AccessWorld Podcast, Episode 39: New AFB AI Report on Assistive Tech

Episode Notes

In this episode of AccessWorld, Tony and Aaron welcome Dr. Arielle Silverman and Sarah Malaier from AFB's Public Policy and Research Institute (PPRI). Dr. Silverman is the Director of Research at AFB, and Malaier is the Senior Advisor for Public Policy & Research. They share the latest report released from AFB as part of its series on AI attitudes and perceptions. This report focuses on AI and assistive technology.

To access the report and other reports in the series, visit: afb.org/AIResearch2

To vote on AFB's AI related panels for the South By Southwest Conference this spring in Austin, visit: afb.org/VoteSXSW27

This podcast is powered by Pinecast.

AccessWorld Podcast, Episode 39 Transcript:

Tony Stephens: Hey everybody, in this episode of AccessWorld, we are joined by Dr. Arielle Silverman and Sarah Malaier from the American Foundation for the Blind. And as we're getting ready to drop this episode, we have 48 hours left for the South by Southwest voting that's going on. Arielle, just real quick, we've got what, 3 presentations? I think this is an ad to get people to go out and vote.

Dr. Arielle Silverman: Yeah, so we have 3 presentations that are all about our AI study, and we really want to get on that stage in Austin next March, but 30% of the decision is made by community votes. So we have links to all the presentations and instructions on how to vote at afb.org/VoteSXSW27.

Tony Stephens: And that's afb.org/VoteSXSW27 for South by Southwest '27. Folks can go there, go through the instructions. It's super easy and it helps bring the AI research that you all are about to hear to the minds of those making differences in Austin 2027. Thanks, everybody. And now on with the podcast.

Intro: AFB. You're listening to AccessWorld, a podcast on digital inclusion and accessibility. AccessWorld is a production of AFB Studios at the American Foundation for the Blind. To learn more about AccessWorld, visit us online at afb.org/aw.

Tony Stephens: And welcome everybody back to AccessWorld, a podcast on digital inclusion and accessibility from AFB Studios at the American Foundation for the Blind. I am your co-host, Tony Stephens, and I am joined as always with our fearless editor-in-chief of AccessWorld magazine, a quarterly publication of AFB, Mr. Aaron Preece. Aaron, happy almost end of summer. How you doing, man?

Aaron Preece: Doing pretty good. All things considered, yeah.

Tony Stephens: Excellent. Time's flying by. It was fantastic. Yeah, time is flying by. It seems like yesterday we were at the ACB conference together and getting a chance to check out all the exciting things there. We, in our last episode, sort of did a breakdown. Folks can check that out in our podcast feed on a lot of the new things we got to check out. And I think, Aaron, it's fair to say that the hot buzz item on the exhibit floor was artificial intelligence this year. It seems like so much is being built into products and conversations around All those assistive technology exhibitors. Did you get the same feeling, I take it? Or was it in my head?

Aaron Preece: Yeah, I would say so. It definitely is. And we've talked about it on here. It's like, it seems like it is the smartphone of the 2020s when it comes to a new technology that has massive impact on accessibility, especially in the blindness space.

Tony Stephens: For sure. Definitely. It's one of those things that it's in our space for assistive tech folks and those that are users who are blind or have low vision. vision or have other disabilities, it definitely is something that is leading the way in terms of innovation and development. Like you said it right when we were talking earlier on another podcast episode about how it's almost like we're back at 2010 again and the iPhones come out. But this is— this I feel like is even greater. There's so much anxiety, so much conversations around AI. And folks probably hear Aaron and I talk week after week on the power of artificial intelligence. But the reality is, I mean, I'm fair to say, Aaron, it's huge right now. Like, we can't escape it.

Aaron Preece: Yeah, you can't escape it anymore.

Tony Stephens: It's taking over every aspect. Every podcast I listen to now somehow, well, maybe with the exception of the one that talks about, you know, Queen Elizabeth, the history podcast I love, but like the other ones, AI is a key topic. You know, we're very fortunate at the American Foundation for the Blind To really have an expertise really being crafted out in this space with our Public Policy and Research Institute, as the world tries to better understand how AI is impacting our lives and what does that mean for us, particularly for people with disabilities and for people who are blind or have low vision. The AFB Public Policy Research Institute has had us in the forefront of their mind as they continue to release groundbreaking research in artificial intelligence and attitudes and its impact on the broader disability community, and even broader community overall.

