Artificial intelligence has one chance to get accessibility right. If we fail, we risk repeating the web's biggest mistake—building billions of digital products that exclude over a billion people with disabilities. The 2024 WebAIM Million report found accessibility failures on 95.9% of the top one million home pages. Meanwhile, 1.3 billion people worldwide live with a significant disability. AI is now writing a huge share of the world's code. The question is simple: will we learn from the past, or repeat it at a massive scale?
Why AI Accessibility Matters More Than Ever
When one developer writes inaccessible code, only one product suffers. When an AI model writes inaccessible code, every product built with that model suffers. This is the compounding problem at the heart of AI-generated software.
Joe Devon, Co-Founder of the GAAD Foundation, explains: "When an AI model writes inaccessible code, it affects every product built with that model. A model that generates low-contrast text doesn't create one hard-to-read page. It creates millions of them."
Lori Samuels, Senior Director of Accessibility at Versant Media, adds: "We're at a critical inflection point. Right now, a small number of AI models are used to generate software applications. It's critical to focus on getting these models to deliver a consistent baseline default of accessibility in everything they create."
The Role of Accessibility Professionals Has Changed
Software engineering looks nothing like it did two years ago. Most teams now use AI coding assistants. Accessibility professionals must learn how to steer these tools.
"Just like accessibility experts became the go-to for semantic HTML, tomorrow's experts will need to know how to guide AI coding tools," Devon says. "Once they do, and once high-quality open-source projects are available, enterprise accessibility will improve."
The bigger challenge? Millions of small businesses. They'll use AI to build websites without knowing the first thing about accessibility. That's why the foundation model companies—like OpenAI and Anthropic—must get it right by default.
What Gets Measured Gets Improved
Foundation model companies compete on benchmarks. Until recently, accessibility wasn't on that list. That's why the GAAD Foundation and ServiceNow built AIMAC—the AI Model Accessibility Checker. It evaluates how well AI models generate accessible code.
"You can't improve what you don't measure," Devon says. "Google's Gemini 3 Pro Preview once finished dead last. Its replacement, Gemini 3.1 Pro Preview, jumped to number 9. That's exactly the kind of progress we hoped for."
The results are mixed but promising:
- OpenAI has three models with near-perfect scores.
- Arcee AI, a 30-person startup, ranks number 5 for free.
- But low-contrast text appears on 84.8% of AI-generated pages—almost identical to human-built sites.
There's also a worrying trend: regression. Models sometimes get worse on accessibility after updates. "That tells you accessibility wasn't part of their evaluation pipeline," Devon warns.
Regulations Are Closing In
Accessibility is no longer optional. Europe now requires digital accessibility for all businesses selling in EU territories. In the US, all federal agencies must comply, and state agencies will follow soon.
"If organizations want the speed of AI, they need to pressure AI providers to generate accessible code," Samuels says. "The same pressure should apply to privacy and security."
What 'Wrong' Actually Looks Like
Ask a leading AI model to build a tab interface. Visually, it looks fine. But test it with a keyboard and screen reader—it falls apart. A sighted mouse user would never notice. A blind user literally cannot use it.
Devon uses a stack of safeguards: open-source Accessibility Agents, component-level checks, and manual review. "AI doesn't get you 100% of the way. It gets you further, faster, but human expertise is required."
Samuels adds: "Accessibility is about creating usable experiences for people with a wide range of disabilities. AI can and must do better. But we also need to validate with real people who have access needs."
Include Disabled Developers From Day One
The fastest way to fix assumptions is to put disabled people in decision-making roles. The Accessibility Agents project was built by Taylor Arndt and Jeff Bishop, who are both blind. They built it because AI coding tools kept generating code that excluded them.
"Rather than building from a checklist, they scratch their own itch," Devon says. "When disabled developers are in the room, assumptions get exposed immediately."
The AI industry needs to actively recruit disabled developers, pay them fairly, and give them power. Samuels sees progress: "Major tech companies like GitHub, Apple, Microsoft, and Google are making developer tools more accessible."
AI as a New Category of Assistive Technology
AI isn't just a risk—it's also an opportunity. People with disabilities are using tools like OpenClaw and AI coding assistants to build software faster than ever before. Voice, natural language, and assistive tech are opening doors.
"When more diverse people build software, the software reflects more diverse needs," Devon says. "A blind developer builds tools that work with screen readers because that's obvious to them."
Samuels calls AI a new category of assistive technology: "It opens up possibilities for better user experiences, productivity, and creative power for people with disabilities."
What 'Accessible by Default' Should Mean by 2031
Devon's vision for five years from now is clear. First, when AI generates a traditional interface, it should be accessible without anyone asking. Keyboard operable, screen reader compatible, proper color contrast, correct structure.
Second, and more importantly, we need to decouple access from any specific interface. "If a restaurant puts its menu in an inaccessible PDF, a blind person is stuck. If an airline has an inaccessible booking flow, a person with a motor disability has to call a phone line. The real promise of AI is that services become available to agents. People with disabilities can use whatever tool works best for them."
That requires industry-wide accountability. We need benchmarks like AIMAC adopted everywhere. We need disabled people in positions of influence. And we need to act while the concrete is still wet.
The web era taught us what happens when accessibility is an afterthought: billions of pages, decades of legal action, and a generation locked out. The AI era offers a rare second chance. Whether we take it depends on choices being made right now, by a small number of companies, with a short window before the defaults harden into infrastructure.
The concrete is still wet. But not for long.
Comment