Federal Agency and Industry Practitioner Hearing on Artificial Intelligence
>> The time is now 2:00 p.m. Hearing chair Elver Ariza Silva, would you please again the hearing? >> Elver Ariza Silva: Indeed. Welcome to the U.S. Access Board's artificial intelligence virtual hearing for the disability community. My name is Elver Ariza Silva. I am a public member
of the Access Board appointed by President Biden in 2022. I serve as the Vice Chair of the board and I live here in the Washington, D.C. area. The Access Board's mission is to create access role. The Access Board is an independent federal agency dedicated to accessible design for people with disabilities. The board is comprised of 25 individuals, 13 of whom, like me, are appointed by the President. The others are representatives from various federal departments. I would like to acknowledge my fellow board members that have joined today's
hearing and thank them for being here today. I would like to take a moment on behalf of the board to additionally thank all of today's presenters and those in attendance for being with us today as we discuss artificial intelligence in relation to the disability community. A few housekeeping notes as we begin. This hearing is being recorded. American Sign Language interpretation and realtime captions are being provided. The hearing will be posted to our website and YouTube channel in the coming days. All attendees will remain muted and with your cameras off unless you are being called upon to speak. You may use the chat feature
to the host if you need assistance. For all board members, presenters, and those that preregistered to offer public comments, please take the time now to ensure that the name you are listed on Zoom is your full time so that we can easily identify you and provide you with the permissions needed to speak at the hearing. On the screen, you will find the agenda for today's hearing. After my welcoming remarks, we will begin the hearing with Alison Levy, Director of the Office of technical and information services for the Access Board who will provide some foundational background on our artificial intelligence AI work in the Executive Order on artificial intelligence. After Alison, we will hear from a wide range of presenters from federal agencies, as well as industry practitioners on AI and Accessibility. Following the presentations, Access Board members and staff will be able to ask panelists questions. Afterwards, there will be a time for public comments from those
that have preregistered to provide public comments. Let us begin by welcoming Alison Levy. Alison? >> Alison Levy: Thank you, Elver. Good afternoon. I'm happy to share some background information with you, starting with the next slide. For those of you who are not familiar yet with the U.S. Access Board, we have basically three primary roles and responsibilities. First, to establish accessible design, guidelines, and standards
under Architectural Barriers Act and the Americans with Disabilities Act, in addition to section 508 of the Rehabilitation Act, among others. Second, we provide technical assistance and training on all aspects of Accessibility for both the built environments and the digital environment. And third, we enforce Architectural Barriers Act, which applies to the federal government buildings. Next slide. So back last October of 2023, President Biden issued executive order on artificial intelligence. Within that Executive Order, the Access Board was tasked
with a few things to help with accessibility of artificial intelligence. Next slide, please. So our three tasks include the following. First, we were asked to solicit public participation and conduct community engagement to learn a little bit more about what folks are feeling and experiencing about their use of artificial intelligence. Second, we've been asked to issue technical assistance and recommendations on the risks and benefits of AI, including use of biometrics as a data input. And third, we're working to help provide people with disabilities access to information and communication technology, as well as transportation services. Next slide, please. To help us with this endeavor, we partnered with
two national nonprofit disability organizations. They are amazing and they are great team players as we move forward in our communication with the disability community, AI practitioners, as well as other federal agencies. Those two organizations are the Center for Democracy and Technology and the American Association of People With Disabilities, otherwise known as AAPD. We engaged in this memorandum of understanding, otherwise known as an MOU, back
in May to really help us connect with the disability community. Next slide, please. One of the key outcomes of this partnership is that we're working through a series of five web hearing based iterations that include hearings. The first was hosted on July 9 of this year, specifically to the disability community and actually the first one is not listed here. The first one was a level setting. It was basic information about AI to help people with disabilities better understand AI and to level set a basic understanding of this technology.
The second one was two iterations with the disability community. We hosted a morning and an afternoon session on July 9th. Next we're hosting today's session for AI practitioners and federal agencies to share what they know about best practices, pros and cons of artificial intelligence. And finally, our goal is to host our last
iteration in November to share our findings, recommendations on the use of AI and Accessibility for people with disabilities. To look for or to view any of the previous sessions, please visit our new U.S. Access Board artificial intelligence web page. The link is provided on this slide. We'll also pop it in the chat box. But if you visit our home page, just look for the link
to our AI hearing information and you'll find a wealth of resources there that will continue to evolve as we move forward with this effort. Now I'll turn things back over to Elver to introduce our next panelist. Thank you for joining us today and we look forward to continuing to support you in this endeavor. >> Elver Ariz Silva: Thank you,
Alison, for that information. We will now transition to the panelist presentations part of the hearing. Panelists will be sharing perspectives op current research, current [audio skipping] AI and other related AI topics. We ask that all panelists keep their presentation to around eight minutes. They should be prepared to share their screen at the time of their introduction. As a reminder, please keep your cameras off until it's your time to present. We ask that all others remain muted with cameras off.