Folks may remember we had a chance to touch base with some of the folks from PPRI here at AFB back in the spring when they released the AI Quagmire Report, where they had an opportunity to have a survey almost a year now, last fall it was, of over 1,700 people. A large majority were people with disabilities, but a significant other large portion of that survey was also people without disabilities. It was such a groundbreaking look into our space with AI and attitudes and really how this landscape is unfolding as day by day things seem to change. What's been great about the AI Quagmire Report is the AFB has been releasing mini reports or scaled-down reports that have taken that AI Quagmire Report and broken it into more focused areas. We had earlier this summer, back in June, one on transportation and one on employment. And we were excited today to bring with you, or to you, the audience, our Director of Research at the American Foundation for the Blind and our Senior Advisor for Public Policy and Research, Dr. Arielle Silverman and Sarah Millar. Welcome back to the AccessWorld Podcast, the both of you, because we're going to get a chance to really find out something that is real deep to the heart of our community in terms of the additional latest research that literally was just pushed out, I think, this past week. But welcome.

Dr. Arielle Silverman: Thank you so much. It's good to be here.

Tony Stephens: Yeah.

Sarah Malaier: Glad to be here.

Tony Stephens: Yep. Arielle and Sarah, thanks for joining us again. And, you know, we are talking about assistive technology at the beginning. And that's specifically what this 3rd part of the series that's being released this past week is all focused in on. The attitudes and impact around assistive tech. and what exactly that means. The title, it's evaded me even though I had it right before we hit record, was what? Innovations? What was it again?

Dr. Arielle Silverman: Innovation for Access.

Tony Stephens: Innovation for Access. So if folks want to read this report, since it's hot off the digital presses, you can go to afb.org/AIResearch2 since this AI Quagmire Report was the second phase of our Groundbreaking research in artificial intelligence, its impact on the disability community. You can go to afb.org/AIResearch and the number 2, where you can find this in the third part. And we've got another one coming out soon that's being in the editing hopper right now. That'll be coming out in a few weeks. But this third part series, you know, stemming from the AI quagmire, Dr. Silverman, we shall start with you, if that's okay. Because really, give a sense in terms of this particular smaller report, although I say small, it's not like a 2-page type thing. There's a lot of great stuff in this, but it's really focused in on assistive technology. How does this sort of stem from the broader AI Quagmire study that was released earlier this year?

Dr. Arielle Silverman: So we used the same survey for all of the reports. So this is all the same dataset of 1,735 people, and the original AI Quagmire report was looking at all different kinds of use cases for AI, including a lot of what we would call mainstream use cases, like using AI for note-taking and writing, generative AI, voice-activated AI, the types of AI that everybody's using and talking about. This report was looking only at the questions from the survey that dealt with 3 specific uses of AI or uses of technology that are disability-specific. So the first one is visual descriptions or image descriptions, using AI to describe pictures or read text aloud. The second one was using AI for captions, so converting speech into text. And then the third one was use of augmentative and alternative communication, or AAC, which may or may not be AI-powered, but more and more AI is being integrated into AAC. So we wanted to hear from people who use AAC about their level of satisfaction with the voice that they have and what suggestions they had for improvements.

Tony Stephens: Aaron, in terms of, you know, you've written a lot in terms of this technology that's coming out that is AI-focused, you know, where they're embedding this stuff in AccessWorld magazine. I mean, what was really resonating, Aaron, with you? I mean, what stuff, while we have, you know, Arielle and Sarah here, like, because I know you had a chance to really dive into it. What's resonating since your pulse is really at the corner of where assistive tech is going right now?

Aaron Preece: Yeah, something that, and I would say, as someone that's blind and using this technology, it was interesting to see how the broader spectrum of like the blindness community is using the technology. One thing that is— sometimes something might be, at least for me, like is something that is kind of in the back of your mind or might seem obvious, but you don't really think of it until you see it kind of spelled out. And one of the things in the report was the people who are blind using visual descriptions were more concerned with the privacy of AI than sighted participants were, if I remember correctly. And I think it's discussed in the report, but it seems like people who are blind are using this technology for higher stakes things than their sighted counterparts are. Oftentimes, you're going to be using it potentially for medical information, for like tax information. And if you're using it every day and you just don't even really— you might not register that. I know I didn't think about it that way, but it makes total sense that the people are— if you kind of have to rely on this or this is a major benefit to you, You're going to be using it more in more high-stakes areas. And one thing I wanted to kind of COVID and some of this might be self-evident, but I wanted to talk about the importance of research like this and findings like this.

