Latest / Accessible Community / AI Trick or Treat in The "AI Learning Circle" with Justin Duran

Transcript

AI Trick or Treat in The "AI Learning Circle" with Justin Duran · Accessible Community

0:00
Eileen: Hi everyone, welcome.
0:01
Eileen: We're happy you're here.
0:02
Eileen: My name is Eileen Dobzinski.
0:04
Eileen: I'm the project manager of the AI Learning Circle at Accessible Community.
0:09
Eileen: Accessible Community is a nonprofit.
0:12
Eileen: We're dedicated to helping disability-inclusive communities grow through socially responsible technology.
0:18
Eileen: And the AI Learning Circle is the direct result of requests from volunteers, members, current and former partners and clients who want to have nuanced conversations on AI and accessibility, and also be able to more easily find resources and information to make personal and professional informed decisions of AI in consideration of accessibility.
0:43
Eileen: And so we've been operating as internal programming for the last four or five months.
0:49
Eileen: And now we're finally excited and ready to start offering events to the public like this one.
0:55
Eileen: And a huge, huge thank you to Taylor and the entire Accessible Community Podcast team for making tonight's live event possible and also the recording, having it out in time for Halloween tis the season.
1:08
Eileen: And so we're really excited for tonight's conversation.
1:13
Eileen: And if you haven't heard of the Accessible Community Podcast before, I highly recommend checking it out.
1:20
Eileen: There are currently two series available.
1:24
Eileen: One is Accessibility and Practice, and the other is Day in the Life.
1:28
Eileen: We're closing in on a thousand downloads.
1:30
Eileen: So if you check us out, that would be really amazing.
1:32
Eileen: And we've had over 30 guests and are currently trying to recruit for individuals who either personally or if you know someone who have um dealt with substance abuse, and that'll be likely a podcast for the new year.
1:47
Eileen: Taylor can talk a little bit more about it later, but it's really just such a great podcast for hearing diverse perspectives and lived experiences.
1:56
Eileen: And I think it ties really well into what Justin and the Noma team is doing and really wanting to have lived experience be at the forefront of AI and accessibility.
2:09
Eileen: And so it's my honor to introduce you all to Justin Duran, who's here tonight on behalf of NOMA, and we'll be doing a presentation for the first half of the hour, and then we'll be transitioning over to an open conversation.
2:23
Eileen: You all are welcome to have your cameras on if you would like, and there'll be an opportunity to engage either in the chat or with your microphone on later on if you're comfortable.
2:34
Eileen: Um, but yes, with all that being said, Justin, the floor is yours.
2:38
Justin: Awesome.
2:39
Justin: Thank you, Eileen.
2:40
Justin: Uh thank you, Taylor, uh, for inviting me on and giving me some time today.
2:43
Justin: Yeah, it's awesome what you guys are doing.
2:45
Justin: You mentioned uh closing in on a thousand downloads.
2:48
Justin: So uh congratulations, it's well deserved.
2:50
Justin: Um so good afternoon, uh everyone who's joining us here.
2:54
Justin: Uh like Eileen said, my name is Justin Duran, and I oversee uh client partnerships at NOMA.
3:00
Justin: Uh that's N-G-O-M-A NOMA.
3:03
Justin: Uh quick audio description.
3:05
Justin: I am a Caucasian and Hispanic male, uh brown eyes, brown hair, combed to the side, back a little bit.
3:13
Justin: I'm in a two-button polo, tan colored with a collar, uh, coming to you live from my home office where I have a uh little bookshelf behind me and uh some drawings from my kids over my right shoulder that uh I've been gifted over the past couple of father's days.
3:33
Justin: One more note uh on the audio description.
3:35
Justin: I've got on some loose, comfy uh Avengers pajama pants that match a set that my son has.
3:42
Justin: So uh no, I'm not gonna show them.
3:45
Justin: Thank you.
3:46
Justin: Uh audio descriptions.
3:47
Justin: Uh, there's your first accessibility treat of the night.
3:50
Justin: Uh I also wanted to kind of set the tone early for vulnerable authenticity in hopes that that carries over into uh the interactive portion of the event.
3:59
Justin: So there you go.
4:00
Justin: Uh thank you all for joining.
4:02
Justin: Uh, very grateful uh to be here with you all.
4:05
Justin: So, like uh seemingly everybody else, uh, I've been thinking a lot about AI or artificial intelligence and specifically how it might impact people.
4:16
Justin: So increasingly we're seeing AIs being used uh in consequential interactions, like conducting job interviews, uh recommending care plans, and proctoring exams, for example.
4:27
Justin: And even everyday conversations are now passing through a machine before they reach a human being.
4:32
Justin: So for people with disabilities, uh, that machine can really open up new opportunities, or it can shut them out at a speed and scale that was previously unfathomable.
4:43
Justin: These are the systems that I spend my days thinking about.
4:45
Justin: Uh at NOMA, we focus on how AI actually behaves for people experiencing various disabilities, and especially where the stakes are the highest.
4:54
Justin: So think uh hiring, healthcare, education, benefits, finance, just to name a few.
5:02
Justin: Our work centers on integrating the lived experience of people with disabilities into the AI design and use processes, and we do that by way of independent evaluation and expert uh improvement consultation.
5:16
Justin: Uh so since these automated systems and tools are making these consequential decisions, uh, we feel that we need a new accessibility standard that doesn't stop at the digital interface level.
5:29
Justin: So our methodology is informed by current accessibility regulations, uh but it does go beyond them in the sense that we seek to assess the overall uh equitability and human impact that these systems can carry.
5:42
Justin: And though the methodology itself is quite complex, uh the idea behind it all is very simple.
5:49
Justin: Let's put AI to the test and see where friction, obstacles, and failures exist for real people before we trust them with live situations that carry real weight.
5:59
Justin: So I'm here because I believe that the disability community stands to be the most affected by these automated systems and tools and therefore should have a respectable seat at the table.