The first presentation, we now welcome Mr. Zach Whitman, Chief of AI officer of the General Services Administration, GSA, on GSA approaches to AI and Accessibility. Zach, you may begin. >> Zach Whitman: Thank you, everybody. Really appreciate the opportunity to speak with everyone this afternoon. Especially regarding GSA's consideration of AI and how that intersects with our accessibility practices. One convenient thing about my role as chief AI officer is that I'm also the chief data officer and in that structure at GSA, we run the accessibility for 508 program, so we're closely involved in how we can best leverage the latest technologies to improve an accessible feature, not only for our public services, but also for our internal team as well. We're really committed to this and having that synergy between the technology
and the 508 office has been a really beneficial relationship that we've been able to bridge. Now, as we look at AI and its potential future, we see this as a transformative technology for some of the solutions that previously may have been a little difficult or prohibitive in our ability to offer. We're seeing a democratization of these services with these general services and applying them to different applications in the work context, but also in our public servicing offerings. Opens up a wide aperture of potential for us. So we see this opportunity, one, with excitement, but also, one, with caution. We don't want to rush into
nicks too quickly. We want to make sure that the events that we're making are doing progress and not doing any disservices as we roll out the new technology, given its relative infancy. Some of the things we're looking at specifically at GSA are things like realtime captioning and transcription services. We know that AI driven tools that are currently commercially available can do realtime captioning for videos or meetings in live events, but we also know that some of these things don't have the accuracy that would be required for our workplace or to offer to the public, and so taking a measured approach in how we evaluate those services is really critical.
Second would be an AI powered screen reader. This is a new possibility for us to be able to not only offer more context around the presentation on the screen, but also provide insight as to graphical information. We see examples of how AI can start to interpret graphics in more meaningful ways than simply representing that this is a specific type of graphic, but what that graphic means. And then that, again, is an opportunity for us to start to lean into this technology and provide a richer experience. We're also looking at automated image and video descriptions. Now, this, again, is going into how we are going to provide better services for either our video presentations or our more static assets that need to be enriched at scale. We have
a lot of digital infrastructure that needs to be looked at in certain contexts to ensure that the images are properly vetted and categorized in a way that is truly accessible. So this is another efficiency that we could potentially see that would improve the quality of services. We're also making sure that we are considering what it would mean to have predictive text or AI based writing support be made available to our workforce. We understand that a lot of
folks can have challenges in writing without having a cognitive load or motor disability making it harder to create or to author text, and we wanted to make sure that the predictive text and AI writing support can be a value add to those services. Now, when we talk about the AI that is currently available, and there are a number of different applications or we can keep going into them, but one thing I would want to call out is we believe that AI needs to be the AI with regards to accessibility needs to be opinionated. What I mean by that is oftentimes we see standards like WCAG trying to outline three basic ideas.
First one would be what is the issue that it's observing? Where the issue is. And then how do you fix the issue? Now, when we're talking about things like HTML, it's critical that we identify these problems, and AI can be of service in identifying issues, but it doesn't necessarily mean that we have a prescriptive answer as to how best to solve the issue. And that's what I mean by the opinionated concepts need to be present in our AI solutions that are going to be supporting in terms of writing, specifically code in this case, context.
Sometimes the solution might not be quite so obvious as to how to fix the problem to an AI assistant who can identify where the issue might be, but might not understand how best to solve that issue. This is one of the biggest shortcomings we see with accessibility testing tools. The fact that they mostly do a good job of finding these issues, but the exact code is sometimes requires subjectivity to ensure that the quality of that solution is meeting the needs of the public. A good example would be these publicly available
LLNs. They are largely populated and trained on publicly available data, public websites. Now, one thing that we have all come to understand is that the data that's used to train these models is public and, therefore, has a lack of duration potential, so we don't necessarily ingest into the system the absolute best solutions for accessible websites, as an example. We want to make sure that GSA plays a role in improving the corpus of data that these public LOMs are training on by providing high quality accessible solution and code so that the better results could be made available through these general tools. I believe that the government should be a source of delivering high quality accessible contents through our websites. And
so that's one of our main focuses going forward is making sure that we are doing everything we can to improve the quality of our digital experiences. Some of the things we're currently exploring is we're using AI to convert PDF holdings into HTML. We have a lot of PDFs on GSA.gov and we'd like to convert into accessible formats. That will take a series of different products and capabilities out I talked about earlier to do effectively. Not only that, though. We want to make sure that the
quality is increasing as we deliver the services. Also the processing that we're going to be put into place to ensure that there's been quality assurance in the outcome and the output of these automated systems. They are also investigating the meeting transcription services and document summarization. These are, again, in the early days because we want to make a very deliberate decision as to how we move forward and not move too quickly and potentially impart some biases, which we will have constrained without proper testing. So we're taking it very slow. And lastly, looking at the AI visual interpretation tooling. This would be for graphical interfaces that have
charts and maps where an interpretation layer on top of the visualization would be assistive in trying to make sense of what the document is going to say. Also, we're trying alternative routes to gather new ideas. We hosted a hackathon recently which had several accessibility submissions specific to AI that were used to assist folks in trying to either understand content that was on federal websites or to make better submissions or make forms more accessible. In fact, the second place winner of that hackathon was one that took jargon language that was hosted on federal websites and made it into clear, plain language through an AI interpreter. So we're really happy to see new ideas coming forward on that front. And lastly, I'll close with we're making a heavy
investment into the U.S. web design system in terms of making sure that any new advancement we make, especially in the domain of AI in the sense of AI interfaces that are specific to web, are accessible. And also, call out an upcoming accessibility forum hosted by archive that's hosted a couple days, I think actually in September. And it's going to be dealing with specifically AI and Accessibility, which I think is a really exciting event that will be hosted by archive out of Cornell. Anyway, that's my time. I really appreciate everyone's attention. Please let me know if you have any further questions in follow up. Thanks. >> Elver Ariz Silva: Thank you so very much, Zach.