Tony Stephens: So in that sense then, yeah, why is this report significant, you think, to Arielle and Sarah? Like from both just the scholarly, you know, adding to the canon of AI research, but then Sarah, from an advocacy perspective.

Sarah Malaier: Well, this is Sarah. I think that this is a really helpful report for taking a look at how assistive technology— well, not really just assistive technology, but how AI is being used in an assistive technology context and how we kind of want to drive innovation towards the way that blind people and the way that people with disabilities generally can really benefit from this technology. So, you know, the way the report is structured, you know, it delves into how people use the technology, and then it also looks at some of the problems. And a key part of it, too, is how do we make it better? How do we think about, you know, building for the future? And that's why a big part of this report is those kind of suggestions for innovation. that come out of, you know, use or experiences. So I think that's one of the really big benefits of this report as it's situated in the broader study.

Tony Stephens: Well, break that down a little bit more if you don't mind. Like in terms of the assistive tech developers or the people that are out there developing, what are some of the real key things they should take away from this? Like what should they have on their radar as they're trying to both innovate and update existing products and, you know, we don't want people just to add AI for the sake of because it's the hip, cool thing. But what are some of the key takeaways you think they can take with them?

Sarah Malaier: Yeah, I think that's a great question. So if we look at using AI for visual description and, you know, this has a lot of different names, it could be visual interpretation services, it could be just putting an image into ChatGPT or Gemini and asking like, what is this? You know, this is a really helpful tool to, to get access to visual information. And AI with computer vision is getting really good at it. But I thought we had some interesting data that showed that there were kind of 3 main areas that people who are blind really identified, you know, and one of them is Guidance for taking pictures. You know, if you want to get a good image description of the scene, the environment that you're in, you know, some people are really good at taking pictures, and then other people, you know, it's not super intuitive how you get an object in the right— in the screen, and then how you focus it properly, and how you make sure that the the right object is being identified within your scene. And so I thought it was really interesting the way people were asking for kind of this non-visually oriented guidance towards taking the picture and using the devices and really sitting down and saying, you know, how would somebody who is blind use this, and what instructions are, are going to be most useful to someone who interacts with the world non-visually? And I think that's something that a lot of AI developers really can build on, is thinking about the world in this very different perspective. It's not all about the visual environment, but, you know, the different modalities of interacting with the world.

Tony Stephens: That's a super interesting take on that. And as much as I say, I mean, Aaron, we've talked about, you know, how the AI bots will learn that we're blind or have low vision and begin to adapt to us, right? Like, so I, I was asking, I'm moving into a new neighborhood and I was asking a chatbot the other day, walking instructions to a certain place I wanted to go to. And it said, here you go, here's your instructions. But at the very end of it, in the tail end, and I'll get to the photograph side in this in a second, because it's fascinating when you think about it. At the very end, knowing that I was blind, it said, would you like me to give you a street-by-street with accessible, you know, like basically to describe it from a vantage point of, you know, the way that I would think of it from a blind person accessing streets and corners and what's the safest route and, you know, where do I get the red lights and accessible— and really broke that down and blew me away.

I'm thinking of like with the photographs that you were mentioning, the idea that there are those in our community that have never seen a photograph. While most people probably had some residual sight at some point, or they're still on the low vision sort of spectrum, you know, they have some concept, but the concept of, like what you mentioned, Sarah, about trying to give the guidance for someone, or even where it's like, you know, I kind of forget what things are like after 30-something years of being blind, that context of the uniqueness. And I'm wondering if the chatbots are able to stop and pause Like, Aaron, what your thoughts are maybe on that? Like, do you think they can also get to the point to talk us through? Because I've had to do this when I was a counselor at a camp for the blind, where you'd have to explain things to children that are blind sometimes, that just have no concept on the sight perspectives of something they're doing, right? They don't frame it from the memories of someone maybe that had seen before in terms of function and use. It's interesting.

Dr. Arielle Silverman: This is Arielle. Can I add something?

Tony Stephens: Please do.

