AI and Accessibility - Part 2

Hello all, welcome back to a course on digital accessibility and in the previous session we spoke about AI and accessibility and what role AI is playing in um advancing accessibility and helping with uh creation of multimodal interactive systems which can be customized uh as per individual users needs and u how access AI is improving accessibility not just by providing tools to enhance uh experiences for digital uh for people with disabilities, but also we spoke about AI powered tools of evaluation and how AI can help with um evaluation of your digital products in order to um know about their compliance standards uh whether they're able to meet uh uh web uh compliance accessibility guidelines. guidelines etc. uh and kind of saving on human effort in order to uh reach to a better compliance. Uh so we kind of now understand that AI has the potential to transform accessibility by offering tools that can support various challenges faced by uh people with cognitive disabilities, neurode divergence and other uh as disabilities which are more commonly known like sensory disabilities. Uh we went through several examples of real time language translation, real time close captioning, um how you know now Windows has a dictate uh plug-in and read aloud plug-in and uh all of those things. So all of these technologies are based on AI based models and they are trying to make um the tools more inclusive, more accessible for persons with different kinds of limitations and um uh however uh we have to also understand that the future of AI uh in accessibility is definitely promising but there are certain risks and uh aspects which we also need to understand when working with AI and accessibility. So we have all learned about accessibility compliance uh in this course in some of the previous week sessions and when we talk about digital accessibility compliance um uh of course we have also tried to understand how AI can help and support evaluation of digital tools so that it can help in uh securing compliance right so uh there can be um comprehensive risk assessment. There can be regulatory reporting. Not just the development uh organizations but also auditors can use uh AI based tools in order to quickly uh audit the uh the the digital products. Uh then vendor compliance regulatory information access. So just um having so I'm sure you might have seen the WAG document it while it is very comprehensive very well annotated but it's still a long document right so uh where AI based tools can help you in accessing the document as per your need and uh the kind of criteria which is required it can directly uh tell the designer or the developer that probably your design is not fulfilling so and so criteria number so make sure that it is um uh it is compliant which the techniques to do it are so and so so I'm sure these kind of tools can be built in order to ensure compliance and due diligence however we have to look at this compliance aspect from the other uh side of it as well so the moment we integrate AI based um agents in our digital experiences, they automatically fall under the purview of digital accessibility compliance rules as well. Right? So any website so for instance any website which has a chatbot integrated uh in their um interface has to comply with WAG uh compliance guidelines and the chatbot also falls under the purview of compliance and regulation. However, there are certain aspects to uh AI based tool boxes which are difficult to manage in terms of compliance. The compliance guidelines themselves um have not adapted yet. So let's talk about all of those aspects. So because AI is changing the rules of compliance, the content which is AI based is more dynamic. It is more personalized. It's not a standard text. It is generated in real time and sometimes it is unpredictable because it's real time generation. So ensuring compliance becomes even more challenging and critical. Automated decision making systems shape who gets access to information jobs or opportunities. This is a very big issue when we are talking about inclusion and AI uh based decision decision making systems shortlisting systems all of those are um uh basically taking away um from the human aspect of access and inclusion and this is something which we need to talk about. So accessibility compliance as we know it today must adapt quickly or it risks becoming irrelevant in the near future because more and more digital experiences will be tailored around AI rather than AI being a tool or a chatbot you know used on a website. It will be the content itself may be generated in real time. The moment you open a URL, it is not a static text. It is a generated website which uh is tailored to your taste, your um uh previous uh browsing history and all of those things. Then compliance regarding layout, regarding perceivability, operability, all of those things comes into question. And that is why current WAK guidelines do not talk about uh AI based or dynamic content a lot and um uh it needs to adapt very quickly uh otherwise it can get irrelevant very soon in the near future. So let us try try to discuss how will accessibility evolve in an AIdriven world. So for decades now we have been uh seeing that accessibility compliance have been governed and regulated by standards such as the WAG guidelines. Uh and then there are laws to enforce the guidelines. So such as the uh ADA which is the American disability act uh the section 508 and then India's right of p rights of persons with disability or the RPWD act which we spoke about. Uh so these are enforce these are laws that enforce the guidelines on to uh developing agencies and uh uh innovators engineers and designers. So the frameworks are grounded in a vision of the web as structured, predictable and largely static which has drastically changed with