Accessibility Isn't Built Into AI, and That's a Problem  🔗 **Source:** [Accessibility Isn't Built Into AI, and That's a Problem](https://share.snipd.com/episode/43b5dc65-9a34-4697-9bb6-75edd1ae492c) ## Highlights - **AI Reflects Historical Exclusion** - AI models mirror historical data gaps so inaccessible pasts produce inaccessible AI today. - Michelle and Samah note training data lacking diversity causes biased outputs that exclude disabled and marginalized users. Transcript: Michelle And funding are shrinking. Support is shrinking. So we looked at all four of those things. And we don't want to talk about layoffs again because, quite honestly, I can't hear it anymore. I'm still laid off. I'm still looking for my next big thing. But we looked at the continued rollback of DEI efforts. Is that the one? Oh, no. Accessibility still isn't being built into AI products. So I clicked through, right? Is that the one? Yeah. So we clicked through and we find this letter, this online thing from Ricky Onsman. Now, if you don't know, Ricky Onsman is a WordPresser. And here we like randomly chat GPT and Googling shows us his article. The first was from May 19th, 2026. The second one was way more frequently than it was a May 20th. So over two days, he wrote this and it's very long, very well-articulated article. There's way more in this article than we will be able to touch base on. And there's a two-part, so you can come back for the second part. So definitely we will link these articles in the show notes. But we did want to talk a little bit about what AI is all anybody talks about. And I don't think we've talked about it that much on this show. So I guess we had to get with the program. Samah Talk about it. Michelle So one of the things that I read in here, which I thought was like, whoa, you know, the first is these tools, the tools that we use can save time, help scale accessibility efforts, but they Do have limitations. For example, never use accessibility overlays because they actually aren't accessible. People think that they are. It's a load of bull that you get sold. Like, just don't, right? And then it goes on to talk about code suggestions and all these other things. It gets to this area of the article called bias. And the first thing that you read is AI systems learn from data, and that data is not always inclusive. As a matter of fact, most of the time it probably isn't because history hasn't been inclusive. ([Time 0:05:46](https://share.snipd.com/snip/e88d473c-0355-4479-a673-773ca2eb586e)) - **Avoid Accessibility Overlays** - Avoid accessibility overlays; they promise fixes but are often not truly accessible. - Michelle warns overlays are a Transcript: Michelle So one of the things that I read in here, which I thought was like, whoa, you know, the first is these tools, the tools that we use can save time, help scale accessibility efforts, but they Do have limitations. For example, never use accessibility overlays because they actually aren't accessible. People think that they are. It's a load of bull that you get sold. Like, just don't, right? And then it goes on to talk about code suggestions and all these other things. ([Time 0:07:06](https://share.snipd.com/snip/de3a14b0-1998-4ed5-b93a-391357e11de3)) - **Real Voices Get Misheard by Voice AI** - Voice recognition often fails people with nonstandard speech, like users of speech synthesizers or those recovering from tongue cancer. - Samah and Michelle reference a TikTok creator relearning speech and Stephen Hopkins as real examples of AI errors. Transcript: Samah Yeah, for sure. There's high error rates or complete failure in voice recognition may be evident for people with non-standard speech pattern. This is common in people with virus disabilities and impairments. As an example, they put Stephen Hopkins uses speech sanitizer that AI might have trouble understanding, which is really nice. Good example. Yeah. Michelle Yeah, for sure. I think it's really interesting to even just accents. Right. So like if you have a Google Home or you have an Alexa speaker in your home, it learns to adapt to your, you know, where you're coming from, what you are. So, and I could speak to it in Spanish, right? It understands my really bad Spanish, but somebody else might come in and not speak their native language, but try to talk to in the car, because I have the car one that I can just play music, Or in the, and it won't understand their American language in a foreign accent and that's just that's not a disability that's just saying simple things but then you get people who have Disability with speech speech impediments there's a woman that I follow on TikTok who had tongue cancer and has had most of her tongue removed and she's learning to speak again and I Have trouble understanding her I can't imagine that AI can understand her very well so that's the first one that's built into this and ([Time 0:08:23](https://share.snipd.com/snip/1b4a56cf-0952-4e15-b9a9-02ff21bfbd7a)) - **Accessible In Theory But Not In Practice** - Physical accessibility that technically exists can still exclude users, like scooters too big for tiny lifts at polling places. - Michelle describes squeezing her scooter into small elevators and being unable to vote at her assigned polling site. Transcript: Michelle To let understanding what I'm saying. Absolutely. It's similar to when I was in Italy for WordCamp Italy that was in Torino. And everything had elevators. But the elevators were so small that I had to take the basket off of the front of my scooter, go all the way, touch the glass or the wall in the elevator, and hope that the door would still Close behind me because they were so tiny. So is it accessible? Yeah. Is it accessible? Like my polling place. My polling place is one of those places that has a lift inside. So it's a church and there's stairs to everything. There's the stairs between levels of all this stuff. There's a lift. My scooter is too big for the lift. So I can't vote there. So now I have to either vote early and go someplace else or vote by absentee ballot because I can't vote in my polling place because even though it's accessible, it's not accessible to Me. ([Time 0:10:50](https://share.snipd.com/snip/55633870-d115-4015-bfc6-c0f9f079d141)) - **Face AI Breaks For Many Real Faces** - Facial recognition and vision AI struggle with nonstandard faces, excluding people with scars, Down syndrome, or face coverings. - Michelle points out a pictured example and shows how masks or