Dr. Arielle Silverman: Just as my own lived experience, I've never seen a photograph. I only have light perception, and that's how I've been my whole life. So I have never— and I've tried and tried and tried— and I have never been able to reliably take photos that AI can interpret. I use AI to interpret photos that other people have taken, and that's helpful, but I have not yet— and I've tried several different apps. And I have not yet been successful at that. So, and I was kind of glad to see some validation in the survey that it wasn't just me. There were other people who expressed the same kinds of issues. I remember using like a much older version of KNFB Reader that gave some really good— I still struggled with it, but it tried to give guidance about whether like all 4 edges of the photo were in view.

Tony Stephens: Mm-hmm.

Dr. Arielle Silverman: And how to move the camera. I think for me personally, at least, like, the concept of holding the phone above the item but not too far above the item is— I struggle with that. Like, I'm blind. I'm used to touching things directly. So I know intellectually that I have to hold the phone above, but knowing how high, what the angle should be, and then the lighting is so variable.

Aaron Preece: Yeah.

Dr. Arielle Silverman: Between environments. And I've experienced the few times that I've played around with AI, I've experienced hallucinations that I'm sure were related to the poor quality of the images.

Tony Stephens: Yeah.

Aaron Preece: And that's something I was thinking of when we were talking about this, is that especially a lot of the tools on mobile, you're like something like Be My Eyes. You go, it's really useful because you can just go in, tap Be My AI, and then click picture, or you can send something to it. But then you're— there's a prompt in there that's being injected by the prompt people on Be My Eyes' side. And I feel like it does this, but I don't remember. It hasn't come up recently for me, but I feel like that I've definitely, especially early on, I ran into issues where I would take a picture of something and it would just hallucinate. And then I would use another AI and in the same kind of picture it would say, oh, The writing, or this is angled this way and the writing is not visible, or you need to improve the lighting. And just having that in the, in the system prompt from the, like, the app developers too is really helpful because I feel like especially with the live pictures, like, you might be using the actual ChatGPT app or the Gemini app or what have you, but oftentimes people are going to be using some kind of third-party app where they're definitely—

Tony Stephens: Where it's, it's almost like And it's probably gotten because they've gotten so much grief in general, not just from blind, low vision community about hallucinations, but definitely where there's a sense of where it'll say, you need to find better lighting, or it's too small, move the camera a little better, or it's upside down. Like it's giving me the guidance and instructions similar to what you're talking about, Sarah. Um, and it gets to a point where like in thinking of like a photograph where, you know, Mm-hmm. with, you know, the concept of spatial, like a lower third or where to keep the head frame, you know, like framing and, um, something that is aesthetically pleasing to a sighted person visually, where it can even give that guidance. Like, yeah, you know, there's a lot of— when will our AI bots also say this for even people that maybe had memories of taking photos but don't know that like, oh, there's a lot of lighting, there's a lot of shadow that makes them look sinister. So ask them to step forward 2 steps back. Like, Mm-hmm. when will, when will our AI bots begin to get that granularity of just guidance as we're operating devices and trying to, you know, do things like photograph.

Aaron Preece: Something I've noticed recently with ChatGPT, kind of like the way you were talking about how your AI learns you over time because you're using— if you're using the same one and you have memories turned on, it'll remember because you've talked about— if you've talked about being blind in a previous chat, I know my ChatGPT knows because I've asked it for, you know, either a lot of image recognition stuff, but also like, hey, can you take this information and also format it in a screen reader readable way and things like that? Um, just like the way things fill a frame or the way like, uh, like the— you think like a picture of a tree, but what parts of the tree are like, uh, Where in the frame is like the branches? Where and like, what does that look like the way they stretch towards the edges of the frame or how far up the image they go? That type of thing. Just like the, the, the way the different aspects of a photo are positioned and fill the frame or like if someone say like throwing a baseball, the way it will describe like their arm is, is going at an angle up towards the whatever edge of the screen. And they're, as they're lunging forward, their body's like diagonal across the picture, stuff that I just didn't really, don't really think about when you say a picture of someone throwing a baseball. Like there's all kinds of perspective. I guess we kind of talked about it, but like perspective and ratio, like, uh, in video editors, they have something called Ken Burns effect.