6:08
Justin: Uh I will say I speak only on my own behalf, uh, so please take what follows as my own analysis and opinions.
6:16
Justin: And with that, uh the question for today is where AI has helped people with disabilities and where it has failed them or even caused harm.
6:24
Justin: I brought three pairs of tricks and treats to share in the form of some real-world examples.
6:30
Justin: Uh, and each pair places a treat, which is an AI tool that helped, next to a trick or an AI tool that has caused harm in a similar environment.
6:42
Justin: So the side-by-side comparison view I think is helpful when considering what uh contributes to the success or uh the beneficial usefulness of a technology.
6:52
Justin: And then after the pairings, I want to share how I see integrating lived disability experience into AI working in practice.
7:00
Justin: Uh, and then some next steps and considerations uh regarding use of automated systems and tools.
7:06
Justin: So the thread through all of it is simple.
7:08
Justin: Uh the same technology can be a door or it can be a wall.
7:12
Justin: And what decides which one it becomes isn't how clever the AI is.
7:15
Justin: It's whether anyone checked how it behaves for people with non-normative lived experiences before it is released.
7:21
Justin: I'm a proponent of the curb cut effect, and I believe that when AI is built for often marginalized users, then the experience will improve for everybody.
7:31
Justin: So for the first pair, I'd like to share two examples of AI in speech recognition.
7:36
Justin: So for a long time, uh voice assistance and dictation worked well for some voices and quite poorly for others.
7:44
Justin: People with ALS, uh, cerebral palsy, Parkinson's, Down syndrome, uh, or stammer were often simply not understood.
7:52
Justin: So in 2019, Google started a research effort called Project Euphonia, working with community organizations that gathered speech data from more than 2,000 people with non-standard speech patterns.
8:04
Justin: And Google reports that personalized models trained on those voices cut median word error rates by more than 80% on average.
8:12
Justin: That work became an Android app called Project Relate, which transcribes a person's speech in real time or repeats it in clear, synthesized voice.
8:22
Justin: And Google has since uh stopped taking new users, but people who are already trained or who already trained a model can keep using it.
8:30
Justin: So those are Google's own numbers about Google's own product.
8:34
Justin: But there is an independent study that tells the same story but a bit more soberly.
8:39
Justin: At the University of Illinois, the Speech Accessibility Project tested a standard speech recognizer, and for typical speakers, it produced about a 3% error rate, meaning it got about three words wrong out of every 100 spoken.
8:53
Justin: And for people whose speech is affected by Parkinson's disease, it produced about a 36% error rate.
8:59
Justin: And then after training on recordings from about 200 people with Parkinson's, the error rate fell to about 24%.
9:05
Justin: So both storylines are important there.
9:08
Justin: The speech recognizer did get much better.
9:11
Justin: And it still gets about one word wrong out of every four spoken.
9:15
Justin: So that's progress, uh, but it's far from a finish line.
9:19
Justin: And what made it improve was that people with disabilities had a seat at the table.
9:23
Justin: So their voices, uh, given with consent were the training data.
9:27
Justin: And now an example involving the same kind of technology uh that has been alleged to cause harm.
9:32
Justin: In March 2025, the American Civil Liberties Union, or the ACLU, and several partners filed a complaint with the Colorado Civil Rights Division and the Federal Equal Employment Opportunity Commission, or the EEOC.
9:45
Justin: It was filed on behalf of a woman identified as DK.
9:49
Justin: She is deaf and indigenous and had worked uh seasonally at Intuit since 2019.
9:55
Justin: So when she applied for a promotion, the complaint says that she had to complete an automated video interview.
10:02
Justin: She asked for human-generated captioning, and according to the complaint, that request was denied, and she was told the software had its own subtitles.
10:10
Justin: She didn't get the promotion, and the complaint says that the automated feedback she received suggested that she should, quote, practice active listening, end quote.
10:20
Justin: The complaint points to research research showing automated speech recognition performs far worse for deaf speakers.
10:28
Justin: These are allegations.
10:30
Justin: Intuit and higher view both say that the complaint is without merit and the case is ongoing.
10:35
Justin: But consider what that feedback would mean if it's found to be accurate.
10:39
Justin: A machine telling a deaf woman to listen better and potentially costing her a promotion because of it.
10:45
Justin: Same technology, different outcomes.
10:48
Justin: So in the first case, uh speech recognition was trained on speech patterns of people with speech disabilities so that it could understand them.
10:56
Justin: And in the second, it was alleged that speech recognition was used to grade a disabled speaker and potentially cost her a promotion.
11:07
Justin: And that's all based on a system that was proved to work far worse for speakers like DK.
11:13
Justin: So moving into the second pair that I have prepared, I'd like to share two examples of AI in decisions about benefits and health coverage.
11:23
Justin: Decisions that can determine whether someone even has access to care at all.
11:28
Justin: So when someone applies for Social Security disability benefits, the weight is usually measured in months.
11:34
Justin: And since 2008, the Social Security Administration has used a computer model called Quick Disability Determinations.
11:41
Justin: It looks at new applic applications filed electronically and picks out the ones where approval is very likely and the medical records are easy to get.
11:50
Justin: Those go to an experienced examiner within one working day.
11:54
Justin: And what I find interesting is how it's designed.
11:57
Justin: So the model can only move an application to the front of the line.
12:01
Justin: It can't deny anyone, and a person still makes every decision using the same rules as any other case.
12:08
Justin: If the model misses someone, that application goes through the regular process with the usual right to appeal.
12:14
Justin: The agency's own research covering 2015 to 2020 found that half of the flagged cases were decided within about two weeks, compared with about three months for everyone else.
12:26
Justin: Those are the agency's own numbers, and it hasn't published an outside review of who the model tends to miss.
12:31
Justin: But being missed means waiting longer, not losing benefits.
12:36
Justin: So that's that's the treaty here.
12:38
Justin: And now the trick.
12:39
Justin: An example involving a similar kind of tool that has been alleged to cause harm.