the evolution of AI based systems and dynamic pro um unpredictable personalized realtime generated content is changing the definition of internet as we know it. So these guidelines and these acts are largely based on the pretext assumption that the uh website or the app or all of those things are predictable. They are structured. The owner or the website um um uh you know the creator has access to change it right or correct it as per the guidelines and comply with the guidelines. But the moment it is generated, you're not really create, you're not really owning uh the content in real time. You don't know if I search I when I uh open the website what I will see uh if somebody else opens the website what they will see. So we we largely don't know every time. So then that becomes a very problematic situation. So that brings us to the point that accessibility compliance itself is at a turning point and that it needs to adapt quickly and accessibility compliance however has always evolved with technology shifts. So now also it can do that and it can evolve with technology shifts. So one of the major technology shifts in terms of digital accessibility have has been from web which was uh landscape based layout which was opened on a laptop or a PC to a mobile based interface or apps and uh the guidelines have evolved with the change in technology and there can be other uh multimodel uh interfaces like VR, AR and all of those have also now come under the purview of WAG guidelines. So that it's not really that the guidelines have not adapted before, right? So maybe it is possible to do it for AI based um interactions as well. So from static web pages to mobile apps uh as we said to dynamic single page applications now uh the blue guidelines has to include all of them but the rise of AI marks a deeper infliction point where both the pace of the change and the unpredictability of outputs challenge exist existing frameworks as you will see And there are certain issues so which we will talk about. The first being that AI generated content is dynamic by nature. It is generated in real time. The semantic essence may also differ uh every time it is refreshed. So traditional accessibility testing assumes that content is relatively stable. Once a page is designed and coded, it can be audited against W CAG. However, uh use of AI uh and AI generated interfaces like chat, GPT powered assistance or adaptive e-learning platforms produce dynamic outputs that vary by user context or query. Right? So it that raises new compliance related challenges. Say for example, how do you certify accessibility when the output changes every time? Who bears the responsibility? The business is deploying AI or the vendor supplying the model. So basically if uh I am using a chat GPT, what does this mean? that if I am embedding a chat GPT based agent on my website who is responsible for an incompliance or non-compliance is it me who is the owner of the website who's created the website and embedded the charge GPT or CH GPT itself like um um uh or the owners of CH GPT who who are deploying the AI model. So that that's a dilemma as well. Uh can the existing standards like WCAG 2.2 adequately capture these fluid interactions? So right now they are not comprehensive. They are not really talking about a lot of dynamic content. Uh but probably in WAC 3.0 to there may be uh some more sections on AI based interactions. Then another um issue is that the standards always lag behind innovation. So this is an unfortunate truth of all industries uh not just accessibility. uh the standards follow while the innovation is leading the change. Uh so WAG is essential but it was not written with generative models or adaptive interfaces or realtime algorithmic decision making in mind. So it is not really talking about them uh right now. They um they probably form a baseline currently but uh they are not really they're lagging far behind the innovation and the current state-of-the-art guidelines cover text alternatives, contrast ratio, keyboard navigation. These are like basic operability issues. um but they yet don't address some of these points which is like caption accuracy thresholds. So whether the caption which is autogenerated is it semantically accurate right biases in voice recognition for atypical speeches with say accents or lisps or stutters all of those things. um transparency of automated decision uh systems they don't uh address uh these aspects. This gaps means compliance frameworks may soon fail to cover the most in impactful accessibility risks which are most of these right we are talking about using AI for making lives easier but who and what regulates it another issue which can come is the rise of algorithmic discrimination. So accessibility isn't only about the POR, the perceivability and operability of the content, understandability etc. But it is also about uh inclusion by definition like including all kinds of people. This um uh disability is one of the facets of inclusion. We have spoken about uh age based uh exclusion, gender based exclusion and u other non-binary u uh gender based exclusions etc. So but because algorithms and large language models work on pattern recognition models right. So basically they are uh working on data which is the commonly occurring opinion and then that human bias can very much reflect in LLMs which is also happening and this is one of the major concerns in the AI community uh which is called algorithmic discrimination. So hiring systems can penalize for example candidates with disability due to speech or gaps in employment. Health platforms that misclassify disabled patients. Chatbot that provide incomplete or misleading information to screen readers. All of these are possible. Compliance will need to expand from technical code checks to ethical auditing of algorithmic systems. Right? So