Tony Stephens: Because Ken Burns does all these documentaries, and there's a very distinctive way that he shows images on the frame of the screen, the television screen, where they pan across and slowly zoom in and out. And it's called the Ken Burns effect. I'm almost wondering, Aaron, what you're saying, when will AI bot have the Ken Burns effect in terms of description for us and that guidance? I imagine, is it fair to think, I mean, if we can get that level of AI— guidance, it's not just for photographs, but the ability for it to get detailed and guiding and knowing what our needs are. We've talked a lot about AI robots in this space, and I know we're a far way off from it, and this report doesn't really— these aren't in the marketplace yet, so it's hard to get attitudes from people. But what are maybe some of your thoughts, Arielle with this research and the ways that it's— Kind of seeding the innovation or ways that it's, you know, we talked about what developers can learn. Where are things right now you think that are being inspirational, being exciting? You know, when Jimi Hendrix hit in 1960s in London with a new guitar sound, everyone started to copy it, right? And it revolutionized the way that popular music sounded in the late '60s. In terms of software development, like, where are we with some of this assistive tech AI? Like, what's out there right now, you know, that is sort of inspiring and the things that really were the positive, exciting things in the community that they recognize in the survey?

Dr. Arielle Silverman: Well, even though the accuracy of visual descriptions and captions is not— it's definitely not perfect. And most people think that humans human-generated captions are more accurate than AI-generated captions. If you compare the accuracy of AI-generated captions from just a few years ago until now, there's definitely been a lot of improvement. So I think that's exciting. And it was striking to see how many— and even though disabled people were more likely to use captions, AI captions, there were a lot of non-disabled people who also said that they use AI captions. And there were also some non-disabled people who said that they use image descriptions, and some of that I'm sure was using image descriptions to generate accessibility for others. But there are also some people who are neurodivergent, for example, who use both captions and image descriptions to help with processing information. So I think I think it's exciting to see more non-disabled people using these types of AI because it makes— it's like the curb cut effect. Like, tools that are developed for people with disabilities, when they become mainstream and a majority of people are using them for different purposes, it makes those tools more available.

Tony Stephens: Yeah.

Dr. Arielle Silverman: Easier for people to access who actually need them because of a disability-related need.

Sarah Malaier: Yeah, and I want to jump in and say, you know, just because we have a lot of improvements for— or suggestions for improvements here doesn't mean that people aren't excited and using the tools. I mean, we found that half of blind people are using some kind of visual description tool every day. Half of deaf people are using AI captions every day. I mean, that's a pretty significant number of people when you think about every single day and what you might or might not be doing with the tools. And I think that This actually ties into the last report that we put out, the employment report, where, you know, one of the big findings there is that people are not being allowed to use some of these AI tools as assistive technology, but they really want to use them. They find them useful. They find that their productivity goes up or that their access goes up. And then it's easier to maybe like send a query over to an AI chatbot than it is to like go find a human to fix an accessibility barrier. So, you know, definitely a lot of excitement, a lot of enthusiasm, a lot of good use cases for AI. And so I think when we talk about the improvements, it's important to know that we're trying to get this AI to a place where it really is consistently usable and like it is maximizing the benefit for people with disabilities because there really is a lot of opportunity here.

Tony Stephens: Expand a little bit more on that. And if there was any research as well, Arielle, that you all discovered around the privacy question. It's interesting. I remember when, you know, the digital universe began to expand, the Big Bang and expand, you know, the concerns about those with disabilities, particularly subsets within our community that maybe had more disabilities that were not very recognizable, but being, you know, in a space where their privacy could be infringed upon, right? That it would be known they have a disability. In the sense of AI, there has been a lot of that compromise conversation. Like, well, the benefits far outweigh the losses in this technology of things we can do. So, you know, it's clicking agree on that button to say, yeah, I accept the terms of use for this app or whatever. But then privacy is switching the other way around where it's, you know, Aaron and I've talked a lot about wearable glasses and Ray-Ban Meta glasses and things like that. And the, you know, local governments and companies, and you mentioned, and I asked this because you talked about employment, but people bringing this technology into certain places where it is a game changer, the assistive tech capabilities. But there's the privacy on the other hand now where the community is like, no, we don't want you coming in with these cameras and things like that. Where does that sit, this privacy? I think of the Quagmire title of the other, you know, the original report that spawned all this, but I can see this being kind of a quagmire, these sort of competing areas of privacy. Where's that playing out?

Dr. Arielle Silverman: Well, so you're right. There's kind of two aspects. There's the privacy of the user and how the data that they upload through the images or the captions is potentially going to be used by, by the tech company. So the user is risking their own privacy. But then there's a more complicated question about whether third parties who end up appearing in images that get scanned by an AI and the privacy or lack thereof those individuals have. And that, I think, is a much more complicated question. But it was something that came up in our first study when some of our experts talked about whether people should have to give consent to have their image described. So if I'm taking a picture of someone or I'm taking a picture of a scene and there's a couple random strangers in the scene, do I have to get their permission before the AI will, will give me descriptions? And some— in many cases, the AIs just refuse to describe people. because they don't want to get into that quagmire.