12:44
Justin: Medicare advantage plans are private plans that provide Medicare coverage, and they have to approve a stay in a skilled nursing facility.
12:53
Justin: So somewhere that someone might go to recover from, say, a stroke or a fall or surgery of some sort.
13:00
Justin: So in June of 2026, the federal government's healthcare watchdog, which is the office of inspector general at the Department of Health and Human Services, reviewed those approval decisions.
13:12
Justin: One company called Navi Health, which is owned by United Health Group, handled about half of those requests.
13:19
Justin: When patients appealed its denials, the plans overturned 97% of them, but fewer than one in five denials were ever appealed.
13:28
Justin: The watchdog said that that raises concerns about the denials that no one challenged, and I agree with them wholeheartedly.
13:35
Justin: Its report doesn't say an algorithm caused those denials.
13:39
Justin: However, a lawsuit filed in 2023 by the family of a man named Gene Logan and others does make that claim.
13:48
Justin: It alleges that United Health used an AI tool called NHPred, which estimates how long a patient should need care to cut off coverage in place of doctor's judgments.
14:00
Justin: The complaint says that Mr.
14:01
Justin: Logan's coverage ended after about two and a half weeks and that his family paid for his care themselves until he unfortunately passed away about a year later.
14:09
Justin: United Health says the tool is just a guide and it's not used to make coverage decisions and that the lawsuit has no merit.
14:16
Justin: Earlier this year, the court ordered the company to share how the tool works.
14:21
Justin: And I'll say it again that these are allegations and no court has decided them yet.
14:26
Justin: But again, we see the same kind of tool producing different outcomes.
14:30
Justin: Social Securities model can only say yes faster, and a person still decides every case.
14:36
Justin: But in the second, it's alleged that a prediction about a typical recovery was applied to a person whose recovery was not typical, and the main safety net was an appeal form that most people never file.
14:47
Justin: And that leads me to my third and final pair of the evening.
14:52
Justin: Again, I have two examples of AI, but this time two that describe the visual world for people who are blind or have low vision.
15:02
Justin: Be My Eyes is our treat here as a free app that connects people who are blind or have low vision with cited volunteers over a video call.
15:11
Justin: In 2023, it added an AI feature called Be My AI.
15:16
Justin: So how it works is that someone takes a photo, say, of their mail or label or a menu and they ask questions about it, and the AI describes what's there.
15:27
Justin: Before releasing it widely, the company tested it initially with a small group of users who are blind, and then with hundreds more afterwards.
15:36
Justin: So what stands out to me here is that the AI didn't replace the people.
15:39
Justin: If the AI isn't enough or sufficient, then a volunteer is still one tap away.
15:45
Justin: And so anyone who has interacted with AI knows that it can be very confidently wrong, and the person using it may have no way to check.
15:54
Justin: So for the things that really matter, you know, say a medication label or a legal notice, then a human stays within reach.
16:02
Justin: And the person using Be My AI decides when to call them in, what to ask, and whether to trust the answer.
16:08
Justin: And now the trick.
16:10
Justin: An example involving the same kind of technology that has caused harm.
16:14
Justin: A company called Accessibility sold an AI plugin to website owners, and it was marketed as being able to make any website meet international accessibility standards automatically.
16:25
Justin: People who are blind and accessibility professionals have been raising concerns about tools like this for years.
16:31
Justin: And in 2021, the National Federation of the Blind uh formally called on these companies to stop making misleading claims.
16:40
Justin: In April of 2025, the Federal Trade Commission, or the FTC, finalized a settlement with Accessibil.
16:47
Justin: The FTC alleged that the company's claims were false or unsupported, and that it had dressed up paid reviews to look independent.
16:54
Justin: Accessibility agreed to pay $1 million and can no longer make those claims without evidence.
17:00
Justin: And it settled all of this without admitting any wrongdoing.
17:04
Justin: So again, same kind of technology, producing different outcomes.
17:08
Justin: Be My AI is chosen by the person using it and backed by a human.
17:13
Justin: The website plugin was chosen by the website owner, and the person who is blind trying to use that site had no say.
17:21
Justin: And according to the FTC, the promises made on that person's behalf weren't backed by evidence.
17:26
Justin: So when looking at these three pairings side by side, a few things stand out to me personally.
17:31
Justin: So our treats or our examples of tools that helped either learned from people with disabilities or it left them in control.
17:39
Justin: In each, there was a person to turn to if and when the machine fell short, and the people behind them were open about the limitations.
17:47
Justin: Our tricks or our examples of tools that caused harm made decisions about people rather than for them.
17:54
Justin: They were built around what's typical, and the human route was missing or hard to reach.
18:00
Justin: And they made promises that were not verified by the people that they affected.
18:05
Justin: So what strikes me the most is that in each harmful case, the problem could likely have been caught earlier.
18:11
Justin: Someone who is deaf could have tested higher views interview software.
18:15
Justin: Someone who uses a screen reader could have tried Accessives plug-in.
18:18
Justin: Someone could have looked at how often people recovering in nursing facilities were being denied before a federal watchdog had to.
18:26
Justin: So asking wasn't hard.
18:28
Justin: It just didn't happen.
18:30
Justin: At least, not of the people best placed to answer that question.
18:34
Justin: Now I did want to spend some time talking about my views on how lived disability experience can be brought into AI in practice.
18:40
Justin: And I see two places where it matters the most.
18:43
Justin: That would be in the data that an AI learns from and in how it's tested before and after it reaches people.
18:50
Justin: So the first is data annotation.
18:52
Justin: Before most AI systems can learn anything, people need to label the data and they write down which words were said in a recording or they describe what's in a photo, or they mark whether a comment is offensive or an answer is correct.
19:06
Justin: Those labels become the system's sense of what's normal and what's right.
19:11
Justin: So when the people doing the labeling do not possess experience of living with a disability, then the machine has no way of accounting for those types of users.