In our second presentation, we would welcome Megan Schuller, legal Director of the Bazelon Center for Mental Health Law on the promises and perils of AI for people with mental health disabilities. Megan, you may begin, please. >> Megan Schuller: Good afternoon. My name is Megan Schuller. I'm the legal Director of the Bazelon Center for Mental Health Law. I use she/her pronouns, white woman in my Forties with mostly blond hair wearing a dark green shirt and gold necklace. I'm coming to you from Acadia national Park, home of the Wabanake nation. Thank you to the Access Board and AAPD for the opportunity to
speak to you all today about the impact of AI on people with mental health disabilities, including both the promise and the perils. The Bazelon Center for Mental Health Law has been fighting for over 50 years to protect and advance the Civil Rights of adults and children with mental health and developmental disabilities and the right to live with autonomy, dignity, and opportunity in welcoming communities supported bylaw, policy, and practices. As mentioned, my focus today is to speak on the impact of AI on people with mental health conditions in particular, including serious mental illness. It is an often overlooked population in the growing discussions of AI policy and regulation. AI, using that term broadly, is now being used for very high stakes decision making from who gets a job or a loan or help in jail or keeps custody of their child. Often with the stated purpose of reducing systemic and unconscious bias. Whatever
you feel about it, AI is impacting everything we do and it's not going away. It's growing and expanding and absent standards and regulations, so is AI bias and what our partners at CDT call technology facilitated discrimination. To make this concrete, I want to talk about pretrial sentencing tools. Courts are now routinely using predictive algorithms to make decisions in courtrooms. Algorithms use pools of information to turn data points into predictions, whether that's for online shopping, hiring workers, or to make bail decisions at the point of arrest. One such popular widespread pretrial sentencing tool gets information from the arrestee, feeds it into a computer algorithm, and that outputs a risk score meant to quantify the likelihood that the person will commit a crime or fail to appear in court. High risk people equals
Yale and low risk equals jail. A widely red pro public an report on such a tool found significant racial bias. It inaccurately predicted that 45% of black arrestees would re offend who did not. While the false positive rate for white was less than 24%. In addition to the horrifying racial bias, note how inaccurate the tool is for everyone on a decision that decides who gets incarcerated, yet these tools are still widely used. Now, to unpack why software that does not know the race of the person would produce such discriminatory and biased results, we should start by looking at the software and how it works, but the courts and judges using these tools generally do not have access to how the software is making its predictions or what the score is based on, because the company that created the tool claims it's proprietary. Think about the due process and constitutional implications of that. Now, despite this black box problem, we do know
the answer to the question of why. The answer is proxies. The tools rely on numerous proxies for race that reflect the societal disparities and institutional racism all around us. And then not only replicates that racism. They increase it. We see similar biases and proxies in the criminal legal system across race and disability. A review of eye few key numbers helps explain why. By one report, people with disabilities account for 30 to 50% of incidents of police use of force. Federal government estimates have found that people in need of mental health support are 20 to 50% of the people shot and killed by police. Black Americans are over three times as likely
as white Americans to be killed by police. And one study found that black people with mental health disabilities were more likely to be incarcerated than any other racial group. Now consider a pretrial sentencing tool that calculates your risk of re offending or fleeing based on factors such as your age at first arrest and prior misdemeanor convictions. Those are going to disproportionately identify black people and people with disabilities as high risk. The same is true of other factors used in these tools due to well documented disparities and biases, both racial and disability based in employment and housing.