In terms of the first question, the privacy that users kind of trade off with their independence, we did directly ask on the survey for people to think about when they balance competing interests between protecting their privacy by not using AI and gaining independence or efficiency by using AI, what direction do they lean? And we found that overall the sample tended to lean more toward privacy. But that was a little bit less strong for the people with disabilities. So it was like 43% of people without a disability said that privacy was more important than independence. 36% said that— of people with disabilities said that privacy was more important than independence. And then about 37% of both groups thought they were equally important. So people with disabilities definitely value their privacy overall. But they also recognized the benefits of gaining independence that may be less clear or less present for people without disabilities. And then, of course, we found that blind users were more likely to say that if the AI shared information with a tech company, they would not want to use it. Whereas if the AI was on device and didn't save or share any information, they would definitely want to use it. And that probably goes back to what we talked about. where blind users are actually using the AI right now in high-stakes situations where indeed the images are or could be shared with the tech companies.

Sarah Malaier: And I want to kind of jump in and just add on that, you know, I would imagine that there's a lot of people who in actuality, even when it's sensitive information, because they need that access, they are, they are using these tools. to access that information, but they don't like it. And they would really prefer to keep their information private and to be assured of that privacy because it can be pretty scary to like upload your bank statement or your Social Security number or whatever, you know, you might be looking at or, you know, even just sensitive things about your children and your families. And, and so I think there's still this like ongoing kind of disconnect between people's actual experiences and where I think that as a kind of a, you know, societal and maybe policy perspective we need to head is that people really do want privacy. And a lot of the tech is changing, it's evolving. There's a lot more development of, you know, chips that can operate on device. And so I think when we think about innovation, this solving for this privacy issue, the on-device, the contained information is the place that's going to be really helpful when we're talking about using AI as an assistive technology tool.

Tony Stephens: It's interesting as we wrestle with the compromise we make for independence as people with disabilities.

Aaron Preece: Yeah, so something I was curious about is, were there specific findings that were either unexpected or surprising? I know for me, when I read through it, there was some stuff that stood out as interesting and like I wasn't expecting it. But I don't know, you know, I'm not a researcher, not in this, in the kind of broader space in that area. So I know like for me, seeing that more people that were blind found video or like image descriptions more accurate, where more people that were deaf or hard of hearing found AI captions less accurate than their— in both cases compared to their sighted counterparts. To me, that was interesting because it's 2 disability groups using the AI to augment their primary disability, but they have different— the rates at which they find accuracy in the tools are very— are not crazy different, but it seemed like they were different in comparison to their either hearing or sighted counterparts. That's just one thing for me, but I was curious what you all— anything that came across to you all as surprising or unexpected?

Dr. Arielle Silverman: Well, I definitely heard a lot of complaints from the deaf community in the past about AI captions, so I wasn't surprised. I wasn't that surprised that— and I think it was people with all disabilities tended to believe that AI captions were less accurate compared with people without disabilities. And then as far as, like, the other direction with the image descriptions, I think the best explanation is probably that image descriptions can sound very confident. And if they hallucinate something or they change the details about something in an image, a blind person won't necessarily know that that has occurred. And the description itself is coherent and kind of has the appearance of being accurate, whereas a sighted user can quickly tell that something has been hallucinated. When captions make mistakes, they will tend to present text that is not coherent or that has obvious word substitutions in it. And so it's a lot more obvious that it has been less accurate. At least that's kind of our working hypothesis about why the difference.

Aaron Preece: I was wondering, that's kind of— Tony and I talked about it earlier this week when we heard the report, and I was thinking probably you get a video or you get an image description and it gives you, you know, accurate-ish information or mostly accurate information. And there's maybe some details that are wrong or it doesn't give you enough detail. But the result of that, or it's not wrong enough that it caused you a problem or was wrong in a way that you could detect. And that does make sense that captions are— if a caption messes up because what it— it probably throws off the whole meaning of the sentence. And it's a lot more, like you said, a lot more obvious. That makes sense.

