you in the way that you need it to be presented, that's a massive step forward. And that is something that I think will happen in the short term. We will see that happening in the next couple of years. 215 00:27:36.960 --> 00:27:47.760 David Sloan: Absolutely, and there's a whole other podcast to talk about how assistive technology can be empowered by agentic AI, and how your assistive technology is 216 00:27:48.320 --> 00:27:53.700 David Sloan: You know, the capability of it is… amplified 217 00:27:54.390 --> 00:28:01.650 David Sloan: you know, I don't know what kind of level of order, that you can ask your screen reader, whatever, to 218 00:28:02.190 --> 00:28:04.249 David Sloan: Go and book my flight to… 219 00:28:04.650 --> 00:28:19.999 David Sloan: to Cancun, or wherever, and you're not even interacting with the website, the agent can do things for you. So, you know, empowering the browser, or empowering the assistive technology, all of the lines 220 00:28:20.280 --> 00:28:34.229 David Sloan: in the model, the user agent, the authoring tool, the content, you know, that classic W3C shared model of accessibility responsibility. All those lines seem to be getting really blurred. 221 00:28:34.470 --> 00:28:35.230 David Sloan: with… 222 00:28:35.230 --> 00:28:35.880 Ricky Onsman: Yeah. 223 00:28:35.880 --> 00:28:38.539 David Sloan: evolution of AI, the way that we kind of 224 00:28:38.950 --> 00:28:44.559 David Sloan: compartmentalize who's responsible for what in the accessibility world. It seems to be. 225 00:28:44.970 --> 00:28:50.770 David Sloan: Getting upended, and who knows where it's gonna be in 2, 3, 10 years' time. 226 00:28:51.580 --> 00:29:03.429 Ricky Onsman: Yeah, there are risks associated with that that we should be aware of. We don't want to get to the point where artificial intelligence is compensating for bad code and bad markup. 227 00:29:03.600 --> 00:29:20.580 Ricky Onsman: Right. So it… there's still… we have to make sure that the web content, digital experiences are presented in a way that meet… that artificial intelligence can present to a user in a meaningful and accessible way. 228 00:29:20.820 --> 00:29:25.310 Ricky Onsman: So, you know, getting rid of… 229 00:29:26.420 --> 00:29:42.230 Ricky Onsman: code spaghetti, is still… in fact, becomes more important than before. So, you know, there is still going to be a role for people to, make sure that they actually write code in a accessible way. 230 00:29:43.380 --> 00:29:51.580 Ricky Onsman: And it's… it's a… that… that is a considerable risk, because we've already seen how browsers will compensate for poor markup. 231 00:29:51.800 --> 00:30:02.770 Ricky Onsman: And it's meant in a positive way of, look, okay, they haven't coded that very well, that's okay, I'll ignore that, and I'll present it as if it had been coded properly. 232 00:30:02.940 --> 00:30:06.609 Ricky Onsman: But that doesn't create a better digital world. 233 00:30:06.900 --> 00:30:16.520 Ricky Onsman: And we've got to make sure that we actually do it correctly in the first place, so that artificial intelligence does it correctly for the individual user. 234 00:30:17.180 --> 00:30:24.070 David Sloan: Yeah, and I guess, ironically, the whole, you know, back in the day when we were using search engine optimization as a… 235 00:30:24.240 --> 00:30:32.580 David Sloan: you know, the SEO is very closely aligned with accessibility best practices, and now we're thinking of AI, or Gen AI, 236 00:30:32.780 --> 00:30:44.799 David Sloan: Optimization is very closely aligned with accessibility best practices when you can make something machine-readable, whether the machine is an assistive technology or a generative AI. 237 00:30:45.230 --> 00:30:49.090 David Sloan: fuel, yeah, it's… it's all… 238 00:30:49.340 --> 00:30:56.590 David Sloan: It's all kind of helping… hopefully helping us all move forward in the right direction. 239 00:30:57.250 --> 00:30:57.810 Ricky Onsman: Yeah. 240 00:30:58.240 --> 00:31:09.789 Ricky Onsman: One thing… That's also a… that's a thing with, like, you know, in the early days of search engine optimization, the techniques that people used to optimize their websites for search 241 00:31:09.910 --> 00:31:20.710 Ricky Onsman: were not all positive. Keyword stuffing. Putting white text on white backgrounds so that machines would read it, but humans wouldn't see it, right? 242 00:31:20.710 --> 00:31:21.130 David Sloan: some humor. 243 00:31:21.130 --> 00:31:21.940 Ricky Onsman: liking it. 244 00:31:22.220 --> 00:31:24.879 Ricky Onsman: Exactly, it's, it's, it's… 245 00:31:25.030 --> 00:31:32.150 Ricky Onsman: It's essential that we kind of think through what we're doing here, and how it might be abused. 