Now, the child welfare context provides another concrete example of the risks of these predictive decision making algorithms for people with disabilities. Many child welfare agencies in the United States are considering using these tools to screen reports of neglect and make child custody and placement determinations. In fact, several are already using them. The Allegheny family screening tool used in Pittsburgh, Pennsylvania, and surrounding areas was specifically designed to predict who is likely to have their child removed from their custody in the next two years. Well, based on historical data, the correct answer is BIPOC parents and parents with disabilities, not because their children are at greater risk of abuse or neglect, but because of racial disparities and well documented biases against parents with disabilities in the child welfare system. So by identifying these groups, the software is working correctly. That's the right answer to the question. And now the tool is going to identify all people in those categories
as high risk and increase the likelihood that they'll have their children taken away from them. The problem begins with the question asked. Why are we not asking which children are at greatest risk of abuse and neglect? Well, because the algorithm cannot predict that, and yet it's still being used. Now, to answer the questions presented to it of who's going to have their
children taken from them, the algorithm has used a trove of detailed personal data collected from child welfare history, birth, Medicaid, substance use, mental health, jail, and probation records, among other government datasets. This tool has all the same proxy problems as the pretrial sentencing tools. But this top has included the fact of having a mental health diagnosis as a risk factor. Grouping together a wide range of mental health conditions with no individualized analysis. It also includes the fact of treatment for mental health as another risk factor, and the same for a past history of substance use and treatment. Treatment is specifically held against you. Based
on one ACLU report on this tool, your risk score can increase by three full points on a scale of one to 20 based on the overt disability factors alone. Never find all the proxies. The Americans with Disabilities Act requires the decisions be individualized. It requires that people with disabilities be given an equal opportunity to obtain the same result. Gain the same benefit or achieve the same level of achievement as provided to others. And it prohibits using criteria that tend to screen out people with disabilities. Or methods of administering a program that result in
discrimination. Now, consider how the algorithms we just discussed comport with these requirements. And to be clear, there is no exception in the ADA or other Civil Rights laws or AI. So this brings me back to where we started. Many of the tools discussed have been held up as a way to reduce human bias and disparities in public systems wrought with bias and discrimination. AI is not going anywhere. How do we address and reduce the perils while also pursuing and realizing their promise? Some of the tools that pose the greatest threat also presents the greatest promise. What
if we could actually use them to reduce bias and discrimination in the child welfare system and in public benefits? We must first understand the very real impact of these tools, stop them from being used behind a curtain and with impunity to strip BIPOC and disabled communities of their rights. And then ensure the people most impact are involved in implementing and developing these stools with solutions. If we are truly to use AI for good, to be clear, this is not a context where we should be moving fast and break things. Those things are people's lives and their families. But
if an algorithm was carefully and thoughtfully developed, deployed and implemented by and with impacted communities, black and brown communities, People With Disabilities, LGBTQIA+ communities, to actually reduce bias in these high stakes systems, that is a world worth imagining. Thank you for the opportunity to speak to you today. >> Elver Ariz Silva: Thank you so much, Megan. Our third presentation, we are going to hear from Nathan Cunningham, senior policy advisor of the Office of Disability Employment Policy within the U.S. Department of Labor and for their project manager of
the partnership on Employment and Accessibility Technology. PEAT. Nathan, you may begin, please. >> Nathan Cunningham: Thank you so much. Good morning, everyone, from Seattle, Washington. My name is Nathan Cunningham. I am a white man in my early Thirties wearing a blue plaid shirt
and a Blazer and I use he/him pronouns. It's a pleasure to give remarks at the U.S. Access Board hearing on behalf of the U.S. Department of Labor. I am a Senior Policy Adviser in the department's Office of Disability Employment Policy or ODEP. Our Assistant Secretary, Taren Williams, is the head of my agency is a federal member of the U.S.
Access Board and through our collaboration, our agencies are able to advance our shared mission to promote equity and access for people with disabilities. So thank you for inviting me and for all the panelist's remarks today. I also want to say I really appreciated the weather of knowledge that disability advocates shared during the hearings on August 8th. Definitely learned a lot. I have low vision and I believe it's necessary for those of us with disabilities to make sure that new technologies work for us as well. So I'm glad that these voices are involved in the hearings. A70 a prime example of this. So my agency, ODEP,
influences federal and state policies to make AI fair and inclusive in the context of employment. We create resources for employers, AI experts, and workers to learn how to use AI in inclusive ways. And whoa look at AI from two main perspectives. So first, as the previous panelist described, AI holds extraordinary promise and potential, and in our case, for improving opportunity and access for workers. It can even enhance accessibility and support reasonable accommodations. Think
of applications like computer vision or meeting transcription, as one of the panelists discussed. However, there's a second part of this angle. Depending on how people develop and use AI, this technology runs the risk of undermining labor rights and causing bias and discrimination against disabled jobseekers and workers. At the Department of Labor, we are committed to empowering our nation's jobseekers and workers. So protecting worker rights and well being helps make employment safer, healthier, and more inclusive. These goals are critical during a time of rapid
innovation in the case of AI, when advanced technologies are reshaping how people work. Earlier this year, the Department of Labor released a set of principles for developers and employers to promote worker well being when they use AI. These principles offer a North Star for inclusive AI governance. And here with the Access Board, we know from the world of Accessibility that robust governance efforts are critical to help people follow policies like section 508 of the Rehabilitation Act. And I will say we need similar guidance to help people understand and follow inclusive AI practices. So a similar standard that