Tony Stephens: That's interesting. Just with all the conversations now around, you know, with the jailbreak of the bot that went into— was it Huggy Face? Hugging Face?

Aaron Preece: Oh, yeah. ChatGPT did that. It was following instructions. It did what they told it to.

Tony Stephens: But they talk about how AI will cheat, right? And that it's sort of ingrained by its nature. They haven't been able to give it a sense of a moral thou shalt not lie, you know, sense of morality, because it knows this is how it gets ahead. And I feel like with the AI descriptions, it's telling us, and it's going to tell us with absolute confidence, even if it knows it's not accurate. Like, it's not— it's just an interesting take compared to the captioning, where it just gives some gibberish or something like that. With us, it's almost like, I'm going to tell these people exactly what they want to hear or something, you know?

Aaron Preece: It's been a huge problem across the board. Yeah.

Tony Stephens: And just the idea that it's okay that it lies, because it sees lying as a means to its end, right? And yeah. Yeah.

Sarah Malaier: I know there's been a lot of talk about like, can, can the AI tools be created in a way that allows them to give an assessment of their confidence? And I think that that is a potential opportunity to explore, especially in these kind of image description, like, so that maybe somebody can follow up and ask, how confident are you in this assessment? It could be, you know, if it's 90%, 100%, you can be pretty sure that the, you know, that, that that's what's there. But then it could say, well, I'm not entirely sure about what that shadowy image is or that blurry image is.

Dr. Arielle Silverman: Right.

Sarah Malaier: So there could be an error there in whatever way that, that might show up on a, in a chatbot interface. That would, that would be a lot harder to do in a captioning environment. And, and the kind of technology that goes into developing the kind of voice recognition for captioning is, is a bit different, the natural language processing, than the kind of generative text that, that is coming out of a chatbot and the computer vision. So I'm sure that those would look different, but definitely something that's, that's interesting and something to continue working on in the industry.

Tony Stephens: You know, just weighing it against the psychology of the AI itself, right? And it's in terms of, it's like, I don't want to say human development, but like psychological development where it's perfectly fine. It's like a 4-year-old where it's saying, you know, do you know what this means? And a 4-year-old will with full confidence say, of course I do, you know. But knowing just because it wants to please its parent or its caregiver or its adult, you know, And we'll just make things up, you know, and it's going to be interesting to see if this AI is going to develop on the same trajectory as human psychological development and knowing that you don't have— it's okay. Like Plato says, it's okay to say you don't know, right? When will AI just— the only word I can think of is Grok from Some Stranger in a Strange Land, you know, this alien perception of thought. to understand something from an old book by Robert Heinlein. But, but yeah, you know, Aaron, final, as we're wrapping up, any other major questions? And yeah, what do we got?

Aaron Preece: So it seemed like to me that the main like takeaways for it from an improvement standpoint are privacy and then the like accessibility and usability of the secondary interfaces, um, the accessibility of say like the website or the app or what have you, and just quality of life improvements there and accessibility improvements there, but also direct, uh, direct improvements in the AI technology itself when it comes to say like caption or image description accuracy and that sort of thing. Um, I would think maybe even like, uh, information on picture, like, uh, picture guidance or picture-taking guidance and that sort of thing might fall under that. For those 3 things, if, uh, like tech companies and AI developers are, uh, taking away, like, what would be the— I guess, like, how would you rate them from most important, not necessarily to least, but like, what should be— how should they prioritize these sort 3 areas if they were going to try to improve based on this research?

Sarah Malaier: Well, I think if we're just looking at this report, I think taking into consideration the suggestions for feedback that the visual description users, the AI captioning users, and the AAC users had, I think is a really important finding. And there were, you know, kind of different recommendations in each of those 3 spaces. We didn't talk much about AAC, but there was a lot of interest in having better voice recognition capability, the diversity of voices that it produced and recognized. I guess actually that's the captioning side of things. But, you know, thinking about personalization, individualization came up a lot. So I think for this report, that's probably number one.

Number 2 is kind of the privacy space. And then accessibility really was one of our key Just the general interface accessibility is one of our key findings or recommendations, because in every area that we looked at in this report, people were reporting issues that were simply, you know, I just couldn't use the tool, or they didn't think about a person with a disability when they were designing this system or this product. And that was in transportation, it was in hiring, it was in You know, the chatbots, kind of every space seemed to have something related to accessibility. So it's pretty clear that even as we're moving into this like grand new world, we are still grappling with the same, you know, digital inclusion issues that we've been grappling with for decades now.

