Accessibility and AI: What’s changing, what must stay

<p><b>Amidst astonishing progress, we mustn’t lose sight of foundational truths.</b></p>

<p>Digital accessibility has never had more momentum or more complexity. Artificial intelligence, particularly the rise of multimodal AI, is unlocking capabilities for human-centered design that were simply not possible a few years ago. And yet, the pace of that change brings real risks: that we move fast and leave people behind, that we automate our way past empathy, or that we inadvertently create new barriers for the very people we are trying to serve.</p> <h4>The multimodal moment</h4> <p>For most of its history, AI operated in silos. A speech model handled audio. A vision model processed images. A language model worked with text. Each was powerful in its own domain, but the experiences they enabled were often fragmented, narrow and difficult to deploy at scale.</p> <p>Multimodal AI changes that fundamentally. Today's large AI models can see, hear, read and speak, sometimes all at once. They can watch a video and generate a transcript. They can look at a complex infographic and describe it in plain language. They can read a dense technical document and rewrite it for a person with a cognitive disability, adjusting reading level, structure and length in real time.</p> <p>For the accessibility community, this is genuinely transformative. Many of the barriers that have persisted for decades—the unlabeled image, the video without captions, the interface that cannot be navigated without a mouse—are now addressable at a scale and speed that manual processes never allowed.</p> <p>But transformation is not the same as progress. And that distinction matters enormously here.</p> <h4>What is changing: The opportunities multimodal AI unlocks</h4> <p>Alt text has long been one of accessibility's most persistent failure points. Studies consistently show that many images on the web are either missing descriptions entirely or carry text so generic it is functionally useless. Multimodal AI can now generate alt text that is not only accurate but contextually intelligent. It understands that the same photograph means something different on a news article versus a product page versus a medical report, and it calibrates its description accordingly. This is a significant leap forward not just in coverage, but in quality.</p> <p>Real-time audio description has traditionally been expensive and largely manual. Narrating the visual elements of a live broadcast or interactive experience for persons who are blind or have low vision required skilled human describers working under considerable pressure. Multimodal AI is beginning to close that gap, generating spoken descriptions of visual content as it happens. For people who have long been excluded from live events, sports, theater and interactive media, this represents meaningful access rather than minimal compliance.</p> <p>Cognitive accessibility is perhaps the most under-served area in the accessibility spectrum. People with dyslexia, ADHD, autism, acquired brain injuries or intellectual disabilities interact with digital content in very different ways, and their needs are extraordinarily varied. Multimodal AI, particularly when combined with user profile data, can dynamically simplify language, restructure layouts, reduce visual clutter and adjust information density. This kind of real-time, personalized adaptation has never been achievable at scale before now.</p> <p>For people with limited dexterity or motor impairments, keyboard and touch interfaces present significant barriers. Multimodal AI is enabling richer, more natural voice interfaces and gesture-recognition systems that reduce dependence on traditional input methods. As these systems become more accurate and more widely deployed, they hold genuine potential to improve independence for a large and often overlooked community.</p> <h4>What must stay: The principles AI cannot replace</h4> <p>The accessibility field is at real risk of losing sight of some foundational truths in the excitement around AI capability. Each of the following deserves its own moment of reflection.</p> <h4>Empathy cannot be automated</h4> <p>AI can generate an alt text description. However, it cannot feel what it means to navigate a world not designed for you. It cannot carry the lived experience of a person who is blind encountering a website that truly works, or the frustration of a person who is deaf who has spent years watching videos with no captions. Empathy, the kind that comes from listening to and designing with people who are disabled, rather than simply designing for them, is irreplaceable. No model can substitute for it, and no organization should let AI adoption become a reason to reduce investment in human-centered research and co-design.</p> <h4>Standards are not optional</h4> <p>The <a rel="noopener noreferrer" target="_blank" href="https://www.w3.org/TR/WCAG21/">Web Content Accessibility Guidelines</a> (WCAG) exist because the accessibility community spent decades establishing what &quot;accessible&quot; actually means across a wide range of disabilities and contexts. They are imperfect and evolving, but they are a critical foundation. There is a risk that, as AI-generated content proliferates across web pages, images, videos and interfaces, those assets bypass the human review processes that ensure standards compliance. AI must be a tool that enforces and supports WCAG, not one that accelerates the production of non-compliant content at scale.</p> <h4>Existing assistive technology must be protected</h4> <p>Millions of people depend on assistive technologies, screen readers, switch access devices, braille displays and eye-tracking systems that they have spent years learning and configuring. New AI-powered interfaces, however impressive, must not be designed in ways that break compatibility with these tools. Innovation that excludes the people it claims to serve is not innovation; it’s displacement.</p> <h4>Transparency is a right, not a feature</h4> <p>One of the less-discussed risks of AI in accessibility is opacity. When an AI system generates alt text, simplifies content or decides that a particular user journey needs modification, on what basis is it making those decisions? Can a person who is disabled understand, question or override that decision? Explainability and user control are not optional extras in accessible design. They are fundamental to dignity and autonomy.</p> <h4>Human judgment still matters</h4> <p>Automated accessibility testing, even at its best, currently catches somewhere between <a rel="noopener noreferrer" target="_blank" href="https://testguild.com/accessibility-testing-tools-automation/">20% and 40% of accessibility issues</a>. The rest require human judgment: a tester who understands context, nuance and the actual experience of a person with a disability. AI will improve that percentage over time, but leaders must resist the temptation to reduce human testing investment as AI capabilities grow.</p> <h4>A framework for what comes next</h4> <p>So how do we embrace what multimodal AI makes possible while protecting what must not be lost?</p> <p>It comes down to three commitments.</p> <ol> <li><b>Lead with the user, not the technology. </b>Every AI capability we adopt should be evaluated not by what it can do in theory, but by whether it measurably improves the experience of a real person with a real disability. Pilots, feedback loops and ongoing research with disabled communities are non-negotiable.</li> <li><b>Treat AI as an accelerant rather than a replacement</b>. The right approach will make human accessibility experts faster, more consistent and more scalable. Automation should take on repetitive, low-judgment work, while human expertise is preserved for nuanced, high-stakes decisions.</li> <li><b>Accountability</b>. As AI takes on more of the accessibility workflow, organizations need clear ownership of outcomes. When an AI-generated caption is wrong, an auto-generated description is misleading or an adaptive interface fails, someone needs to be responsible. Accountability structures must keep pace with automation, not trail behind it.</li> </ol> <h4>A final thought</h4> <p>As we move through 2026, it is striking how much has changed in a short time, and also how much has not. The fundamental challenge of accessibility has always been one of will, not capability. The tools available to us today are extraordinary. But tools don’t build a culture for all. People do.</p> <p>Multimodal AI is the most significant development in accessible technology in a generation. Used well, it can close gaps that have persisted for decades and bring genuinely accessible experiences to scale. But that requires clarity about what we are automating, what we are amplifying and what we might inadvertently be leaving behind.</p> <p>The question we must keep asking is not simply whether our AI can do something, but whether a person who is disabled would actually choose the experience it creates.</p> <p>If the answer is yes, we are on the right track. If we do not know, we need to ask them.</p>