---
id: "workshop:invisible-barriers"
type: "InnovateUS workshop"
title: "Invisible Barriers: Disability Rights and Public Sector AI"
url: "https://innovate-us.org/invisible-barriers"
date: "2026-06-25"
duration: "60 minutes"
series: "Elephants in the AI Room"
instructors: ["Ariana Aboulafia","Jay Wyant","Aaron Page","Marie Cohan"]
moderators: ["Deborah Stine"]
recording: "https://www.youtube.com/watch?v=W3yNzJ-UDW8"
transcript: "automatic, from the recording"
source: "InnovateUS (innovate-us.org)"
licence: "CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/), as stated on innovate-us.org"
---

# Invisible Barriers: Disability Rights and Public Sector AI

InnovateUS workshop · 25 June 2026 · 60 minutes

As governments adopt artificial intelligence to improve efficiency and service delivery, disability rights and accessibility are increasingly becoming central governance concerns rather than niche technical issues. Yet many AI systems are still built on datasets, assumptions, and design practices that fail to account for the diverse ways people interact with technology, including through screen readers, voice navigation, captioning, alternative input devices, and other assistive technologies. These invisible barriers can unintentionally exclude people with disabilities from public services, employment opportunities, education systems, healthcare access, and civic participation.

This workshop moves beyond high-level discussions of responsible AI to focus on practical public-sector challenges and decisions. Participants will examine how AI-enabled systems can either improve or undermine accessibility depending on how they are designed, procured, implemented, and monitored. Drawing on emerging public-sector guidance, accessibility initiatives, and disability-rights research, the session explores practical strategies agencies can use to build more accessible, equitable, and trustworthy AI-enabled services.

## Learning goals

- Provide an overview of the challenges and opportunities that people with disabilities face when using generative AI at the desktop level and its applications in state and local governments

- Identify organizational and operational challenges governments face, including workforce capacity, procurement constraints, vendor oversight, legacy systems, accessibility testing, DOJ Title II digital accessibility requirements and WCAG standards

- Explore practical ways AI may strengthen accessibility, such as improving compatibility with screen readers, supporting document remediation, generating captions and alt text, enhancing navigation, and assisting communication now and in the future

## Speakers

- **Ariana Aboulafia**, Disability Rights Advocate and Project Lead, Center for Democracy & Technology
- **Jay Wyant**, Chief Information Accessibility Officer (CIAO), State of Minnesota
- **Aaron Page**, Vice President Accessibility, Allyant
- **Marie Cohan**, Statewide Digital Accessibility Officer, Texas Department of Information Resources

## Moderators

- **Deborah Stine**, Senior Fellow, Innovate US

Part of the series [Elephants in the AI Room](https://innovate-us.org/workshop-series/elephants-in-the-ai-room).

Recording: https://www.youtube.com/watch?v=W3yNzJ-UDW8

## Transcript

Automatic transcript of the recording: names and terms can be misheard, and the times can be a few seconds off. Each time links to that moment of the video. The transcript alone: https://innovate-us.org/invisible-barriers.txt (plain text) and https://innovate-us.org/invisible-barriers.json (every segment with its start and end second).

**Deborah Stine**

[0:00](https://youtu.be/W3yNzJ-UDW8?t=0) And now to our very interesting topic for today. The Elephants in the AI Room is a three-part series that addresses the most frequently requested and most difficult questions about responsible AI in the public sector. Rather than focusing on tools or use cases, this series addresses the systemic issues governments must confront as AI becomes embedded in public institutions. So far, we've discussed workforce displacement and energy and environmental implications of AI data centers. About 300 people attended each of those sessions, illustrating the level of interest in this topic, and we are hoping for the same today. We will be discussing invisible barriers, disability rights, and public sector AI. We'll be discussing the topic in general today, and we are planning to offer more in-depth workshops in 2027. Our speakers today are Ariana Aboulafia, who's a disability rights advocate and project lead for the Center for Democracy and Technology. Marie Cohen, who is statewide digital accessibility officer for the Texas Department of Information Resources. Aaron Page, who is vice president of accessibility for Alliant. And Jay Wyant, who is chief information accessibility officer for the state of Minnesota. Our session will begin with each of our speakers' perspectives on the topic.

[1:34](https://youtu.be/W3yNzJ-UDW8?t=94) I'll ask them a few questions, and then we'll answer questions from the audience that are posted in the Q&A box. So we have a lot to do in this hour. With that, we'll go to our speakers, and Jay, I'm going to turn it over to you.

**Jay Wyant**

[1:51](https://youtu.be/W3yNzJ-UDW8?t=111) Hello, everyone. This is Jay Wyant. Let me share my screen real quick, if I can remember how to do that. And that seems to have gone away. More options. No Q&A.

**Deborah Stine**

[2:16](https://youtu.be/W3yNzJ-UDW8?t=136) Okay, so Jay, like if you go to the more box.

**Jay Wyant**

[2:22](https://youtu.be/W3yNzJ-UDW8?t=142) I am. And I've got breakout, poll, Q&A, resources, doc, note, and whiteboard.

**Deborah Stine**

[2:28](https://youtu.be/W3yNzJ-UDW8?t=148) Okay. Okay, so Christine, why don't you see if you can fix the problem for Jay? And while we're doing that, let's go to Marie.

**Marie Cohan**

[2:39](https://youtu.be/W3yNzJ-UDW8?t=159) Thank you, Debra. My overall perspective on AI and public sector and how it's impacting people with disabilities is it's a double-sided thing, right? There's a lot of benefit that's coming. You know, public sector can do a lot more efficiently by embedding AI into its digital government services. However, we do need to remember that AI is not a cure-all, fix-all. It's not the one and done. So we need to be aware of how it can also impact our users negatively if we overlook something or don't have a human in the loop.

**Deborah Stine**

[3:20](https://youtu.be/W3yNzJ-UDW8?t=200) Okay, great. Thank you. Jay, are you? Yeah. Okay, great.