now let us talk about emerging compliance models for an AIdriven world. Right? So as AI becomes central to digital experiences, accessibility compliance cannot remain tied to static checklists or one-time audits. Several models are already taking shape pointing to where compliance is heading next. So let us talk about those aspects. So one is continuous and realtime accessibility monitoring. So static audits are no longer sufficient. So as AI systems generate new outputs every second, it's dynamic. Compliance also needs to be dynamic in nature. Compliance will have to shift from a one-time audit to a realtime auditing mechanisms and which is continuous monitoring frameworks. Um so for example realtime caption validation measuring word error rate uh WER in automated captions dynamically and flagging when thresholds are breached voice command testing pipeline. So stress testing AI agents with diverse speech samples including uh disarthic and accented voices. Uh adapted UI monitoring tracking how interfaces uh reflow and adapt when personalized by AI. So businesses may need to implement automated monitoring agents that tests AI outputs continuously much like cyber security intrusion detection systems. The other emerging uh aspect is AI specific accessibility standards. The accessibility community is already debating what comes after WCAG 2.2 which is the current uh guide book. The W3C accessibility guidelines the WAG 3.0 0 draft acknowledges the need for outcomebased flexible testing, but we may see entirely new standards that cover accuracy benchmarks for uh AI captions and transcripts, data set documentation, disclosure of what user groups were included, excluded, bias auditing protocols to measure fairness for disabled cohorts. Um so bias auditing pro protocols means that uh one is that the uh that the owners have to audit themselves and document what kind of users are uh getting included and excluded and then probably doing accessibility uh auditing is one thing but also adding bias auditing. bias auditing protocols uh is based on the algorithmic discrimination that we spoke about earlier and that whether uh those AI systems AI algorithms are objectively doing decision making or not that is something which needs to be taken care of. Explainability requirements for automated decision systems is also required. In short, the compliance will expand from content to the algorithm itself. So it will get intertwined with not just the front end, not just the experience part of it but more deeper uh computer science aspects of development and um design itself. Then uh we spoke about one of the issues that who shares the responsibility of a compliance failure. Right? So then that the acts the rule the laws um which are formed to regulate and enforce the guidelines have to um keep the responsibility shared between vendors and deployers. So in an AIdriven ecosystem accessibility responsibilities have to be distributed. uh vendors that are AI model developers must disclose uh training data limitations and accessibility performances. Businesses deploying AI must validate those outputs in real context and provide some fallback mechanisms. Uh, of course they may not be able to uh judge all of the use cases, but uh the testing um uh you know methods can include a diverse set of uh audiences who may ask a query to your AI system. Regulators will need to clarify liability and who is accountable when an AI powered exclusion happens. So this part also becomes a little bit tricky with open-source models. Uh but we'll see how it evolves. I don't really have a right answer to it right now and uh maybe that is something open to discussion uh in the um discussion session. So this shift echoes GDPR's uh shared accountability and privacy law. Uh so similarly we have shared accountability in security and privacy. The same needs to be extended to accessibility as well. So with compliance of course will increase the number of accessibility audits and the traditional audit checks code and documents in the future. Audits may include data set inclusivity reviews to check if disabled users were represented at all. So user demographic may need to be disclosed by companies. Bias testing to check how AI outputs differ for disabled versus non-disabled users. Transparency assessments that check if model limitations are disclosed. Fallback evaluation to ensure that the system provides accessible alternatives when AI fails. Accessibility orders will evolve into multidisciplinary assessments, blending technical testing, data ethics and user research. Risks of ignoring accessibility in AI. Uh we cannot risk ignoring accessibility in a in the use of AI. The first and the foremost risk being the legal exposure uh as AI becomes more central to most businesses operations excluding disabled users could trigger lawsuits under existing disability laws. So one uh organizations need to be very careful about um you know including AI models particularly in decision making uh aspects of your organization. In the US there is already um uh department of justice's rule for public websites explicitly covering dynamic and AIdriven content. You can uh visit this link. In Europe, the EAA will demand accessible design for AI powered services from June of this year. In India, is 78 uh SE 17802 which is developed by the BIS already set sets national accessibility standards that ex extends to all digital platforms. And this standard already speaks about dynamic content. So it is already under the legal exposure. However, the guidelines are not in place as much. So then as designers and as developers we have to tread carefully because the laws cover it but the guidelines currently do not. So we have to be really careful about it. Then of course there can be other risks such as reputational risks. The viral story of AI mocking