lays out people's responsibilities to enact inclusive AI policies at an organization, whether from an employer side, vendor side, or even worker rights. New tools powered by AI are in the spotlight, because they can impact so many aspects of work. AI can help people carry out job tasks, take meeting notes, provide chat bot services, process large datasets, scan applicant resumes, and even match people to jobs. Many of us are encountering AI all the time without realizing it. And that's part of the problem here. Many of us have talked about transparency as a key issue. Transparency is one of the central pillars of
inclusive AI governance. So jobseekers and workers need to know when AI is present, how it affects them, what accommodations they can request, and even how they can opt out. Being Able to call in a real person or a human for oversight and assistance is not a separate function. This is a key element of inclusive AI systems. AI is technical, but also very social. People build AI systems and decide how to use them. Unfortunately, the people most at risk from these
systems are often left out of the conversation. But not today. At ODEP we address this issue head on. Our mission is to develop policies and practices that increase the number and quality of employment opportunities for people with disabilities. AI is one issue that crosses many of our policy areas, from education to employment to workforce training. Through this work we fund an initiative called the partnership on Employment and Accessibility Technology or PEAT. PEAT works
toward a future where all technology is born accessible so disabled workers can succeed in their careers. And for many years, PEAT has influenced accessibility practices and policies for traditional forms of technology, such as computers and websites. As technology evolves, we have also created resources that give employers practical guidance on disability, inclusion when they choose to adopt AI. These resources guide employers through each step of choosing inclusion I have technology, implementing it, and training staff on best practices. To create more useful and robust materials, we collaborate and partner with a range of policy makers, disability advocacy organizations, technology companies, employers, AI experts, and researchers. One of our first public resources on this topic was in 2021, AI and Disability Inclusion Toolkit found on our website, and it lays out for employers steps they can follow if they're interested in using AI in recruitment and hiring. This is one of the biggest use cases that we're
seeing in the employment context right now and one that is deemed high risk generally, and it's tricky, because according to data from the society for human resource management, I think 92% of employers are procuring these AI tools from vendors, so there may be a lack of understanding about how the tool functions, what its risks are, especially to disabled workers. So there are responsibilities that we're laying out for employers to ask the right questions, to put in place the right governance practices so that these tools are not discriminating, because as the previous panelist said, the presence of AI does not negate employer's responsibilities to nondiscrimination. We have also put out resources on the promise of AI. So there are some disability led startups that are using AI in exciting and innovative ways to
match people with disabilities to jobs, to train people with disabilities to fill in demand jobs, and we have this research on our website as well. In 2023, we also put out a series of articles on how automated surveillance can create barriers to workers with disabilities, so moving out of the hiring and recruitment context, looking at the issues with ability based monitoring of workers to energy productivity, advancement decisions, and even term I have nation. There are many accessibility challenges with these automated tools that employers should be mindful of. We are also working on new policy resources with federal partners overseeing AI risk management. These upcoming materials will guide employers that use AI tools in recruitment and hiring to follow concrete steps to maximize the benefits of this technology and better manage risks. And this will be a much more robust policy resource that we're aiming to launch this fall.
Ultimately, employers can use AI to reduce bias and even open the virtual door to recruit more workers with and without disabilities. Disability led startups are leveraging AI to advance inclusive hiring and major hiring platforms are advancing inclusive strategies in their own networks. To learn more about our resources, I encourage you to visit our website at www.PEAT works.org. In closing, I want to reiterate that we all have a
duty to decide when and how to use new technology. All uses of AI are not inevitable. Many are good. Some require more attention. I hope the work I shared today can help with how to address the risks and benefits of AI for disabled people. I want to thank the American Association of People With Disabilities and the Center for Democracy and Technology for their partnership with the Access Board on this issue. And I look forward to hearing from the rest of the panelists. >> Elver Ariz Silva: Thank you so very much, Nathan.
Now for the fourth presentation, we will welcome Sarah Decosse, Assistant Legal Counsel of the ADA and GINA division within the Office of Legal Counsel in the U.S. Equal Employment Opportunity Commission. Sarah, you may begin, please. >> Sarah Decosse: Thank you so much. Let me just share my screen. Thank you so much for this opportunity. My name is Sarah Decosse. I'm a disabilities attorney with the Office of Legal Counsel at the EEOC. I use she/her pronouns and I am a white woman wearing a blue scarf and a blue sweater. I'm going to address AI and other algorithmic decision making tools and employment for people with disabilities. And
I am delighted to follow on Nathan's presentation, which is very apt in terms of what I'm going to be discussing. I'm going to go into a little more detail about just how the ADA raises concerns about the use of AI in the workplace. So first, I just want to discuss a few of the types of tools that we're seeing. There are many and they are moving into the workplace rapidly. So a few of those are video interviewing, chat pots, resumé readers, productivity monitors, key stroke counters, and gamified tests. In addition, there is a very large group of tools that are called wearables. These are AI driven devices that literally attach to someone's body
and they can perform many different tasks. They're designed to perform many tasks. So some examples of wearables include eye tracking glasses, driver fatigue detectors, posture and limb strength trackers, workplace wellness monitors, movement disorder detectors, and things like social behavior monitoring. For example, monitoring body language and tone of voice. Or there are several devices that use headbands, headsets, or helmets to detect emotion, attention, or mental focus, and they also can perform things like EEGs. So where do we see potential concerns arising about the ADA? We see them arising in three different categories and because time is very short today, I'm just going to quickly run through those categories. The first, and Nathan brought this up, is the failure to provide reasonable accommodations. There are many different circumstances in which individuals with disabilities may need an accommodation to interact productively with an AI tool. So for example, if an individual with a visual impairment cannot