**Jay Wyant**

[3:28](https://youtu.be/W3yNzJ-UDW8?t=208) I'm ready to go. Okay, sorry about that. I couldn't see very well for a minute. So I just have a couple of slides I'm going to share just to kind of throw out a few ideas and things about the inter-assessment of AI and people with disabilities. Let me see. I'm having other issues now. There we go, I think. So first of all, wow, why am I having so much trouble here? Okay. First of all, AI and people with disabilities, there are two broad impacts with AI in focus, like Marie said. One is AI as a tool to improve efficiency and service delivery, which is what we said in the summary of the presentation. And the other is AI as a tool to improve existing technology for people with disabilities. I think we'll talk about both of those today. Secondly, AI, and then A1 and Y, the numeronium for accessibility. So I say accessibility, A1 and Y, versus or and AI. They can work together or they can oppose each other.

[4:29](https://youtu.be/W3yNzJ-UDW8?t=269) I'll remind everybody that accessibility is a tool, it's not a solution. You use accessibility as part of a profession. You use accessibility to create principles, best practices, and guidelines. And by applying accessibility to your work, it helps narrow down the variables that you have to deal with when you're creating, making things more usable, or applying human-centered design into your work. But it itself is not a solution. And I also want to add to that, that no one really sets out to build inaccessible technology. I would hope not. But we all know in the accessibility field that accessible technology performs better, it reduces rework, and it's more future-proof. So why does it happen? Why is there so much inaccessible technology out there? Well, one answer might be our tech blows in Silicon Valley with a favorite phrase, move fast and embrace things. As we all know, AI is now moving even faster and continues to accelerate. As shown in the WebAIM, they do a report every year where they do a quick accessibility survey of a million websites. And this year, they noted in the report that there was a 10% jump in errors and a 22% jump in page simplicity. And they said that they thought this being potentially due to an increase in byte coding and other AI practices. So if AI is moving so fast, can we use AI to reverse the train and move fast in the other direction? So one theory would be to maybe use what they call model contest protocol to build an agent that helps developers create accessible code. And there's a company working on that now. We will have to see how they're going to do with that. Just to turn it on, I have been banging the drum about AI bio for years now. And people are finally catching on. And what she talked about is the Pareto principle with the 80-20 rule in which people say, hey, if you spend 20% of the effort,

[6:33](https://youtu.be/W3yNzJ-UDW8?t=393) you should get 80% of the population. That's good enough. And so what happened is she showed the scatterplot diagram here. Great. She showed 80% of the population. And things don't work so well for the next 10%. And things tend to be horribly wrong for the last 10%. And the problem is that last 10% tend to be people with disabilities. And so the 80-20 rule really does a disservice to people with disabilities. And AI is accelerating that trend. So why not invert the 80-20 principle and use AI and AI technology to actually focus on that 20% and thereby create innovation for the 80%? Here's an example. We all know that Alexa, Siri, and Google were built by technology and really created for blind people. And suddenly, hey, everybody else wants to do that. Or text messaging. Just folks like me. We were the ones banging our fingers and thumbs on those small phones. And everybody else said, so silly, deaf people. Actually, no. Guess what? And finally, Octorips were created for people with arthritis.

[7:42](https://youtu.be/W3yNzJ-UDW8?t=462) And now, they're the most popular testing tool ever. So I'm going to wrap up with, I'm going to post some links in the chat regarding the State of Minnesota AI program that you can follow up with later. And that's my brief summary of a couple of things that came to

**Deborah Stine**

[7:57](https://youtu.be/W3yNzJ-UDW8?t=477) my mind today. Thank you. Great. Thank you so much, Jay. Very interesting. Ariana, let's go to you.

**Ariana Aboulafia**

[8:07](https://youtu.be/W3yNzJ-UDW8?t=487) Sure. Hi, everyone. So my name is Ariana Abulafia. I lead the Disability Rights and Tech Policy Project at the Center for Democracy and Technology. CDT is a tech policy organization that thinks about tech policy from the perspective of civil rights and civil liberties. My work specifically focuses on the ways in which AI and algorithmic tools and other emerging technologies impact people with disabilities. I'm an attorney by training. I've been a disability advocate for over 10 years now. And my work has pretty much always focused on multiply marginalized people with disabilities. And specifically, sort of how, again, disabled people interact with not only AI tools, but also with the systems that these tools are deployed within, right? Something I say a lot is that AI is not good, it's not bad, but it is also never neutral, right? And so what that means is that the tools themselves have no sort of moral coding.

[9:13](https://youtu.be/W3yNzJ-UDW8?t=553) But when they are deployed into systems that can already be potentially discriminatory towards people with disabilities, they can serve as force multipliers and they can make things worse. There are, of course, many, many uses of AI for accessibility as well. And so kind of the, like, North Star behind a lot of my work is thinking about policy solutions, recommendations for activists and advocates, recommendations for policymakers, for developers, for companies, pretty much anyone who plays a role in the AI ecosystem of how to make it so that these technologies that people with disabilities can maximize the benefits and minimize or mitigate the harms. That has sort of always been the guiding light of the work that I do. And I think that applies very much in the public sector sort of use of AI and its impact on disabled people as well.

**Deborah Stine**

[10:12](https://youtu.be/W3yNzJ-UDW8?t=612) I'll turn it back to you, Debra. Okay, great. Thank you very much. Aaron, what's your perspective?

**Aaron Page**

[10:18](https://youtu.be/W3yNzJ-UDW8?t=618) Awesome. Thank you. Thanks, Debra. Hi, everybody. My name is Aaron Page. I am Vice President of Accessibility at Alliant. If you are not familiar with Alliant, the brand has only been around since 2022. We're actually a combination of three legacy accessibility companies, T-Base Communications, who does alternative formats, Common Look, who does electronic documents and document remediation, and Accessible 360, who focuses on web and digital. At Alliant, I run our internal accessibility program, which we call Live Our Values. I do a lot of accessibility. I call it evangelism, talking webinars, conferences, presentations like this. And I run what we call our accessibility claims team. It's our team of auditors who helps customers who've been targeted by a lawsuit, demand letter, or other form of legal action. So in terms of AI, you know, to just start off with, as somebody who is blind, you know, the possibilities and the potential of AI as an assistive technology is just incredible. You might be able to tell I'm wearing a pair of MetaGlasses. I just got these a few days ago, and they're already changing my life, just in terms of matching socks, reading the labels on letters, and things like that. So AI as an assistive technology itself, I think, is an immensely powerful. For government entities, I think AI really has the potential to raise the baseline, right? You can have much better, much improved automated remediation. You can have much better automated accessibility scanning and maintenance, right? So if your average PDF is 20% compliant, we can raise that, you know, potentially raise that floor with automation to 60% to 80%, maybe, right? So it's a great way to raise the floor.