19:19
Justin: And from the examples above, uh harm occurred as a result of people with disabilities not being included in the design of the system or the tool.
19:27
Justin: Conversely, the more helpful examples included uh said experiences like the speech accessibility project, or they refrained from allowing the AI to make consequential decisions like the quick disability uh determinations model.
19:41
Justin: The second part is evaluation, uh testing how a system actually behaves for real users.
19:47
Justin: So in my opinion, uh, this is what good testing looks like.
19:50
Justin: Uh first, can people actually use it?
19:53
Justin: Sounds like common sense, and that's because it is.
19:57
Justin: Uh not, you know, does it pass a checklist, but can someone perceive it, operate it, understand it, and complete the task?
20:05
Justin: Uh two, what does it cost a person when it goes wrong?
20:09
Justin: So finishing isn't always the whole story.
20:12
Justin: Uh how much extra effort did it take somebody?
20:15
Justin: Uh did someone need help that they shouldn't have needed?
20:18
Justin: Did they keep their dignity throughout the entire interaction?
20:21
Justin: Uh three is you know, who carries the errors?
20:26
Justin: So a system can look accurate overall while its mistakes keep landing on the same people.
20:30
Justin: Statistics and averages can hide individuals, uh, but for the person who can't use the product, uh it's simply a failure.
20:38
Justin: Four, is it dependable enough for what's at stake?
20:42
Justin: A slip in a music recommendation by an AI is very different from a slip in a benefits decision.
20:48
Justin: And five is when it gets something wrong, can a person understand why, challenge it, and reach another human?
20:57
Justin: Those questions are best answered by people with disabilities, matched to what the task actually asks of someone.
21:03
Justin: Something cannot be considered safe unless it has been tested.
21:07
Justin: And a group of testers doesn't speak for everyone with a disability, of course.
21:11
Justin: Uh evaluators find is a sign of risk, not the final, not the final word.
21:15
Justin: And one last practical point uh companies sometimes ask people with disabilities to label data or test AI systems.
21:22
Justin: And when that happens, I just wanted to offer my take on what I think evaluators should be offered.
21:28
Justin: Uh a few points here.
21:30
Justin: So one, being paid for all the time involved, including training, ideally by the hour, not per task.
21:37
Justin: Uh task that takes 12 minutes with a mouse can take 30 minutes with a switch device or screen reader.
21:42
Justin: And that difference is part of what's being learned right now.
21:45
Justin: So it shouldn't come out of uh an evaluator's pocket.
21:48
Justin: Uh two, accommodations like interpreters captioning uh in alternate forms are paid for by the organization, not taken out of the evaluator's compensation.
21:59
Justin: Three, pay that doesn't depend on what the person finds or how they label.
22:04
Justin: Insights should be genuine.
22:07
Justin: Point four, personal and disability information should be kept private unless the person agrees otherwise or the information is uh legally mandated to be shared.
22:18
Justin: And point five is being free to pause, take a break, or outright discontinue an engagement at any time.
22:25
Justin: Uh at some point, I feel we started to put corporate well-being over personal well-being, and I feel that needs to be reversed.
22:32
Justin: So as I start to wrap up here, um, I want to bring this down to everyday life.
22:38
Justin: Uh, most of us will never see how an AI system was built or tested.
22:43
Justin: Uh, but there are signs that I've learned to look for that say a lot about whether a company or a product takes accessibility seriously.
22:50
Justin: And none of them is proof on its own, uh, but together they can start to tell a story.
22:54
Justin: Uh the first would be honesty about AI and a real way to reach another human.
23:00
Justin: Uh so a company that says plainly when a chatbot or an automated system is involved and how to get past it to a human is thinking about the people on the other end.
23:10
Justin: So if there is only one path, and that is through the automated tool, then that's worth noting.
23:16
Justin: Uh, two would be a clear way to ask for an accommodation and how to fix mistakes.
23:23
Justin: So when an automated step is part of a job application or a benefits form, there should be an obvious way to ask for an accommodation and to correct or appeal a decision.
23:33
Justin: If those are hard to find, then the system probably wasn't built uh with everyone in mind.
23:39
Justin: Three is an accessibility statement that is specific.
23:43
Justin: So, in my opinion, uh a good one should name the standard that it aims for.
23:48
Justin: It should say when it was last reviewed, it should admit any known problems, and it should offer a way to report new ones.
23:58
Justin: Uh four is kind of along the same lines.
24:00
Justin: Uh a vague promise or a claim of being, quote, fully compliant, end quote, uh, with no detail, can say a lot without really saying a lot.
24:09
Justin: So, for instance, a widget that promises to fix everything, uh, maybe an accessibility button that claims to make a whole website accessible automatically, uh, those to me are reasons for caution rather than reassurance.
24:22
Justin: Um, I think this is five.
24:25
Justin: Uh whether it works with the tools uh that people already use.
24:28
Justin: So, can the page be navigated with a keyboard alone?
24:32
Justin: Uh, do videos have captions?
24:34
Justin: Does it work with a screen reader uh or with large text?
24:38
Justin: And some of those initial uh digital interface accessibility tests can be a good clue as to the amount of thought uh that has been given to the overall accountability and human impact of an AI process uh for a diverse user group.
24:53
Justin: And lastly, uh who they say has tested the product.
24:56
Justin: Companies that have involved people with disabilities in the uh development process usually say so, and they share how.
25:04
Justin: So this is something that we see NOMA being able to contribute to and even make it standard of practice.
25:10
Justin: Our documentation is designed to relay this information to executive boards, uh, lawyers, and everyday users alike.
25:19
Justin: And we think that diligent efforts made in good faith should be expressed and celebrated.
25:25
Justin: So noticing is a start.
25:28
Justin: And before I close, I'd like to share two things that I think uh we can do with what we notice individually and collectively.
25:35
Justin: So the first is to report what doesn't work and to keep a record.
25:40
Justin: Uh when an AI system or website gets in the way, telling the company in writing, describing what happened and what was needed, does two things.