navigate a hiring process because the algorithmic tool is not fully screen readable, they may need an accommodation that will allow them to do so. Or if an individual with limited manual dexterity cannot complete a timed knowledge test that requires the use of a keyboard, even though they're well qualified to the position, they may need an accommodation to allow them to perform in a way that will be assessed accurately. So we have a number of recommendations promising practices for employers. As a note, just as I was starting, I put in the chat links to our two technical assistants materials that address AI and the Americans with disabilities act. One is more or less focused towards employers. The other provides several tips for employees and applicants. So I hope that you will look apt those. Those go into much more depth about what I'm discussing today. So what are some promising practices with respect
to ensuring that people who need reasonable accommodations when they're interacting with AI tools can get them? First, making sure that job applicants and employees know that reasonable accommodations will be made available. Making sure that HR staff know to recognize asked for reasonable accommodations. And ensuring that that alternative test formats, alternative formats for any different kind of processes are available. And of course, those same premises would apply with a third party is undertake being the hiring or employment task for an employer. The second category of potential ADA violations in employment relates to the ADA's limitations on employers seeking individual information about someone's disabilities or their medical status. Medical information. Congress recognized that allowing employers to get this information might subject individuals in the workplace to discrimination on the basis of disability, and that's why these particular three provisions I'm going to mention apply not only to people with disabilities, but to everyone in the workplace. So there could be potential concerns if AI tools,
for example, make disability related inquiries. In other words, they ask questions that are likely to elicit information about someone's disability. Or they're also prohibited, and again, this is in most circumstances. There are some exceptions. From collecting information that
qualifies as a medical examination under the ADA. And similarly, if employers click medical information, the ADA requires them to keep it confidential with very limited exceptions. So what are some promising practices for employers so that they can avoid either inadvertently collecting this type of information or failing to respect the ADA's requirements to keep medical information confidential? Some promising practices would be communicating with any Investigators who might be selling these tools to ensure that the tools do not ask questions that may elicit information about disability. To ensure that the tools do not engage in medical examinations unless something like a request for a reasonable accommodation has been made. And similarly, ensure that any medical information collected remains confidential and is only used for appropriate purposes, according to the ADA. The third category of potential violation is
what the ADA calls screen out of a qualified individual with a disability. The ADA prohibits the use of selection criteria that deny employment opportunities to qualified individuals with disabilities, whether the screen outs are intentional or not. For example, if an employer uses a chat bot that is trained to reject applicants with significant employment gaps, individuals who may have such gaps, because of their disability, may be screened out, even though they're qualified to do the job. Or an individual may have a disability that affects their speech, and a video interviewing tool may rate them poorly, because they do not speak as quickly as other candidates. Again, even though they're able to do the job.
So here, too, we have a number of promising practices for employers to reduce the risk of screening out qualified individuals with disabilities. Among those promising practices, it would be helpful to clearly state that reasonable accommodations are available. To give all applicants and employees as much information about the assessment tools as possible in advance of the individual beginning the assessment, we've noted that sometimes people don't know what they're about to be asked to do and may not recognize that they need an accommodation until they're already midway through assessment, which makes it a little bit more difficult.
Further, it's important to make sure that the assessment tools reflect actual essential job qualifications so that those tools are only measures qualifications that are truly necessary for the job. And of course, those qualifications should be measured directly, not simply through correlations that may prove to be inaccurate. And employers may wish to directly ask software vendors questions such as whether or not the user interface can effectively interact with people with as many disabilities as possible and if the materials are available in alternative formats should they be needed. It's important to note that there are occasions when both an employer and an AI vendor may both be responsible under the ADA. So when AI algorithmic
decision making tools are designed or implemented by a vendor and the vendor acts as the employer's agent, the vendor, in some circumstances, may have the same legal obligations as the employer to job applicants and employees under the ADA. Some quick final notes, I don't want to go too long, first just to note that the EEOC is very interested and has put a great deal of energy into advancing our work on clarifying not only the ADA implications of AI, but also looking at other equal opportunity laws and the impact that AI halls in other protected factors that are covered under those laws. These are just some of the references to some of our initiatives on that front. And finally, I just want to make reference to those two technical assistance materials that appear on this slide with hyperlinks that we can share with you, as well as a very recent settlement just two weeks ago that we made with respect to individuals who needed accommodations of effective screenreaders to apply for jobs. You
were not provided those accommodations. As well as a link to our library of disability related resources, which we welcome you to visit. So thank you again for this opportunity. I'm delighted to answer questions later. >> Elver Ariz Silva: Thank you so much, Sarah. Very appreciate it. Now our next presenter, Josh Mendelsohn, Acting Chief of the Disability Rights Office at the Federal Communications Commission. Josh, please proceed. >> Josh Mendelsohn: Hello, everyone.