[12:10](https://youtu.be/W3yNzJ-UDW8?t=730) You know, what concerns me about AI in this space, and I'm sure we'll talk more about this when we get into the questions, is really just the overpromising, right? People kind of hope that it will solve all of the problems related to accessibility, and we're just not there yet, right? There is no AI that is going to make a website or make a document 100% compliant, right? I think that confusion in the marketplace is really tough for entities in the Title II sector right now. With that, I'll pass it back to you, Deborah. Okay, great. Thank you so much. So I'm

**Deborah Stine**

[12:46](https://youtu.be/W3yNzJ-UDW8?t=766) going to ask everybody a few questions, and so we can really kind of get into the topic. So Marie, let's go back to you to start off answering these questions, which is, what accessibility challenge related to AI and government concerns you the most right now?

**Marie Cohan**

[13:08](https://youtu.be/W3yNzJ-UDW8?t=788) You know, one that would concern me is, if we're using AI in, again, our digital government services, and one of our, you know, citizens accesses that agency, or whether it's higher ed, think like chatbot or something like that, we think we've set it up correctly. We've built it to serve the end user. However, that person may encounter a problem, like a screen reader user may not be able to actually read a table or a graphic that we've created there, and we have no way of knowing that that is not actually working for them. So what we're trying to create to serve our users is actually creating that barrier for them, and how do we create a feedback loop so that our citizens can contact us and let us know that they are having an issue with whether it's a bot or, you know, some kind of text translation? How do we create that feedback loop? I think that is really critical as we're rolling out some of these new services and using

**Deborah Stine**

[14:13](https://youtu.be/W3yNzJ-UDW8?t=853) AI technology to do that. Okay, great. Ariana, what about you? What do you think are the biggest accessibility challenges on AI in government?

**Ariana Aboulafia**

[14:27](https://youtu.be/W3yNzJ-UDW8?t=867) Yeah, so the first thing I'll say is that a big part of the work that I do is actually thinking beyond accessibility. It's sort of a controversial statement, but sometimes I talk about accessibility as a floor, right? That the ability for disabled people to access tools should not be thought of as sort of the holy grail of it, and instead we should be thinking about that as kind of the bare minimum, and then it opens the door to then think about all the other ways that these tools can impact people with disabilities, including ways that might be discriminatory or unlawful. Awful or something like that. But regarding accessibility, I mean, in general, for those of y'all who are following the web access rule and the saga therein under Title II of the Americans with Disabilities Act, we know that just basic web accessibility, web and app accessibility, we are pretty far behind in where a lot of folks would want to be on Title II entities, web and app accessibility, right? And of course, the pushback of the deadline on that rule recently released for the Department of Justice isn't necessarily going to help bring us into compliance or enhance accessibility in the short term, right? And so when we're thinking about like, what is a good model of AI accessibility look like? I think it's important to sort of think, take a step back and think about how far behind we even are just on the internet kind of in general.

[15:56](https://youtu.be/W3yNzJ-UDW8?t=956) You know, the other thing I'll say kind of like, as a threshold matter, right, is that CDT did a study that we published, I believe it was last year, maybe in the year before, focused on employment hiring tools. And these are used both in the private sector and to some degree in the public sector as well. And we found that some of those tools specifically gamified assessments. So like these little tools that look like, they look like computer games, and sometimes they're used allegedly to test things like risk averseness or things that are questionably related to job functions. But we found in our research that those tools posed not only potentially discriminatory effects, but that they were just not accessible with things like screen readers. You know, and then I'll say also just from the public sector perspective too, I think there are real questions as to Section 508 accessibility and AI that maybe we can talk about later too, and sort of what the best way is to square that all together. But yeah, I mean, I think I'll say again, like, knowing that it's a bit controversial, yes, we need to think about access with AI. But we can't stop there. Because it's not enough for any other group of people, people without disabilities to say, well, they can use it. And so we don't have to think about it anymore. Right? Like, no, when we're talking about AI outside the context of disability, we're talking about the ways in which it impacts the whole person. And so I think we need to extend those conversations to disability and think about accessibility, yes, as something that matters, people with disabilities, particularly who want to use AI tools, or who are interacting with these tools and systems wherein they may not have a choice, right? Like in healthcare or transportation or something like that. The accessibility is necessary, but it's not sufficient. And I'll

**Deborah Stine**

[17:51](https://youtu.be/W3yNzJ-UDW8?t=1071) turn it back to you. Okay, great. Aaron, what about you? What do you think are the biggest

**Aaron Page**

[17:56](https://youtu.be/W3yNzJ-UDW8?t=1076) accessibility challenges in AI in government? Yeah, I think, you know, there's two problems or two challenges that come to my mind. And one, one is just, as more entities are adopting these AI powered tools, knowing how accessible is the content that these tools can generate, right? That alone, I think is a very difficult question. We're seeing more and more organizations using AI tools to say, build their websites, right? Build and maintain their websites. Well, how accessible is the content that those web that the AI tool is generating and putting up on the website? You know, a good resource that I would highly recommend focusing, take a look at is amac.ai. Christine, if you don't mind dropping that link in the chat, that is a link to essentially a leaderboard that compares the accessibility of web content generated by different AI tools and different AI models, right? It's at least a good kind of baseline to know, okay, if we're, if we're adopting GPT 5.0 at our organization, how does it do in terms of creating accessible content, right? So I think that's one, one challenge is just knowing how accessible the content is that these AI tools generate. And then I think the other, the other big challenge is, you know, we're, we're having AI really come to fruition, seeing much wider adoption, right at the same time that the title two regulations are coming into effect, right? There is the IFR that is pushed out the deadlines, but the timing, you know, they're happening really at the same time. And so what's happening is, is a lot of organizations are really looking to AI to say, solve this problem for me, right? And the hope is just that, you know, by leveraging some AI product, implementing it into a workflow, it will handle everything you need in terms of accessibility. And, you know, while I definitely agree that AI tools have increased the power of what can be fixed and