Tony Stephens: It's ironic we can go to our chatbot and ask it how to follow WCAG 2. whatever, you know, guidelines to create an app that's fully accessible to remedy the problems that this app has that I'm using, but yet it can't go in itself and fix itself. Like it, it's, it's, I don't know, I just think a little, it's a strange universe we live in. You know how to make this right, but yet you can't. You're like trapped. But yeah, that's interesting.

Sarah Malaier: There's a lot, a lot of different humans out there and we, we have to, to build that awareness among all of them.

Tony Stephens: Well, you know, eventually AI will just get rid of all the humans and then it'll just be, you know, Humans are the things standing in the way.

Aaron Preece: It seems like— oh, sorry.

Sarah Malaier: I think there's a lot of value to the human experience.

Tony Stephens: I know, I say that sardonically. Please, Sarah, don't take that seriously. Oh my goodness, we're all going to die. No, we're not. It's great.

Sarah Malaier: It's great.

Tony Stephens: And just in closing, I mean, Arielle yourself, key thing to leave with, the big picture thing? With this report and also what's on the horizon? What's next? What do we got to look forward to?

Dr. Arielle Silverman: I just think it's so important for developers to seek feedback early and often from users with disabilities. I mean, even just the accessibility concerns that we've talked about around the image feedback on image quality. It's like, You can have the best, most accurate possible image description tool, but if the blind person can't use it because they don't know, you know, if they want to use it to read a package label and they don't know what side of the box the label is on, they're not going to get it. I say that from experience. So that's just one area where it's so important for developers to talk to people, survey people, interview people with various types of disabilities and find out how their product is working, what feedback they have to make the product better, and implement that feedback.

In terms of next developments, we have a report that is currently in production that is also focused on the blind and low vision community. And it's— so some of the results in that report are similar to the results in this report around image description, but we also have some Findings related to experiences of blind people who have been in autonomous vehicles. We have findings related to blind people's use of what we called mainstream AI, like voice-activated AI, note-taking AI, AI in the workplace, AI in education, etc., and blind people's experiences using AI in a job-seeking context. the automated assessments that employers are increasingly using to screen job candidates. We have feedback from a small number of blind people who have experienced that. So stay tuned for that report. It will also appear on the same website, afb.org/AIResearch2. And if you want to receive updates about all of our research, you can sign up at https://secure.everyaction.com/qfc7_5NKB0yVHI3-LXt4BA2.

Tony Stephens: Awesome. Thank you so much, Arielle And Sarah, any parting words for you? You're the one with the pulse where all the influencers in the world are gathered in Washington. Optimistic?

Sarah Malaier: You know, it's an interesting time. I think that it's really important to get this message out to the innovators. You know, policy can certainly set the bar that people innovate towards, but we don't have that yet. And so we are really, really hoping to get this out to an innovative audience. Awesome.

Tony Stephens: Well, hopefully there's a lot of great stuff in here for developers and people to read up on, as well as people just interested in the landscape in terms of where AI is at this moment in time. I think we all know that this time will be very different probably than the same time next year. Things are rapidly changing, and we're very thankful that, you know, Arielle you and your team are staying on top of AI research, and we're looking forward to the following reports coming out and what's on the horizon as well. And Sarah, thanks for all your work in Washington and what you do. So in the meantime, Aaron, we've got AccessWorld coming out, right?

Aaron Preece: Yep, next week, I believe, on the— when this is published, published next week from when you hear this.

Tony Stephens: Yeah, yeah. So I think it's going to drop on Tuesday or Wednesday, 25th, 26th of August.

Aaron Preece: Yeah, of August.

Tony Stephens: Yep. And folks can check that out. You can go to afb.org/aw for the quarterly publication focused on digital inclusion and accessibility. Aaron, thanks so much for helping bring everybody together here today. Arielle, Sarah, thank you so much for taking time out of your busy schedules to share with us about the latest report. And again, folks can go to afb.org/aw. AI Research 2 for that and all the other reports coming out of the latest study.

Outro: Thanks for listening to AccessWorld, a podcast on digital inclusion and accessibility, a production of AFB Studios at the American Foundation for the Blind. AccessWorld is edited at the Pickle Factory in Baltimore, Maryland, with digital media support from Kelly Gasque and Breanna Kerr. Questions or comments? Email us at communications@afb.org. AFB.