246 00:31:32.700 --> 00:31:34.670 David Sloan: Yep, yep, absolutely. 247 00:31:34.890 --> 00:31:36.339 David Sloan: So one thing I wanted to… 248 00:31:36.630 --> 00:31:43.419 David Sloan: ask about your, you know, as you've been kind of surveying and researching into evolution of AI to support 249 00:31:43.670 --> 00:31:45.419 David Sloan: Content creation. 250 00:31:45.880 --> 00:31:55.350 David Sloan: You know, one… clear opportunity, and also a threat to accessibility, is accessibility of AI tools, 251 00:31:55.760 --> 00:31:59.360 David Sloan: You… what evidence do we have that they… 252 00:31:59.670 --> 00:32:04.540 David Sloan: user interface of generative AI tools and the agents. 253 00:32:05.250 --> 00:32:24.920 David Sloan: is getting easier to use by people with disabilities. I mean, I've heard kind of conflicting evidence that tools are more or less accessible to people with disabilities, and maybe if you're comfortable using a command line interface and you're a screen reader user, then everything's good, but again, that also requires a level of 254 00:32:25.240 --> 00:32:33.180 David Sloan: technical expertise that may… maybe can't be assumed by everybody. So, have you… have you seen any trends there in terms of accessibility of the 255 00:32:33.430 --> 00:32:35.139 David Sloan: Gen AI tools. 256 00:32:36.030 --> 00:32:40.469 Ricky Onsman: Yeah, and I think that that works on a number of levels. 257 00:32:40.990 --> 00:32:47.100 Ricky Onsman: One… one aspect is just the… the simple usability of the tool with the user interface. 258 00:32:47.200 --> 00:32:49.619 Ricky Onsman: There's kind of… it's… 259 00:32:50.280 --> 00:33:07.359 Ricky Onsman: it's, again, it's a… it's two sides of the coin, because a lot of people with disabilities have become used to using assistive technologies, that people who don't have disabilities, for them it's a complete mystery, you know? Using a screen reader is like… and… and I have… I have… 260 00:33:07.570 --> 00:33:23.550 Ricky Onsman: A used a screen reader myself quite a lot, but I've seen people who are very adept at using screen readers, and they, you know, operate on a completely different level. So there's a level of comfort with technology that assists them in their daily tasks. 261 00:33:23.750 --> 00:33:29.560 Ricky Onsman: that… Coincides with the way that you might use assistive technology. 262 00:33:29.790 --> 00:33:37.080 Ricky Onsman: That might use artificial intelligence. So on that level, there is… there is a kind of a… 263 00:33:37.240 --> 00:33:43.030 Ricky Onsman: A prospect there for people who are comfortable with using what you might call unusual tools. 264 00:33:43.520 --> 00:33:46.279 Ricky Onsman: At the same time. 265 00:33:46.410 --> 00:33:57.279 Ricky Onsman: I think when artificial intelligence kind of broke through the, what you might call the conversational banner, it became a lot more usable by people in general. 266 00:33:57.440 --> 00:34:01.559 Ricky Onsman: There was very little thought given to accessibility at that point. 267 00:34:01.870 --> 00:34:13.890 Ricky Onsman: I think things that are happening now are influencing that, that if you are going to produce a tool that is meant to be used by everybody, it has to be designed to be used by everybody. 268 00:34:14.489 --> 00:34:23.069 Ricky Onsman: And there are efforts in that direction. I think the companies that create the artificial intelligence tools are becoming aware that 269 00:34:23.139 --> 00:34:36.350 Ricky Onsman: As with everything else, there is a market for people with disability, and you don't want to ignore that. There is a cultural imperative, there is a social need to make those tools accessible. 270 00:34:36.760 --> 00:34:42.789 Ricky Onsman: That's a… an evolutionary thing. And it takes pressure to make that happen. 271 00:34:43.010 --> 00:34:46.230 Ricky Onsman: The efforts of people like, 272 00:34:46.510 --> 00:35:05.050 Ricky Onsman: Joe Devon in setting benchmarks for how AI tools approach accessibility. That's going to be very, very important. The good thing is that most of those companies creating the AI tools are open to that. They're not deliberately trying to create inaccessible tools. 273 00:35:05.090 --> 00:35:09.519 Ricky Onsman: They want it to be accessible, they just haven't given it a lot of thought up until now. 274 00:35:09.690 --> 00:35:16.119 Ricky Onsman: So I see that, that as being, an area of, of progress. 275 00:35:16.260 --> 00:35:22.319 Ricky Onsman: There are some things that are built in that make that process a little easier. 