[19:59](https://youtu.be/W3yNzJ-UDW8?t=1199) identified automatically, you know, a simple example is on a website, right? A status message that doesn't read out to a screen reader, a traditional web scanner would not find that, but an AI powered one can, right? So it has the potential to improve what is found, improve the baseline of accessibility, but it doesn't have the ability to just determine whether or not somebody with a disability can use that resource, right? It cannot achieve 100% WCAG compliance automatically. You still have to involve persons with disabilities, right? And so just understanding the limitations of these tools, given the high number of tools that are coming to market and the kind of marketing around them as, you know, embed this code and it solves all your problems. It's very hard for organizations to actually know what risk are they taking depending on what solution that they go with, right? And so from my point of view, leverage the AI where you can, right? The big backlog, but understand that, you know, key workflows, the forms people need to fill out, right? These are things that still need to be manually remediated or manually looked at, tested by persons with disabilities, right? Accessibility is about people and AI is not going to replace the need for people in that chain. So that is, those are the two big challenges that I see right now.

**Deborah Stine**

[21:25](https://youtu.be/W3yNzJ-UDW8?t=1285) Debra, I'll hand it back to you. Okay. So Jay, what about you? What do you think are the biggest accessibility challenges in AI in government today? Well, this is Jay speaking. Marie, Eliane, and Aaron

**Jay Wyant**

[21:37](https://youtu.be/W3yNzJ-UDW8?t=1297) covered a lot of the stuff that I think about. I want to bring up, I will point back to that million website survey that WebAIM did. It reminded everybody that much of AI is built on a, we all know, large language model. Those large language models are basically made by hoovering up data, just all that data out there. And 95% of that data they hoover up are inaccessible websites. So large language models are predicated on things that don't work for people with disabilities from the beginning. And if you're just using those standard AI models and the AMAP site kind of illustrates that. So it's basically that alone. It's going to create a lot of problems. Then when you move from just general AI to what we call agentic coding, agentic coding is where the AI model uses a large language model to kind of think and how to make that code happen. And right now, there's a lot of hallucinations happening with agentic coding. So again, that's why I mentioned the content protocol to kind of, there's a way to maybe use AI to make that better, but nobody's doing that.

[22:45](https://youtu.be/W3yNzJ-UDW8?t=1365) And then I'll remind people, Aaron referred to the overlays. So the big overlay trend in the last five years or so where companies were saying, like Aaron said, here's a few lines of code. This is your website. It's magic. It'll work great. And because people need, they see the Title II issue. They see these problems. They want to solve it quickly. They were fixing quickly and overlays are a massive bullet. AI is now the new massive bullet. And just like overlay, AI alone will be very problematic as a solution. So that really bothers me a lot. That's why we have to talk about how we actually flip AI around either with a content protocol or something else. Finally, I just want to note that, what was I going to say, that we can then think about ARIA. So a lot of websites right now use way too much ARIA. A-R-I-A, Accessible Risk Interface, something, I forget what it is. It's a very popular tool in a lot of platforms to make websites work. And the first rule of ARIA is to not use ARIA. The problem is websites use too much ARIA. And so I can not only imagine that most AI-based websites that are, or AI-based PDF solutions, for example, converted PDF, would use ARIA to fix things. And that would not ensure accessibility. So to kind of reiterate what everybody else is saying, kind of just a different word. Thank you. Okay, great. Thanks.

**Deborah Stine**

[24:16](https://youtu.be/W3yNzJ-UDW8?t=1456) Okay, so the next question is, well, first, let me remind everybody that, you know, when I finish with these few introductory questions, we're going to open it up to audience questions. So make sure you put those in the Q&A box. And just always remember that the first ones in are the ones that I see first as we're prioritizing questions. So be sure to get those in. So Aaron, let's go with you for the second question here, which is, what do you think are the most promising current uses of AI to improve accessibility in government settings

**Aaron Page**

[24:48](https://youtu.be/W3yNzJ-UDW8?t=1488) and services? I apologize. My screen reader was reading out the chat to me when you said the question. Could you read it to me one more time? Okay, yes. And I actually was going to Ariana

**Deborah Stine**

[25:00](https://youtu.be/W3yNzJ-UDW8?t=1500) first, but I will go ahead and go to you. Oh, you said Ariana. I thought you said Aaron. Apologies again. Okay. Okay. So but I tell you what, Ariana, why don't you go ahead and answer it and then I'll go to Aaron and I'll repeat the question. But it's basically the most promising

**Ariana Aboulafia**

[25:13](https://youtu.be/W3yNzJ-UDW8?t=1513) current uses of AI. Sure. Yeah. This is Ariana. So yeah, I mean, I think that there are promising uses of AI in public service. But, you know, as with anything else, there's just a lot of caveats there, right? And so the example that I give sometimes is like chatbots, right? There is a world in which chatbots, so meaning like large language models, generative AI chatbots can be used for customer service. And they can be, you know, integrated into government services in ways that, you know, could minimize wait times, prevent, as an example, people with disabilities from having to go to a place in person, which a lot of folks with disabilities can't do, right? Standing outside, being in the heat, et cetera. And so that's like a potential use case, right? But there are real caveats that like to do any of this, it has to be done well and right, right?