25:48
Justin: It gives them a chance to fix it while simultaneously creating a record.
25:52
Justin: DK's complaint from our example uh earlier points to concerns that she had raised with her employer's accessibility team a year earlier.
26:02
Justin: Sharing these experiences with disability-led organizations can help too.
26:07
Justin: Problems that are easy to dismiss one at a time are much harder to dismiss when they're bound together.
26:14
Justin: The second is to ask the question before a system is chosen, not after.
26:20
Justin: Many of us have a voice somewhere along the line in the decision process, um, whether that's you know in your workplace, uh, at a school, a community group, uh, a local government meeting.
26:32
Justin: And when a new AI tool is being considered there, one of the simplest and most powerful questions is was it tested by people with diverse experiences, including disabilities, and can we see what they found?
26:45
Justin: And when people with disabilities are invited to help test systems or label data, taking part on fair terms is one of the most direct ways to influence what gets built.
26:54
Justin: So I started with the idea that the same technology can be a door or it can be a wall.
26:59
Justin: And in each of these pairs that I shared earlier, uh the door opened because people with disabilities took part or stayed in control.
27:07
Justin: And each wall came to light because someone spoke up.
27:10
Justin: That's what I hope uh you all will take away from this presentation.
27:14
Justin: We cannot accept an automated system's fairness on faith.
27:18
Justin: They need to be evaluated by the people that they are meant to serve.
27:22
Justin: Thank you.
27:25
Eileen: Thank you so much, Justin.
27:27
Eileen: That was that was amazing.
27:28
Eileen: I really loved the way you presented the pairs, especially, because oftentimes if we look up how do we test if a website's accessible, it's just like a pros and cons list or just a straight list, but really understanding how the same kind of technology can either be a step forward or a major step back, and how those consequences can start small but really be amplified when it comes to healthcare or other really serious, not always life or death per se, but really serious um topics and things that could honestly be avoided or minimized just by including people with disabilities in the conversations throughout and making more informed choices.
28:09
Eileen: So thank you so much.
28:10
Eileen: Um I invite anyone who's on the call if you want to turn your camera on.
28:14
Eileen: If not, that's okay.
28:15
Eileen: We're just gonna transition into a bit of open discussion here, just starting with if anyone has any initial thoughts or reactions to all of the wonderful content Justin just shared.
28:29
Emma: Hi, so my name is Emma.
28:31
Emma: Um, I just want to say thank you for like sharing all of that info, especially at the end.
28:36
Emma: Um so I work in research ops at my company, and I'm still I'm new to the world of accessibility, and I'm still trying to understand um the best way of you know making it easier for researchers and designers when they're doing testing, doing usability tests.
28:54
Emma: So I really found like what you said helpful, and I'm sure like we could definitely apply this into like our principles, especially, and having this like awareness is really definitely like such a different color for me.
29:10
Emma: So thank you.
29:13
Justin: Yeah, I appreciate that, Emma.
29:15
Justin: Um something, you know, as I was listening to you mention that, you know, some of the things that you're you're working through um at your business, I I think that the regulations um, you know, and and standards we place around things, they they definitely are of benefit.
29:32
Justin: Um, but an adverse effect that they might have is I think that they complicate some really simple circumstances.
29:39
Justin: And you know when when we bring in compliance, um that's a a big scary word.
29:46
Justin: And I think at the end of the day, what we're what we're getting at, um, the spirit of the law, if you will, is are you making a good faith effort um at doing what ought to be done?
29:59
Justin: So yeah, accessibility really comes down to is it accessible?
30:04
Justin: And I think how you answer that question is you ask the people that would have the most um barriers to entry, right?
30:12
Justin: So it's I think it's just um you know who's in the room when we're when we're having these conversations.
30:19
Emma: Yeah, thank you.
30:21
Taylor: I want to talk to you, Justin, but at the same time, I don't want to keep anyone else from chiming in if they want to.
30:29
Taylor: Uh, but I just want to thank you.
30:31
Taylor: That was uh really enlightening, even for someone like myself who's been in this space for almost eight or nine years now, and something I learned it's a lot easier to showcase different topics when you put them in context.
30:46
Taylor: So I really appreciate you putting all these different things in context so we can better understand them and the impact they have, like real life impact and not just theoretical impact.
30:58
Taylor: Uh, because as a person who identifies as blind and low vision, I encounter a lot of this stuff as well.
31:06
Taylor: And I like how you talked about AI and like the hiring process.
31:11
Taylor: And that's something I've seen happen a lot, and it's something I really wish we move away from for the reasons you pointed out, but also I feel like it's it's reducing applicants down to a number.
31:27
Taylor: And people with disabilities already have a hard time as it is finding employment, and when you cut out the human side of that process, it can make it even more challenging for people with disabilities to get in that door and at least get the ball rolling.
31:45
Taylor: And so I I I love how you uh pointed that out, and uh, hopefully that's something that we'll move away from, but I it's out of the box box now, so I'm not sure if it will, but like you were saying, it just needs to be better policed in the earlier stages of development.
32:03
Justin: Yeah, I I I definitely share your concerns, and I I want to go back to the analogy of the door versus the wall, right?
32:13
Justin: So something I I really liked about showing the pairs was it it really doesn't become a binary, like good, bad.
32:22
Justin: Um, it does really show the nuances that are at play in all this stuff.
32:26
Justin: It's it's really complex.
32:28
Justin: Like there are a lot of angles to this, and that just kind of speaks to the nature of community, right?
32:34
Justin: Like there's a lot of different perspectives and experiences to consider um at any given time.
32:39
Justin: So I think that with AI hiring, um yeah, it we're we're seeing it cause real harm.
32:45
Justin: Like there are there are a couple different lawsuits I I probably could have cited.
32:49
Justin: And it's just we're not at the helpful point yet.
32:53
Justin: Um, and part of that is continuing to include the diverse perspectives, right?