injuries break face login systems. Transcript: Michelle AI facial recognition may not recognize non-standard faces. AI vision recognition tools struggle with non-standard or diverse physical appearance. Again, this is common in people with various disabilities and impairments. And the non face they're showing on the screen is somebody who looks like they might have Down syndrome. And I'm absolutely drawing that just from look. So forgive me if that's not what they're trying to show. But that is what it looks like they're trying to show. You know, maybe massive scarring or women who have to wear veils or choose to wear veils and face coverings, you know, and that's not always going to be something that's accessible for Them either. Like my PC laptop, which I gave to my boyfriend, because why am I using a PC? I don't game. His computer died, so I gave it to him. Anyway, when you first looked at it, I forgot what password I set up, but it had to do with your Facebook recognition. And so his brother was over trying to use the laptop, and he had to keep running out and putting it in front of my face because I couldn't remember the password. So, you know, if you are somebody who knows my face, but let's say I've got a black eye or, you know, I've got a mask on or anything, and then it's not going to work. ([Time 0:11:51](https://share.snipd.com/snip/c29c3df9-f6ee-4b8f-8e91-e4713e106930)) - **Touch Interfaces Often Ignore Mobility Needs** - Touch-based interfaces and tiny touch targets exclude people with mobility and dexterity impairments. - Samah and Michelle describe tiny targets, arthritis, and needing a stylus to avoid accidental taps in mobile games. Transcript: Samah AI interaction that relies on touch interfaces, often inaccessible and unforgiving for people with mobility impairments and lack of digital uh dexterity dexterity yeah yeah i can Make touch targets too small and place them too close together i can agree on that because if you have big hands or or sometimes with the screen touches or if you have a tiny hand you have To really push hard on order to see it, this is really inaccessible. Yeah. Michelle It really is. And for people with arthritis and people who, you know, for whatever reason, aren't able to use that touchscreen or see it, understand where the goal is. I play one game on my phone that I have to use a tiny little stylus with a little tiny tip underneath the plastic at the end otherwise it hit my finger will hit four things at the same time So yeah um gen ai that responds to user prompts frequently at this stage perhaps even predominantly ignores the lives of people with disability in its response yeah that's crazy that I don't even know what to say to that like let's just pretend that an entire group of people don't exist. Exactly. Yeah. It's so sad. Yeah. And these are people who pay money for all their services. So pay money to buy clothes, pay money to buy food, pay money, pay money to buy all the things in their life, just like the rest of us do. ([Time 0:13:14](https://share.snipd.com/snip/09d0ae8c-2d08-4185-b867-e94d2fa452f1)) - **Generative AI Treats Disability As An Exception** - Generative AI often ignores lives of people with disabilities unless explicitly referenced, treating them as exceptions. - Michelle highlights commercial harm: excluded customers still spend money but are left out of AI outputs and product models. Transcript: Michelle Exactly. Yeah. It's so sad. Yeah. And these are people who pay money for all their services. So pay money to buy clothes, pay money to buy food, pay money, pay money to buy all the things in their life, just like the rest of us do. And you're just going to ignore them. I don't I don't like that. Samah And also AI trained on board LLM largely treat people with disabilities as exceptions and don't include, consider, or even recognize them unless specifically referenced. This is really annoying because like for me it is, how can I say, it's like you're being putting one specific group, you're not targeted group as a customer you're not including them Michelle It's kind of a little bit yeah it is it is not as not for sure it even says this kind of bias has very serious implications when AI is used to generate models of products or services intended To meet the needs of the whole community and we've talked about this before. We've talked about how advertising only reaches certain groups of people, how when there is advertising, the very first time that Cheerios, and it was less than five years ago, the Very first time Cheerios showed an interracial couple with a mixed race child, they got serious backlash because how dare a white person be married to or be in a relationship with a black Person and produce this child you know um that's my child my child's a mixed race child you know like shut up cheerios is doing the right thing you know but there are so many people that Are just in their own little world and when you use, I'm trying to think, when you only use people who don't require accessibility, when you only use people who are white, Anglo-Saxon, Protestant, whatever, you know, the people at the top of the food chain kind of thing, you are missing out on so much rich history and heritage and people who want to buy your product, But will look for somebody else to buy it from because you're not inclusive. So I think it's important. I think it's really important that we think about how we use AI and what story we're getting back from AI, what information we're getting back, and is it accessible or not? Does it include accessibility? So for people like me and you, we're using it. We don't have any of those issues. We don't have to worry about touching on the screen. We don't have to worry about hearing things. We can see everything perfectly fine. ([Time 0:14:27](https://share.snipd.com/snip/f73e8e9d-b223-4d57-a176-f920ce642a20)) - **Train AI With Inclusive Accessibility Data** - Train AI on inclusive data and add accessibility-focused examples to reduce bias. - Samah stresses that while current data reflects inaccessible history, developers can augment training sets to include disability experiences. Transcript: Samah And also, as you said earlier, the AI taking from the data which really exists now, and the history and what it's happened now, it's like a lot of things are not accessible, taking it from It, but also we can train it, we can add more information. ([Time 0:17:24](https://share.snipd.com/snip/dfb880b3-4bc8-45d4-bb64-295c916af1c7))