[26:21](https://youtu.be/W3yNzJ-UDW8?t=1581) And so with the chatbot, as an example, that seems like really straightforward, right? Like, of course, we could integrate this into government services and it would reduce wait times and it would make things faster and I wouldn't have to pay staff as much and whatnot, right? However, there are privacy concerns, first and foremost, right? If you're creating a large language model, what are you doing with the very sensitive data that people, particularly people who may be putting health-related information or, you know, information related to their social security number or their, you know, their tax identity, whatever, right? Into this chatbot, what data privacy, you know, mitigations are you putting into place? Are you doing something that we call data minimization, where you're only collecting the information that you need? Are you doing something that we call purpose limitation, where the information that you are taking, you're only using for the suggested purpose? You know, we have found in one of the surveys that my organization did in the education context, we found that schools that used AI were more likely to report data breaches, right? That's not necessarily a causation, but it's worth noting, right? And so there's privacy concerns and there's, you know, just data governance concerns and cybersecurity concerns. But then there are also just concerns about efficacy, right? Do these chatbots work on, let's say, people with multiple disabilities or people that have whatever their situation is?

[27:55](https://youtu.be/W3yNzJ-UDW8?t=1675) It might be a little less common or popular, right? And the reason I bring that up is because generally speaking, the way that large language models work is that they create outputs based on inputs, right? And so when their outputs are under-inclusive or non-inclusive, that's typically because the training data, right? Like the body of data that they're pulling from to create the output is also under-inclusive, right? And we see this a lot with disability, that the training data used to create these chatbots is not sufficiently inclusive of disability. And I wrote a report on this that I published two years ago for the ADA anniversary, thinking about like the myriad of reasons why disability data isn't where it should be and how to, you know, get better data for these purposes, but in privacy protective ways, right? And so we have to think about not only the privacy and the cybersecurity, but also like, does this even work for the purpose that I want it to work for? Is my training data sufficiently inclusive? If it's not, how do I make it that way?

[28:57](https://youtu.be/W3yNzJ-UDW8?t=1737) The chatbot, is it accessible, right? Like, which again, like we're talking, this is like the bottom floor of what we should be thinking about, but like, does a chatbot work with a screen radar? Is it providing me with outputs in plain language such that someone with intellectual developmental disabilities would be able to access it, right? And then, you know, and this is more the purview of one of my colleagues who works on multilingual AI, but, you know, we should also always think about whether these tools are able to provide services in multiple languages, right? And there's quite a bit of evidence that shows, particularly with under-resourced languages, that they're not good at that, right? And that, again, it goes back to the training data. And so there's, and I'm seeing these comments in the chat too, and of course, right? Yes. Is it biased? Is it discriminatory? And that also, not for nothing, tends to go back to the training data, right? If you ask a large language model to create an image of a person with a disability, you can see the way in which it can

**Deborah Stine**

[30:00](https://youtu.be/W3yNzJ-UDW8?t=1800) So Erin, what do you think are the most promising current uses of AI to improve accessibility in government services and settings? Oh, you're muted, Erin.

**Aaron Page**

[30:36](https://youtu.be/W3yNzJ-UDW8?t=1836) Sorry about that. Thanks. So I think the most promising use of AI right now is to improve automated accessibility scanning. You know, for a very long time, automated accessibility scanning tools, whether they were for documents or for websites, were very limited, right? If an accessibility SME was being direct, being kind of honest, the answer was usually somewhere between a quarter and a third, right, of accessibility violations could be potentially detected by conventional automated accessibility scanners, right? So it left a large, large gap of things that it could not find, right? Things like a button that's not marked up as a button, right? An image that lacks alt text. Traditional scanners could find that image, but it couldn't suggest alternative text for it. Issues like that couldn't be found by those traditional scanners.

[31:37](https://youtu.be/W3yNzJ-UDW8?t=1897) They could only be found by manual accessibility auditing and review. And a lot of these things that used to be dependent on manual accessibility auditing and or remediation now can be detected by these AI-powered tools, right? You know, I apologize, Mr. White, he mentioned the ability to use AI to check content via model context protocol, right? Using that, AI-powered accessibility scanners are able to go in and actually perform the actions that a manual auditor would potentially do to check for accessibility issues, right? You click a button and a dialog box appears. Well, a traditional automated scanner can't click that button to trigger the dialog to check its contents, but an AI-powered one leveraging MCP tools can, right? So it now has the potential to check more content more accurately, identify more issues, right?

[32:42](https://youtu.be/W3yNzJ-UDW8?t=1962) So I think that that is probably the biggest opportunity is to be able to improve just ongoing monitoring and maintenance around accessibility. And as I mentioned before, when it comes to document accessibility, to raise the floor, right? To be able to take, you know, maybe something that is completely inaccessible, an image-only scanned PDF, right? And to be able to take that and turn that into a text-readable PDF, perhaps with some basic headings and some basic alternative text, right? It might not be 100% compliant, it might not be perfect, but it's better than what it was before, right? And so I tend to think about that as raising the floor, but what it doesn't do, again, is guarantee 100% compliance or eliminate the need for the input of persons with disabilities who, in the end, are the ones who have to actually use these documents, right? The tools are fantastic, and they might say this has been validated 100% compliant, but you don't necessarily really know that until somebody with a disability has gone and used it.

[33:46](https://youtu.be/W3yNzJ-UDW8?t=2026) And I've had plenty of experiences where a tool, AI-powered or not, has said something is compliant, and I, as a screen reader user, have found it to be inaccessible or a poor experience, right? You can't necessarily take them at face value. So I'll pass it back to you, Debra. Great. Thanks, Aaron.

**Deborah Stine**

[34:03](https://youtu.be/W3yNzJ-UDW8?t=2043) Jay, what about you? What do you think are the most promising current uses of AI?

**Jay Wyant**

[34:09](https://youtu.be/W3yNzJ-UDW8?t=2049) Sure. Thank you, Jay Bryant. Well, right now, we see one thing that's already happening with the automated captions. So the automated captions have become much, much better over the last five years. We're using it right now with this presentation, but it still has a ways to go because, you know, I have a deaf accent, and it has a hard time with my voice, but it's getting better and better as that goes along. Another one, our young and missing bots. A lot of organizations, like Data Minnesota is one of them, are using what we call retrieval augmented generation, or RAG, bots. And those are bots in which you can authenticate data that you own, data in your system. So, for example, a website, a document that you created, and have the bot just go to that data, and nothing else with the public or whatever. And that's the source of search bots. And then you can offer a search bot on your website, and we're going to just go to that database.