32:59
Justin: Because I really do think it has real um door potential.
33:03
Justin: It can open opportunities for people.
33:05
Justin: It's just that right now it's acting more as a wall.
33:08
Justin: So we need to get the voices in the room.
33:11
Justin: Um, that's something that we're trying to do at NOMA.
33:14
Justin: But yeah, I I hear you.
33:15
Justin: Um there's there's a lot more work to be done there.
33:19
Eileen: A comment I have there with AI and hiring, and there's really been a lot of pushback, not just for people with disabilities, but really just in general, taking that human aspect out of an already grueling and more complex by the day process of just looking for employment.
33:35
Eileen: Um and I think that's possibly a great example of like the curb cut effect in our time and being able to see how the advocacy for making a process more accessible and maybe moving away from AI in the hiring process or at least like in making those critical decisions is something that'll not just benefit individuals' disabilities, but really everyone.
33:55
Eileen: And I liked your point earlier, Justin, about how if you are going to use AI as a virtual assistant as a chatbot feature, that there are not necessarily safeguards, but there's an opportunity to talk to a human, get to a human quicker, and it's not a whole hassle.
34:13
Eileen: Like, for instance, I was just on CVS, like calling CVS the other day to get a prescription refilled, and they had to contact my doctors, this whole back and forth.
34:21
Eileen: And I literally was talking to this AI assistant for so long, and they kept being like, You have nothing to refill.
34:26
Eileen: And I was like, Can I talk to a human?
34:27
Eileen: And they're like, You have nothing to refill.
34:29
Eileen: And I was like, That's the point.
34:30
Eileen: Like, I need to have them do this and healthcare in America, whole thing.
34:34
Eileen: Um, but I really think that was just a good example of having this really benefit just a larger population.
34:41
Eileen: And I think that um there's yeah, like you said, potential for it to be a door and not just um for individual disabilities, but really for everyone to benefit.
34:50
Justin: Yep.
34:51
Justin: Yep, could not agree more.
34:54
Taylor: Yeah, I like how y'all are both talking about it being a door because I fully agree.
34:58
Taylor: And like you were saying earlier, there are a lot of nuances to this topic and really accessibility as a whole.
35:05
Taylor: So looking at it through that perspective, we've talked about the tricks and treats of green flags and red flags.
35:13
Taylor: What is your overall stance on AI in general?
35:18
Taylor: Are you fully pro I wouldn't say fully pro or fully con, because I'm sure you have kind of a love-hate relationship like we all do.
35:26
Taylor: Um, but what is your like current stance on AI in general and where it's moving?
35:33
Justin: Um, I would say cautiously optimistic.
35:36
Justin: Um I think that technology, um, much like life, moves in cycles.
35:43
Justin: And the technology is something new, but the the process is not.
35:50
Justin: Right.
35:51
Justin: So we we have seen this with the internet.
35:53
Justin: You know, there's an explosion in techno in technological uh advancement and capabilities.
35:59
Justin: And some groups get left behind, unfortunately.
36:03
Justin: And so that's that's kind of our our stance now is um, you know, let's let's learn from the mistakes of the past when we've gone through these technological rollouts and let's let's move trust uh upstream, right?
36:16
Justin: Let's get the voices in the room upstream in development.
36:19
Justin: Because it's it's much harder to try to fix problems downstream after harm has been caused.
36:27
Justin: And and then you're doing it in a bolt-on fashion, if you will.
36:33
Justin: But the system is already the system, right?
36:36
Justin: So that's our stance.
36:37
Justin: Is we need to we need to move it upstream.
36:39
Justin: We know what happens when it doesn't.
36:41
Justin: Um the snowball gets bigger and bigger as it rolls downhill.
36:45
Justin: Said this earlier, but there there's a lot of work to do.
36:48
Justin: And I think um there's a lot that we can learn if we're willing to look back with an honest uh perspective on some uh instances where we've missed the mark in similar cycles.
37:01
Justin: So cautiously optimistic.
37:05
Taylor: Yeah, I'd say I'm I'm pretty similar to you on that sand.
37:11
Justin: Yeah, I I I feel like um kind of have to be, right?
37:15
Justin: I think uh the optimism allows you to um move without dread, and the caution prompts you to do the work.
37:26
Justin: The the skeptic is um aware that it doesn't get better without contribution, but that contribution from a fear state is not always the most helpful.
37:40
Eileen: That's a great way of looking at it.
37:42
Eileen: I've been asking myself some of those questions recently how about how do I want to define my position on AI, not just externally to other people, but personally, like how do I truly feel about it?
37:53
Eileen: And I could go on and on.
37:55
Eileen: I mean, I my background's in like anthropology, so it's it's very easy for me to get down that sort of like ethical side of things, but I think with the both the caution and the optimism, being able to really get work done and recognize the benefits and then also constraints and ways to um be critical when needed, but not have that limit progress and opportunity.
38:17
Justin: Yes, ma'am.
38:19
Eileen: I'm curious if either you, Justin, or anyone else who's with us right now, if you have any AI tools or extensions that have been working surprisingly well, that are accessible, that you want to maybe shout out or um recommend to other people.
38:39
Taylor: Uh for for me, not necessarily a certain product, but a certain type of AI.
38:46
Taylor: I find that I'm often getting misled or fed false information or even sometimes lied to.
38:54
Taylor: I feel like it's has no shame when used the voice mode of certain AI tools.
39:01
Taylor: I feel like it's much easier for me as the user to control the output of the AI when I'm in like text mode because I can have more control over the prompt I give it rather than just speaking that prompt.
39:17
Taylor: Because I feel like it's a lot can be lost in translation when you're doing that.
39:21
Taylor: Like Justin was pointing out when you're doing the dictation, a lot of errors can happen and things can get misinterpreted, and it'll fill gaps that it shouldn't be filling.
39:34
Taylor: Like if it's unclear on something, I always tell it like ask me, but it's I feel like I have to constantly tell it what I need and it's it's not listening.