276 00:35:22.430 --> 00:35:41.460 Ricky Onsman: The fact that you can use voice access to use artificial intelligence is a very positive thing, and it's something that actually moves the dial forward for people with physical disabilities who can't use their hands, you know, can't use a keyboard at all, can't use a mouse, so use voice access. 277 00:35:41.620 --> 00:35:44.150 Ricky Onsman: That's tremendously empowering. 278 00:35:44.380 --> 00:35:56.730 Ricky Onsman: Having said that, you've then got to make sure that the AI tools don't rely on voice alone, because then people who have voice impediments can't use the tools. 279 00:35:57.080 --> 00:36:04.910 Ricky Onsman: So it has to be something that is customizable to meet the needs of individual people. And that's… we're a little way off that yet. 280 00:36:05.080 --> 00:36:19.600 Ricky Onsman: I think as we get to the things we were talking about with browsers being AI-influenced and customizable by individuals, that will influence a lot of the more general AI tools that are available. 281 00:36:19.660 --> 00:36:28.310 Ricky Onsman: And it will become… there's no reason why artificial intelligence tools can't be customizable to individual needs. 282 00:36:28.470 --> 00:36:36.569 Ricky Onsman: But that's a process, and it's something that people in our industry need to exert influence on that to make it happen. 283 00:36:36.790 --> 00:36:53.590 David Sloan: Absolutely, yeah, the accessibility field is an important role to play there. So, we're approaching end of time. I'm going to ask you a couple of very quick questions with quick answers, see how you respond. Firstly, do you see the 284 00:36:53.780 --> 00:37:04.170 David Sloan: trend for… around AI. Are people moving from skeptic to enthusiast, or vice versa, or do you see a kind of 285 00:37:04.950 --> 00:37:07.150 David Sloan: Sort of stability in terms of 286 00:37:07.730 --> 00:37:16.530 David Sloan: enthusiasm. What, you know, do you see any change in how people line up to how AI can help accessibility efforts? 287 00:37:18.090 --> 00:37:33.850 Ricky Onsman: Okay, quick answer, yes, I do, but it depends on what form of artificial intelligence you're talking about. For things like translation, absolutely, that's… that's a positive move. For things like creating plain 288 00:37:33.880 --> 00:37:47.070 Ricky Onsman: easy-to-read language versions of complicated information? Yes, absolutely. We can see the positive sides of artificial intelligence in making things easier to understand. 289 00:37:47.480 --> 00:37:54.780 Ricky Onsman: In some other areas, where it oversteps the boundary from 290 00:37:56.560 --> 00:38:09.840 Ricky Onsman: a transformative mechanism to something that generates something from nothing, there is a long way to go, and I think there is a level of distrust there that is 291 00:38:10.410 --> 00:38:18.730 Ricky Onsman: justified, quite frankly. So I think it depends on what kind of artificial intelligence you're talking about. 292 00:38:19.090 --> 00:38:19.690 David Sloan: Right. 293 00:38:20.160 --> 00:38:29.039 David Sloan: We'll arrange another conversation in 6 months and see if that answer has changed. I bet it has to some extent. 294 00:38:29.240 --> 00:38:45.769 David Sloan: And last question, just one piece of advice you'd give to organizations who are looking at ways to use AI to help with their accessibility efforts, rather than hinder those efforts? What would your one piece of advice be to organizations? 295 00:38:46.780 --> 00:38:51.560 Ricky Onsman: I think it would be… Use AI as a tool 296 00:38:51.800 --> 00:39:01.550 Ricky Onsman: to assist human processes. Don't rely on it as something that replaces human functionality, particularly in regard to digital accessibility. 297 00:39:01.660 --> 00:39:06.300 Ricky Onsman: Use it as something that helps, not something that replaces human effort. 298 00:39:07.830 --> 00:39:11.670 David Sloan: Yeah, no, well said. I think that's great advice. 299 00:39:11.800 --> 00:39:26.690 David Sloan: Well, I think we've… we'll stop it there, but I know that we could carry this conversation on for another 2-3 hours, but the good news is I'm sure we'll find more time to catch up later in the year and see how things have changed. 300 00:39:26.980 --> 00:39:32.679 David Sloan: Thank you, Ricky, so much for sharing your perspective today. That was a great conversation, and… 301 00:39:32.790 --> 00:39:40.129 David Sloan: like I say, I know that AI and its influence on accessibility efforts is only going to change and evolve over time. 302 00:39:40.760 --> 00:39:52.200 David Sloan: So, well, now you know the state of accessibility. I'm David Sloan, thanking Ricky Odinsman, and reminding you that the state of accessibility is always changing, so please help us effect change.