[35:12](https://youtu.be/W3yNzJ-UDW8?t=2112) And so AI is very good at that, and it's getting better. And it's a good way to train that model to have your own unique set of language models, for example. And then another one is AI assistive technologies. Several folks today have already mentioned the term raising the floor. And actually, Raising the Floor is an organization by Greg Vander Heiden, who works at the Trade Center in Wisconsin. And he has been working for years on a tool called Morphic, M-O-R-P-H-I-C. And it's not AI-based, but I can see it becoming that in the future. And the idea of Morphic is that anyone with a disability who has to set up their browser and their tool, whether it's a screen reader or Zoom or whatever, and they have to customize things for themselves. If they walk into a library, they can use Morphic to create their own setup at any browser, on any computer, at any given time. So I can see AI being used in that way, where people with assistive technology need don't have to carry everything with them.

[36:16](https://youtu.be/W3yNzJ-UDW8?t=2176) They can go into any website or any application or any system, and their AI, their assistive technology is available to them. I can see AI doing that in a very big way.

**Deborah Stine**

[36:27](https://youtu.be/W3yNzJ-UDW8?t=2187) Great. Thank you, Jay. So, Marie, what do you think are the most promising current uses of AI to improve accessibility?

**Marie Cohan**

[36:34](https://youtu.be/W3yNzJ-UDW8?t=2194) I think my colleagues here have mentioned a lot of them. Like AI-generated captions and transcripts has been better and better. I like the live transcript feature because that is not only supporting people with a hearing disability, but that can be used for people with a vision disability because the transcript can be read by a screen reader. So that's something that we can provide. You know, we've talked a little bit about, you know, the chat bots. One thing I'd like to say there is government, this is an opportunity for public sector to build its own large language models instead of leveraging like public large language models. I see opportunities there to improve the kind of data that people with disabilities can access because then the bot is just pulling from your own model, your own data. So there's a lot more control for public sector there. And then, you know, building in that process automation through AI that if a user with a disability suddenly hits a wall, you don't have that looping functionality that we all like hate when we get a bot, whether it's on the phone or on our computers.

[37:43](https://youtu.be/W3yNzJ-UDW8?t=2263) You can have it just immediately connect that person with a disability to an actual human being to support them. That solves a lot of, you know, other issues that come with that too that Ariana mentioned like privacy, maybe security. So I like that we have that opportunity. I think we should leverage that more. You know, Erin mentioned automated scanning tools. Those have been super helpful and not just automated scanning, but automated remediation tools. Government loves PDFs. We love our documents, right? Millions and millions of documents in public sector. And then now we've got to make them compliant with Title II. So how do we do that? We're relying heavily on those automated scanning tools and remediation tools. However, government may become too reliant on those. So, you know, there's that caution there to not just rely on the automated tools, but actually go in and do testing. I think Erin mentioned it as well. Testing with people who have a disability to ensure that we are getting that most accurate, most accessible percentage, if you will, as possible.

[38:48](https://youtu.be/W3yNzJ-UDW8?t=2328) You know, and then I think I was going to talk about, you know, graphics. We do have a lot of graphics in our agency because we set policy for the state of Texas. We do the strategic planning. So, you know, helping us with the color contrast, the alt text, the data that accompanies a graphic, AI has been really helpful with that. And lastly, procurement, because I do like to go back to procurement. We need to start there when we're purchasing AI tools or services to build our own AI model. I'm currently leveraging the data that I have collected for the last four years and evaluating products and services and vendors. And I'm going to leverage an AI tool to help me analyze that data. I think there's an opportunity there to speed up that evaluation process to see which vendors are actually providing the most credible documentation around their product accessibility or their services accessibility, like developers.

[39:52](https://youtu.be/W3yNzJ-UDW8?t=2392) I think that could help government save a lot of time when we're going through and vetting those products and services. If we can leverage AI to help us cull through all of the data that we've collected and help us identify faster whether or not that vendor most likely is going to have the best accessible product or service.

**Deborah Stine**

[40:10](https://youtu.be/W3yNzJ-UDW8?t=2410) Great. Thank you, Marie. So we're now going to go to the audience Q&A. So while we're doing that, Christine, why don't you go ahead and post the results from our survey. And we'll do a speed round so that we can cover as many questions from our audience as possible. So, oops. So the first question, let's go ahead and start with Jay on this one. Is there a sound framework on balancing the efficiency versus assistive nature of AI in the context of public impact and equity? Is there a good criterion to share? So Jay, you're muted if you're answering.

**Jay Wyant**

[41:12](https://youtu.be/W3yNzJ-UDW8?t=2472) Yes, I think a good first basis is, just like I say, when people are starting the accessibility program with AI, start with the policy. You need to build an effective policy in your organization so that you actually have certain guidelines and guardrails and other resources to make sure that you know what you're going to do in a meaningful way with AI. So because, again, it's the 80-20 rule in a way. You can try to be efficient, but you need to put that efficiency around to make sure that whatever you're doing is not isolating or marginalizing people with disabilities. If you cannot ensure that, then you really can't search for efficiency. So that's where the policy goes into play. You have a policy, you understand what you're taking on, how you're handling it, and you're focusing more on the efficacy rather than the efficiency. Right now, I think that's a very difficult thing to balance.

**Deborah Stine**

[42:09](https://youtu.be/W3yNzJ-UDW8?t=2529) Great, thanks. So Arianna, if accessibility for 80% of the population is considered inadequate, then what percentage is legally considered fully accessible? For example, I just attended a public hearing that was in person but advertised as fully accessible.

**Ariana Aboulafia**

[42:29](https://youtu.be/W3yNzJ-UDW8?t=2549) I don't know the answer to that. First of all, I'm not an expert in digital accessibility. I am a civil rights lawyer that focuses on, like, discriminatory impacts of AI. I really don't know the answer to that. But I would say that I would be very careful about advertising things as fully accessible because accessibility is, to a certain extent, like a person-by-person thing, right? Different people sometimes need different things to make an event accessible. I say, not as legal advice, but as, like, a best practice to leave things very open for people to say what it is they need. And this is one of the reasons why I am a big proponent of transparency when it comes to the use of AI tools, right? So that folks know what tools are being used so that they can ask for an accommodation.