39:44
Taylor: So I I think what Justin and you all are doing on getting people with disabilities involved on the back end is is amazing because for a lot of different reasons, and you had mentioned it a little bit ago, it's much easier and a lot less expensive to get ahead of these issues early on rather than get to the end of the road and be like, oh, we have all these issues, now you have to hire people and train people on how to fix these issues, and that's a waste of time, reason and resources that you could have gotten ahead of by just involving people with disabilities much earlier on.
40:24
Justin: Yep, absolutely.
40:26
Justin: I think someone I would point out as um doing some really interesting work are uh, you know, there's some folks at Microsoft that are working on um sign language AI that can interpret in real time as somebody is uh dictating and transform it into the various uh signs for uh ASL and for other uh dialects.
40:53
Justin: I'll also call out uh Calvin Young, who is working in a similar space, and also he is trying to help the deaf community um find ways to adopt AI that are the most beneficial um to given situations.
41:10
Justin: And that's something that is really uh fascinating to me is uh you know taking speech in real time and and um translating it into uh ASL and other uh dialects of sign language.
41:23
Justin: So yeah, I I wanted to call that out.
41:25
Justin: That's something that uh has been on my radar recently.
41:29
Justin: And that's uh, you know, that's that's a good example of trying to meet people where they're at, right?
41:36
Justin: Not just uh thinking of translation from spoken language to spoken language, but uh into different forms of communication.
41:44
Justin: So a good example again of of people really getting outside of their box, so to speak.
41:54
Eileen: And then um Torin, who's um one of the founders, correct, of of Noma?
42:00
Justin: Yeah, he's our founder.
42:02
Eileen: That's right, that's right.
42:02
Eileen: Sorry.
42:03
Eileen: Uh Torin just hopped off, but he put in the chat before he left since he was on the move, uh, shouting out a uh, I believe it's a wearable called Hapwear.
42:13
Eileen: And so it's an opportunity for people who are blind or low vision to experience nonverbal communication.
42:19
Eileen: I'm just reading the brief description you gave.
42:22
Eileen: I'm not super familiar, but that sounds really exciting with um the ability to help share some of that um, as I know that's an ongoing challenge, and you know, how can we address this and what what can we use technology for good and not uh yes, just like putting it to good use.
42:38
Eileen: So I mean that's those are all great examples.
42:41
Eileen: And um I'm curious if also and Justin, or feel free anyone else to um jump in on this.
42:50
Eileen: But you you mentioned you know some of the red flags that kind of come across quickly, whether it be just quick visual or keyboard checks.
43:00
Eileen: But do you think there's any um maybe in the development of Noma's methodology, you've maybe run into any really like hidden red flags, things that seem normal but are actually really um uh impactful in the ways that they're preventing accessibility of AI products or services?
43:21
Justin: Yeah, so the first thing that that comes to my mind is you know, it's not something that is uh is kind of top of mind, I would say.
43:32
Justin: Um but thinking about maybe folks that are um neur have neurological differences, um, you know, maybe uh ADHD or autism.
43:43
Justin: And if you're working through a an AI process, or really any process, the ability to pause when needed and not lose your place or and completely have to restart the process is a really Really, really big deal.
44:01
Justin: And that is something that's becoming more and more uh present and prevalent given security concerns, right?
44:09
Justin: So they don't want, and this makes sense to me, right?
44:13
Justin: You you don't want to leave too much time because then there's room for um you know maybe somebody who's not the uh applicant or the the one who should be filling out the form.
44:25
Justin: So it's a security concern, but the unintended consequence is that you know, if if somebody has to stop for really anything, right?
44:33
Justin: You have to go pick up your kid at school, all the way to you know just being fatigued and needing a break before you complete it.
44:39
Justin: You know, if you're if you're getting kicked out and having to constantly restart, that's a huge barrier to accessibility.
44:46
Justin: And there's a a good likelihood that somebody who has to keep starting over is is gonna give up.
44:54
Justin: Right.
44:54
Justin: So you're alienating a large um group of people that could be in need of the service or have a desire for the service um that you know can't complete it, may look elsewhere, um, or even worse, just not have access to something that's um that's vital.
45:11
Justin: So um, yeah, that that is something that um again has been kind of top of mind and kind of keeps popping up for me.
45:19
Eileen: That's a great example.
45:21
Eileen: And another example that really highlights how it might come from addressing a need of a specific population or specific disability, but in reality benefits so many different use cases and situations.
45:35
Eileen: And that reminds me, I can't remember who posted it, but I saw something on my LinkedIn feed the other day about how we should maybe reconsider the idea of using the word user testing and have it really be focused on usability testing.
45:48
Eileen: Because sometimes user testing, it's like, can this user use this?
45:52
Eileen: Whereas the idea of usability testing is like, is this functionally available and accessible to the wide audience it should be?
46:01
Eileen: And so that's something that's also been like on my top of mind of recognizing there's so many things I had never considered before, and that's one that's really stuck with me of like, oh yeah, like that makes sense.
46:11
Eileen: Like it should be usability testing and not be on the user to um have the fault or the blame pointed at them, but rather at the people building these systems and products of, hey, it's not working, let's change this, let's make it better.
46:24
Justin: Yeah, thank you for that.
46:26
Justin: Um, that's awesome.
46:27
Justin: I think I'm I'm gonna have to start using that.
46:29
Justin: I myself and am a huge fan of language.
46:33
Justin: I think that the language we use is very important because ultimately we're trying to convey ideas and ways of thinking um and stories, really.
46:42
Justin: So uh yeah, usability testing.
46:45
Justin: I think I think you're right on.
46:48
Justin: And I think um you're right, we should start using that.
46:51
Taylor: For me, I I think something that came to mind when you're talking about those like hidden red flags that may not really be considered like errors on the AI side um is more on the usability side, like you were talking about.
47:08
Taylor: Um very often different AI models give you the option to customize the model and give it different preferences and all these different things, but they'll hide it in like I'm like, well, why have that as an option if you're not gonna present it as an accessible option to do?