or failing a disabled users can spark backlash. Accessibility failures in AI aren't invisible. They make headlines. And then uh because AI is the most talked about aspect of technology in today's day and age uh it can risk reputation of your organization. market laws definitely we've been talking about that the fact that disability community represents over 1.3 billion people worldwide and u they have a huge spending power failing to make AI inclusive it's not just a legal risk but a missed market opportunity and we have to address this gap uh with conscious efforts and we've We've also spoken about uh several times the losing temporary and situational customers uh and users over incorrect usage of AI in your uh interactive systems. Internal workforce exclusion. So AI bias in HR tools or ERP systems can exclude talented disabled employees. While we were on the one side in the previous session when we were talking about the potential of AI, we were talking about how AI technologies can now enable people with disability to attend meetings with equity to have a seat at the table and companies can confidently hire them because tools are in place uh to facilitate easy communication between person who is has a sensory hindrance with a person who may not have a sensory hindrance. However, if and when AI is integrated in hiring mechanisms, shortlisting algorithms and all of those, it may end up um you know kind of excluding uh those people with the bias and accessibility compliant in AI isn't just defensive. It's a way to build better products, broader markets, and stronger trust. So these are some points which can be considered if you are a new business owner which is working in the domain of AI uh based tech or innovators which are looking to work in the domain of accessible AI or AI for accessibility. Uh of course keep in mind a few of these things. Step one is audit the current AI systems. Make sure that all your um AI powered tools which are in use are uh audited. Access accessibility assess your accessibility gaps both technical and ethical. Step two is build accessibility into procurement. So make sure that the vendors, theus that you sign, the agreements that you sign cover clauses which mention accessibility in order to ensure shared accountability that we spoke about earlier. It has to be tailored into the agreement with legal statements in place. Step three is implementing continuous monitoring. Deploying automated testing tools and monitor you know captions and everything needs to be established. So as we spoke earlier uh audit auditing accessibility audit will surely shift from a onetime audit to a continuous auditing system a continuous monitoring system. So make sure that you uh are also deploying such systems in your product so that even if there is a third-party monitoring system installed, your product is able to uh pass accessibility compliance. Train your teams on AI accessibility risks. Tell your development teams, tell your design teams about these risks. Communicate with them, discuss with them about possible ways to uh curb those risks. Uh uh what are the most optimized solutions to um move forward uh discuss and move forward with your teams. Engage disabled users in testing. Involving people with diverse disabilities in design as well as evaluation will surely help you in your journey and pay participants fairly uh and integrate their feedback in your iterative design. Prepare, brace yourself for regulatory shifts. They may come sooner than you know it. Track developments in WAG 3.0 0 and other national AI acts if they might uh you know add some more um addendums because uh because it will happen soon that the accessibility compliance may include algorithmic auditing as well. So what is the role of policy and advocacy in this scenario? Governments and advocacy groups are pushing for stronger oversight of AI systems. Likely developments may include AI accessibility certifications similar to VPADS today. But so VPAD is again your privacy related certification but extended to AI bias and data set inclusivity. uh global harmonization. So alignment between W CAG uh EAA and other AI based regulations uh have to uh you know kind of all of these organizations have to come together and form um interdisciplinary guidelines. Actually public private partnerships uh can help in maintaining and ensuring accessibility across platforms. Investments in inclusive data sets representing disabled users as well. Forward-looking businesses should engage in policy discussions not wait for enforcement as the future of accessibility compliance will be shaped by AI. This is really very important to consider and other aspects include that traditional frameworks at W CAG remain foundational but they're not enough to cover the risks of algorithmic exclusion. Uh compliance in the AI world will be driven by continuous um evaluations. uh have to be they have to be algorithm aare not just content focused shared responsibility spread across vendors deployers and regulators. So these are the three primary aspects of up andcoming WAG guidelines. Businesses that would adapt early will not only avoid legal risk but also unlock innovation and reach new markets, be the first movers uh to ensure inclusivity and maybe uh you know avoid the risk of shutting down altogether. To summarize today's session, uh the future of AI is definitely promising particularly in accessibility with the possibility of integrating different assisted technologies to create seamless more effective solutions for all kind of users. uh in an AIdriven world. However, accessibility compliance is not optional and it is the foundation of trust, equity and sustainable growth. Uh thank you for joining us in this session. I hope it had a clear takeaway this session and I will see you in the next session. Thank you for joining.