[43:30](https://youtu.be/W3yNzJ-UDW8?t=2610) But, yeah, I mean, I would say someone in the comments just brought up the accessibility overlays. And, you know, if the FTC accessibility settlement is something that we want to keep tabs on vis-a-vis deceptive business practices, I would be very careful about making claims like that. But I don't know if there is a particular term of art that would put you at a particular percentage that would say, yes, that's when you can or can't make that advertisement. But accessibility is a shifting and growing sort of thing. So, yeah, I do not have a satisfactory answer for that.

**Deborah Stine**

[44:12](https://youtu.be/W3yNzJ-UDW8?t=2652) Okay. So, Marie, another question is, state CIOs hold a massive market-shifting lever through procurement. Are states using their collective purchasing power aggressively enough to reshape the accessibility market? If not, what specific procurement practices would help states send a stronger signal that inaccessible technology is no longer acceptable?

**Marie Cohan**

[44:41](https://youtu.be/W3yNzJ-UDW8?t=2681) Yeah, this is a great question. So I'll just start with the state of Texas. We have done a collective purchasing power. So my agency has all the technology contracts for the state of Texas. And we share those contracts, like meaning you can. Ava DuVernay, Ph.D.: With all public sector in the state of Texas and outside of the state of Texas. We've done that, that hard work of ensuring their security and privacy and accessibility and best pricing and etc. Ava DuVernay, Ph.D.: I've made sure that, you know, all of our master contracts are covering our customers, whether they're buying a product or a development service. Ava DuVernay, Ph.D.: We also leverage the policy driven adoption for accessibility, which is a Nacio sponsored tool J state uses it as well. And that just test the Ava DuVernay, Ph.D.: Accessibility maturity of the responding vendor. So it's not tied to their product or service offering. It's more around how is that vendor employing accessibility within their organization.

[45:46](https://youtu.be/W3yNzJ-UDW8?t=2746) Ava DuVernay, Ph.D.: And that's a that's another lever that helps us gauge whether or not we want to approve that vendor moving forward with their product or service. Ava DuVernay, Ph.D.: As far as, you know, procurement practices again having that language built into the master contract. So we have all of our governance language in there. We have our WCAG standard that we follow. We have Ava DuVernay, Ph.D.: Language in there that states, even though our agency is asking you for this accessibility documentation right now at this point in time to evaluate Ava DuVernay, Ph.D.: Whether or not you get a master contract, be prepared to provide all of this documentation again at any point in time during the life of the contract. Ava DuVernay, Ph.D.: When one of our customers is purchasing off of it because technology changes all the time. So we want to make sure we're getting the most current most accurate data at that time of purchase Ava DuVernay, Ph.D.: So, you know, as far as technology no longer being accessible. I've heard of accessibility warranties. I've heard of Ava DuVernay, Ph.D.: You know, getting accessibility statements from the vendor vendors are not really cool with that. So once you once you set up a statement of work off of the contract that even has all of those safeguards in place.

[46:58](https://youtu.be/W3yNzJ-UDW8?t=2818) Ava DuVernay, Ph.D.: Be really specific about what your accessibility requirements are in in the requirements of that statement of work be very clear Ava DuVernay, Ph.D.: Ensure if it's a development project you're testing for accessibility at every sprint. Don't wait till the very end, make sure it's clear on who's remedied remediating what and be very clear on that care period after you've deployed or implemented, whatever it is that you've purchased

**Deborah Stine**

[47:22](https://youtu.be/W3yNzJ-UDW8?t=2842) Jennifer Hootman-Pickert, Ph.D.: Okay, Jay, you want to chip in on this one too, since you're also a CIO

**Jay Wyant**

[47:26](https://youtu.be/W3yNzJ-UDW8?t=2846) Jay Chowdery, Ph.D.: Briefly, yes. I'm not a CIO Jay Chowdery, Ph.D.: But

**Deborah Stine**

[47:31](https://youtu.be/W3yNzJ-UDW8?t=2851) Jennifer Hootman-Pickert, Ph.D.: I missed a letter there. Sorry.

**Jay Wyant**

[47:33](https://youtu.be/W3yNzJ-UDW8?t=2853) Jay Chowdery, Ph.D.: Yeah, I'm a child. That's okay. Um, we did everything brilliantly and very much. We do have colossal power through NASCO, National Association of State CIOs and NASPO, National Association of State Procurement Officers, which is great. Jay Chowdery, Ph.D.: But unless the company is specifically targeting the government sector, we're really not very big. Jay Chowdery, Ph.D.: Most of the private companies, you know, the Target, the Amazon, they have a much bigger market share than we do. So what we can do is lead the way in terms of what Marie said. Jay Chowdery, Ph.D.: Set the standard, show what we're looking for, hold them to their feet to the fire with accessibility and so forth. But what we really, really need is to persuade all the Jay Chowdery, Ph.D.: private companies out there to also include accessibility in their requirements and what they're asking for. And that is Jay Chowdery, Ph.D.: a big lift. And that's something where we need all the community, everybody else to push their companies and their vendors, anyone else to do it. So it's not just us to have that kind of leverage.

**Deborah Stine**

[48:44](https://youtu.be/W3yNzJ-UDW8?t=2924) Mary Hayhoeser & Scripps, Ph.D.: Okay, great. Thanks. So, Aaron, we talked about how AI can be used to increase access, but how do we ensure things are accurate. Mary Hayhoeser & Scripps, Ph.D.: There's real potential that when an AI system is used to write text or provide any output a person with a disability who utilizes that output may not even be able to judge its accuracy.