I am Josh Mendelsohn and I am the Acting Chief of the Disability Rights Office of the Federal Communications Commission. I actually white bald middle aged man with a salt and chili colored beard. I'm wearing a black suit with a green shirt. I would like to thank the Access Board for hosting this hearing today to cover this very important issue. The FCC has had a few ongoing initiatives that have involved the use of artificial intelligence, which has an impact on individuals and people with disabilities, and I would like to talk about three specific areas today. The first being modern communications or access to communications. The second being video programming and the third being emergency
access or emergency access communications. First of all, communication access or modern communications cover a wide range of applications and uses ranging from telephones and telecommunications relay services. P this also includes hearing aids, as well as interoperable video conferencing platforms, like what we are using right now at this moment. And all of these temperatures been impacted by the recent FCC action involving artificial intelligence, otherwise known as AI. I am going to start by talking about robo calls and robo texts. This is becoming a huge issue with almost everybody who has a cell phone, whether it be a landline or a mobile phone. Many individuals are being deluged by these
robo calls and robo text messages. And the FCC is very well aware of this issue. We are also aware of the use of artificial intelligence on the side or, rather, by those bad actors who are using AI to send even more robo calls or make them more realistic, and also, the use of AI to inhibit or reduce and pre vicinity robo calls and robo texts. In using this technology or using artificial intelligence in robo calls, especially we are concerned most particularly as we have recently been collecting comments and soliciting those comments on the accessibility and taking that into consideration in defining those technologies, particularly on the development of how to avoid discouraging the development of beneficial AI tools to detect and block unwanted and fraudulent calls and text messages. So how can AI be used to improve the ability of people with disabilities to communicate with called parties? Such as the ability to revoke consent to future calls and messages. And to work effectively with telecommunications and relay services and to generate or translate those messages. We have also been seeking comments on steps that we can take in order to ensure that important accessibility tools such as voice altering individuals or, rather, voice altering technology for individuals with disabilities are not negatively impacted by our rules, such as the TCPA rules, regulating calls using artificial or prerecorded voices.
We recently, earlier this month, held a vote to adopt an NPRM or further rules and sought comment on the positive use of AI to improve access to the telephone networks for people with disabilities. Especially those individuals who are using artificial or AI generated voices or synthesized voices to make these calls. Another area that we've been looking into and that we have been regulating is telecommunications relay services, which enabled people with disabilities, people who are deaf, hard of hearing, or speech disabled, to be able to use a telephone network to make calls to people who are not deaf or hearing on the other end. One example of telecommunications relay services is internet protocol captioned telephone services, otherwise known as ICPTS in which a person who is deaf or hard of hearing will be able to speak for themselves and then the other person on the other end of the line will hear that speech, what the person is saying, and then when they wish to speak back, what they say would then be typed and sent as a text message back to the person who is deaf or hard of hearing. Six years ago, the FCC ruled that ICPTS could also be provided on a fully automated basis using only ASR or automatic speech recognition, which we, at times, have referred to as a type of AI. This has been used to generate captions without the participation of a
communications assistant or otherwise a relay agent, which could serve as an intermediary. Recently, just last month, we adopted a new plan for compensation regarding IPCTS or, rather yes. The ICPTS. IP captioned telephone service providers using different rates. One would be a higher compensation rate for that which used communication assisted technology or communication assisted captions, otherwise a person, a human who would be printing the captions. And a lower rate of compensation for ASR only captions. That way, we sought to reduce the incentive to provide only the lower costing ASR caption service when captions by a human intermediary would be better in some circumstances or preferred by the consumers. Recently, just several months ago, a consortium of consumer groups filed a petition requesting that the FCC initiate a rule making to require ICPTS providers using ASR technology to also provide consumers the option to switch to a communications assistant at any point during an IPCTS telephone call. Now, we currently have a number of IPCTS providers and all of those providers are able to
provide IPCTS using ASR, but only a few of those providers also provide the option to switch to a communications assistant. The petition asked or states, rather, that ASR being used as a part of IPCTS frequently misinterprets speech with accents, dialects, or patterns which deviate from standard American English or when used to recognize speech in environments with background noise. We are currently in the process of soliciting comments and replies and all of those comments and responses are due or, rather, comments, I should say, are due September 3rd and supplies are due September 16.
Another form of telecommunications relay services is known as IP relay in which a deaf or hard of hearing person or a person with speech disability would type a message and then that typed message would be spoken to a person on the other end, then that person on the other end of the line would be speaking and what that person speaks would then be transcribed into text, which then I caned read as a TRS user or IP relay user. We currently have two applications by current IPCTS to also provide IP using speech synthesizers and ASR. Another area that's been impacted by AI is interoperable video conferencing system or video conferencing services. Just like this platform, such as Zoom, teams, WebEx, and other types of platforms. We at the FCC are seeing an increase in the use of AI and ASR on these types of platforms. And the STC has already ruled that these platforms must be accessible to people with disabilities. So we see AI to be used to generate automated captions, also AI summaries of conversations, and also being used for transcription services. They also can be used to
automatically dead ignite signers as speakers. One emerging area that we are seeing the use of AI have an impact on is automated ASL interpreting services, otherwise known as AIS, 234 which AI is being used to recognize sign language and translate those signs into speech or text. And the same is true in the other direction. We are seeing speech and text being then translated to American Sign Language or other signs using an Avatar or photo realistic personas or figures.
We recognize that this holds a lot of promise for the telecommunications landscape and other contexts, and for other contexts as well. The use of AI is also emerging in video programming. The use of audio descriptions, for example, which interprets what's being shown on the screen for those who are blind or have low vision. Wore seeing AI be used to generate these types of services and audio descriptions. Even though we recognize that there are some complaints
regarding the quality of these audio descriptions and the quality of speech, ASR is also being used to generate closed captions for video programming more and more in recent years. However, we're also seeing complaints regarding the dictionaries that are being used by these services. Some are not up to date or do not have the accurate dialect or vocabulary that's necessary with those contexts.