47:32
Taylor: Like, I'll I'll try and customize the model because I'll hear like a colleague say you should try it this way, and I'm like, Well, I don't know how to get there.
47:40
Taylor: I'll go to settings and it has like this very complex menu, and it's not very straightforward.
47:47
Taylor: And so like I just find that very frustrating because I feel like I'm able to get more out of these tools than I'm getting, but I just don't know how to take it to that next level because they don't really make it the most accessible to get to.
48:04
Taylor: Yeah.
48:05
Justin: Yeah, you're you're you're totally right.
48:07
Justin: Um yeah, that's that's why usability testing is so important, right?
48:12
Justin: Because you know, if you're you're running out a VPAT or you know, your your own assessment of your product's accessibility um during you know procurement, you can check the box that says, you know, we have accessibility um widgets and customization.
48:29
Justin: But yeah, the the widget or the customization itself might not be accessible.
48:35
Justin: So yeah, it's it's uh usability all the way around.
48:42
Eileen: I wish we well, might we might be able to make the public or the recording public, but we had an internal session recently for the AI Learning Circle talking about text-to-speech and how it started and now how it's evolving with AI and with that customization and and personalization that now oftentimes these AI features are behind paywalls as well.
49:04
Eileen: And so they're able to have these cool features presented, but then they're costing more, and so they're actually being used for things like the I don't know, AI fruit videos or things that are like entertainment and not for people who um are nonverbal or have other difficulties communicating and would really benefit from having a voice that sounds more genuine or authentic to who they are and how they identify.
49:26
Eileen: And so um, I hope maybe I can share that link with you, Justin, or anyone else, um, because I think it's it's very much connected to what we're talking about about something starts out such a good idea, useful tool, and now it's becoming harder and harder to access, um what somewhat ironically, since it began as an accessibility feature.
49:45
Justin: Yeah.
49:45
Justin: Yeah, yeah, please do.
49:47
Justin: Oh, go ahead.
49:48
Justin: I was just gonna say, please do share the video.
49:51
Taylor: Oh yeah, and I was gonna say, I fully agree.
49:54
Taylor: It's unfortunate that that stuff happens because like we were talking about, there's there's a lot into AI.
50:02
Taylor: There's not all good or all bad, but for me, my perspective is AI, when used correctly, can be like the ultimate assistive technology, even if it's not an AI model that was specifically designed to help people with disabilities.
50:19
Taylor: Like for me, I use Microsoft Copilot at least a dozen times a day to spell track documents, uh make my grammar sound better and an email to a client, that kind of stuff.
50:32
Taylor: And like like you were also saying it's behind paywalls to get these certain like use usability limits and stuff like that.
50:40
Taylor: And so it's unfortunate that something that has the potential of being that ultimate assisted technology is becoming less and less financially accessible to people with disabilities.
50:57
Justin: Yeah, I I think that um priorities play a big role in that, right?
51:04
Justin: Um, and that things that start out with good intention can get hijacked.
51:09
Justin: I don't really have an answer.
51:12
Justin: I hope that um we get to one or maybe there's not an answer.
51:15
Justin: Um but yeah, I think that the incentives may need to shift a little bit.
51:22
Justin: The incentives right now for accessibility are mainly around compliance um and you know fines and uh other repercussions for not meeting specific standards.
51:33
Justin: Um, but my hope is that we can get to a point where those are obsolete, maybe, because we are all just uh doing things to benefit our neighbor.
51:47
Justin: And I think that you know, if we can give that a shot, we'll see that benefiting our neighbor actually does come around to benefit us as well.
51:54
Taylor: Um, so I I hope that's tangential, but uh yeah, and and I appreciate you answering it, and I just want to make sure you know that sometimes I might make a point and it's more so just event about it.
52:09
Taylor: Because I certainly don't have all the answers either.
52:12
Taylor: So um yeah, but talking about this kind of stuff is frustrating because again, like I see so much good potential in AI, but I'm just seeing it being used so irresponsibly, and then you have like those paywalls that Aline was talking about, and all the I feel like the tricks are starting to outnumber the treats when it comes to accessibility in in AI, and that's my biggest thing with that.
52:44
Justin: Yeah.
52:44
Justin: Right, there's the the US frontier models lobbying for more um regulation around the open weight models, right?
53:04
Justin: And it's uh it's seems to be more of a competitive move.
53:08
Justin: But I think that you know, if we have more of that open weight structure, then you will see more um innovation for benefit sake, people starting to do really cool things with it um to find ways to better their own situation or the situation of uh loved ones or or people around them.
53:30
Justin: So I I do hope that uh the open weight models don't get um crushed or you know extinguished before they can really be used uh for some good.
53:42
Justin: And I'm I'm I'm not saying that they're entirely used for good.
53:46
Justin: I know that there's um always maybe some bad actors, but yeah, let's not I hope we don't take the tools out of people's hands before we really see what um humanity can can do with it.
54:01
Eileen: I think that's a great note to end on with that uh cautious optimism, as always, and being aware of you know, where can we go from here, and just with our conversation today, and anyone who might be listening in on Thursday or at a later date to really recognize that even just taking away one thing, two things, something to share with someone else, that's what's going to be creating change and making more people aware of these things.
54:24
Eileen: And um, I have to shout out the Accessible Community Podcast again if you want to learn more about these sort of dynamic conversations.
54:31
Eileen: I've truly learned so much from them, and I know I'm a bit biased being on our leadership team, but I think it's seriously such a great resource.
54:39
Eileen: And I really look forward, Justin, to keeping in contact and seeing how NOMA continues to grow and how Accessible Community can continue to be in partnership and support of you all.
54:50
Justin: Yep, we're around.
54:52
Justin: You guys are awesome.
54:52
Justin: Thank you.
54:54
Eileen: All right, thank you so much for joining us.
54:55
Eileen: We're gonna end the recording here.