**Aaron Page**

[49:06](https://youtu.be/W3yNzJ-UDW8?t=2946) Aaron Mazzrillo, Co-Founder of Aaron Mazzrillo, Ph.D.: That's an interesting question. You know, I guess my my response to it would be, I think that that kind of speaks to, you know, being careful about how you go about using AI tools, right. Aaron Mazzrillo, Co-Founder of Aaron Mazzrillo, Ph.D.: I myself don't usually use an AI tool to generate content from scratch. Right. Instead, what I am doing is I am leveraging the AI tool to have it build what I want. Right. So, for example, Aaron Mazzrillo, Co-Founder of Aaron Mazzrillo, Ph.D.: You might notice from my my talk today here that I can be a little verbose. Right. And so when I write a post on LinkedIn, it tends to be very, very long and very, very wordy. Right. And so Aaron Mazzrillo, Co-Founder of Aaron Mazzrillo, Ph.D.: I like to use AI to take what I have written and refine that into something that is more clear, it's more concise, it's more polished polished and professional. Right. But in the end, it's still Aaron Mazzrillo, Co-Founder of Aaron Mazzrillo, Ph.D.: What I told the AI. I want to say it's not what the AI came up with for me. Right. And so I think Aaron Mazzrillo, Co-Founder of Aaron Mazzrillo, Ph.D.: Using AI in that kind of a way is is a much safer use of it right you're you're running an awful big risk if you just simply say Aaron Mazzrillo, Co-Founder of Aaron Mazzrillo, Ph.D.: You know, AI tool generate a LinkedIn post for me about how to build an accessible carousel.

[50:33](https://youtu.be/W3yNzJ-UDW8?t=3033) Aaron Mazzrillo, Co-Founder of Aaron Mazzrillo, Ph.D.: Right, that that would be very dangerous because you're going to probably get different recommendations from different tools. It won't be consistent. Aaron Mazzrillo, Co-Founder of Aaron Mazzrillo, Ph.D.: Right. And unless you know about carousel accessibility, you might not realize that what it's saying is wrong. Right. And so Aaron Mazzrillo, Co-Founder of Aaron Mazzrillo, Ph.D.: I think it's just important to use it to refine be very careful if you're using it to research. If you do use it to research use actual research modes where it sites. So you can look at the sources of where the information is coming from. Aaron Mazzrillo, Co-Founder of Aaron Mazzrillo, Ph.D.: I think that's the best way to ensure that you have a clear understanding of whatever it is that AI happens to be providing you

**Deborah Stine**

[51:11](https://youtu.be/W3yNzJ-UDW8?t=3071) Okay, great. Thanks. Well, we got through as many questions as we as we could today, but I wanted to give it a chance for our speakers to give one final word so If our participants remembered only one thing from this discussion. What should it be just give me like one sentence. So let's start with you, Ariana.

**Ariana Aboulafia**

[51:33](https://youtu.be/W3yNzJ-UDW8?t=3093) I think one of my colleagues is the one who mentioned this, but that the most important thing when you're thinking about the use of AI and how it impacts people with disabilities is to ask people with disabilities to center people with disabilities in these conversations, because we can tell you

**Deborah Stine**

[51:51](https://youtu.be/W3yNzJ-UDW8?t=3111) Okay, what about you, Aaron. What's one thing to remember from today.

**Aaron Page**

[51:54](https://youtu.be/W3yNzJ-UDW8?t=3114) You know, I would say, use a hybrid approach right and you want to implement processes and solutions that leverage AI leverage automation. But still keep people with disabilities in the loop. Right. If you are trying to go with a solution or a process that is just simply fully automated from start to finish. You're, you're setting yourself up for failure, you know, as my colleague on the panel here mentioned back with overlays, you know, we saw a great example of that then so hybrid approach manual and automated is still the right way to go, even with AI tools, having become as powerful as they are.

**Deborah Stine**

[52:34](https://youtu.be/W3yNzJ-UDW8?t=3154) Okay, Jay. What about you. What's one thing that people should remember from today.

**Jay Wyant**

[52:38](https://youtu.be/W3yNzJ-UDW8?t=3158) Pretty similar to what Eliana and Aaron have said. First, again, I would really encourage people to have a policy for AI and I point everybody to a link to the policy that we have in the state of Minnesota. The second thing is, be sure that whoever manages that AI program, and whoever manages accessibility, talk to each other. So like I work very closely with a person who manages the AI program and that's been fantastic. I've learned a lot from him. He learned a lot from me. And we work together to make sure that the organization is as equipped as it can be to handle that whole AI onslaught.

**Deborah Stine**

[53:16](https://youtu.be/W3yNzJ-UDW8?t=3196) Okay, great. And Marie, what's your final word.

**Marie Cohan**

[53:19](https://youtu.be/W3yNzJ-UDW8?t=3199) So I'm going to do a culmination of what all of my colleagues have said specifically, and I actually wrote it down so I'm going to read it to you all. People with disabilities are often the earliest adopters and most sophisticated users of AI and simultaneously the population most exposed to its failures. Public sector IT leaders are in the rare position of deciding which of those realities their residents experience.

**Deborah Stine**

[53:42](https://youtu.be/W3yNzJ-UDW8?t=3222) Okay, great. Thank you so much. So thank you, Ariana, Marie, Aaron and Jay. I'm sure I and everybody else in our audiences really have a great deal to think about from this very informative session. And thank you to all our participants who took time out of their day to attend this session. If you enjoyed it, this is more to come. So on July 1, we'll begin our series on public contact centers. On July 13, we'll begin our results lab section that I'll actually be moderating about how to conduct continuous evaluation of state and local government programs. Christine is putting information on the workshops and how to register them in chat. And finally, please fill out the post workshop survey and suggest topics you'd like us to cover in the future. We use your feedback to plan future workshop series like the one you participated in today. As I mentioned earlier, this was in response to participant requests and all the chat and questions that came in today I'll use to plan the future activity we'll have in 2027 on this topic.

[54:51](https://youtu.be/W3yNzJ-UDW8?t=3291) So I'm ending the meeting now and I just like to do a final thank you to everyone for speaking and attending our meeting. Thank you so much.

**Aaron Page**

[55:00](https://youtu.be/W3yNzJ-UDW8?t=3300) Thank you all.

* * *

Original page: https://innovate-us.org/invisible-barriers · Licence: CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/), as stated on innovate-us.org