Now, there was a lot of promise when it comes to new next generation televisions or ATS [inaudible] 3.0 in terms of how AI can be used to generate overlays of information when it comes to signed information or automated sign language or captions, various graphics or you know at this pre stations of graphical information on the screen. And I recognize that I'm soon out of time, so I would like to encourage people to visit our website, www.FCC.gov/accessibility. And there you'll be able to find more information about our recent initiatives involving AI and people with disabilities. The FCC is committed to the use of AI in order to
enhance accessibility for people with disabilities in modern communications and video programming, as well as for emergency communications. >> Elver Ariz Silva: Thank you so very much, Josh. Now in our next presentation, we will hear from Rylin Rodgers, Disability Advisor for Microsoft Accessibility on AI and Accessibility at Microsoft. Rylin, you may begin, please. >> Rylin Rodgers: I'm the disability policy directorate Microsoft. I'm a middle aged white woman with brown and white hair wearing a Variety of blue patterns today. I'm excited to join this conversation and build on some of the
great resources already shared. At Microsoft, we believe accessibility is a fundamental right. It breaks down barriers and opens doors to a more equitable future. We're committed to ensuring that people with disabilities have access to accessible technology. Our approach takes on four different areas. We're really interested in the challenges
people are facing and we understand that they are complex and there's no single company or sector that can solve them. So we're committed to working with our global ecosystem of people with disabilities, partners, and customers. We really work together in four key areas: Developing inclusive technology, working with disabled people as the talent that drives the [inaudible] forward. Modernizing Public Policy to ensure access to fundamental right of Accessibility and accelerating awareness and connectivity through partnerships. It's key that these pieces work together and are grounded in the needs of disabled people.
I want to take a step back to our conversation around AI, because we've spoken a lot about the term and what it means and the Access Board is responding to the Executive Order on AI, but holistically, it's important to understand the history and that there are many types of AI that are currently driving accessible features in products and systems. AI has existed since 1956. We've had multiple waves of advancement from machine learning to deep learning, and to our current age of generative AI. I think this is an important point of clarification when we're talking about risk, opportunity, and regulation to be as inclusive and clear as possible about what the risks are and what types of technology we're targeting.
At Microsoft, we have a long history of really thinking about accessible and responsible AI. We're committing to advancing AI through ethical principles, and I think it's important to call out that within those principles, fairness and inclusivity are key components. This has been an ongoing practice for multiple years, and our accessible AI principles are regularly updated and reviewed and used in all of our product development. I encourage everyone to view those principles, guidelines, and toolkits on our transparent website so you can learn more about our approach and consider it. The other part that's helpful if we've taken
some of our learning and have been sharing it in terms of a blueprint governing AI, sharing what's possible and how can we think collectively about ensuring privacy, security, and accessibility in this New World? Where he find ourself at this moment of generative AI, which really is proving to be a paradigm shifting innovation. There's an opportunity to create a more inclusive experience at tremendous scale. Individual assistants can be available to millions of people across the spectrum of disability in ways it hasn't been previously available. And there's the potential to transform accessibility as a practice and making design more inclusive.
And I think that's a really important place to start. Our partners at GitHub have really been leading the way in terms of using AI in coding modeling. We've seen two massive steps forward in these efforts. One ensures that new code is built accessible by design as AI can acquire it and check is it going forward. And the other, as it's transformed the access to this practice of coding and other parts of technology development, to a wider range of disabled technologists. It's been critical to drive forward the future of innovation.
I want to take a minute to talk about what I think is a bit of the elephant in the AI room, and that is the foundational models and the reality that they're built on the world as it exists. I frequently say that the world as it exists is racist, sexist, and ableist, and that really gives us an opportunity to take a moment and think about what that means in terms of the models and how we need to be addressing it now and in the future. We are very clear that data libraries need disability data to empower representation and to protect against ableism. We're also aware that historically, data libraries under represent people with disabilities, disability experience, and disability expertise. We've been tackling this in a Variety of ways. As previously mentioned, there's quite a lot of learning in the pace of translating language access in the space of AI, particularly in a Variety of sign language, and an effort to get to more equity any terms of language access for people. We have ongoing research activities at Microsoft and
a blog really outlining where we see it as possible and what needs to still be done to make sure that those practices are culturally inclusive and meet the needs of consumers. The other part of the learning is that it's not just about filling and creating new data. It's about making sure that we're testing and modeling and really correcting the current data to prevents harm. This includes a critical need to test for disability bias in all parts of the design process, the need to tune the foundational models to more accurately represent who we all are and Highway we all interact in the world, and again, that ongoing commitment to filling the data gaps. I'd like to point to one example of that commitment to fill the data gap. It's one of my favorites. It follows so well after our previous speaker, because it really is
around what is that challenge about getting to accuracy for non typical speech in AI models? But it's also one of my favorites, because it's about partnership. We are going to get to more inclusive datasets faster if we work together. So the speech accessibility project is a project across researchers and all of the major tech companies working with disabled people to gather high quality representative speech s
2024-08-26 00:53
