Artificial Intelligence Has One Chance To Get Accessibility Right

Digital accessibility has been failing People with Disabilities for thirty years, with WebAIM's 2024 research showing that nearly 96 percent of top websites contain barriers blocking access for those who rely on screen readers, keyboard navigation, and other access technology. The World Health Organization estimates the global disability community at 1.3 billion people, making continued digital exclusion one of the most widespread injustices in modern technology. Artificial intelligence (AI) now writes a growing share of the world's production code, and a handful of foundation model companies are making the design decisions that will shape billions of digital experiences. Unlike a single developer producing one inaccessible page, a model generating low contrast text, unlabeled form fields, or broken keyboard navigation will reproduce those failures across every product that draws on it, compounding harm at a scale the web era never reached.

To force accountability into AI development, the GAAD (Global Accessibility Awareness Day) Foundation built AIMAC, the AI Model Accessibility Checker, in partnership with ServiceNow, scoring foundation models on how consistently they generate accessible digital interfaces. Once those scores entered public view, results changed: a Google model that had ranked last on the benchmark advanced to ninth place in its next release, and the A11y (Accessibility) LLM (Large Language Model) Eval developed at Microsoft confirmed the same pattern under public scrutiny. Some models now achieve near perfect accessibility scores, and a 30-person startup released an open source model ranking fifth at no cost. Yet the most common failures have not changed: low contrast text shows up on roughly 85 percent of AI generated pages, matching rates already documented on human built websites. Rather than correcting these problems at the point of code generation, AI models are reproducing them at scale, when accessibility checking tools can detect and address common accessibility failures including missing form labels and unlabeled buttons before any product reaches People with Disabilities.

The most direct path to AI tools that genuinely work for People with Disabilities is placing People with Disabilities in decision making roles before any product ships. When blind developers are present from the start, flawed assumptions surface immediately: someone who uses a switch device refuses a drag and drop interface outright, and a developer who navigates by voice demands to know which button a tool is referring to when none carry labels. A tab interface that looks correct on screen completely breaks down when a Blind user tests it with a keyboard and a screen reader, and a sighted mouse user will never catch that failure. The Accessibility Agents project built by Taylor Arndt and Jeff Bishop, both blind, arose from this reality; AI coding tools kept producing interfaces they could not use, so they built their own. The AI industry must recruit People with Disabilities, pay them fairly, and give them authority over product decisions rather than treating their inclusion as an optional advisory role.

Beyond the harm of inaccessible AI generated code, a parallel transformation is taking shape: AI is becoming a new category of access technology that creates pathways into software development for People with Disabilities who previously had no route into building digital products. People with Disabilities are already using AI tools and coding assistants to build software faster than was previously possible, and when blind developers build those products they naturally produce software that works with screen readers; when developers with motor disabilities build interfaces those interfaces naturally work without fine motor control, because both groups design from their own lived experience. As voice based and natural language pathways for interacting with AI tools expand, software development becomes more accessible to People with Disabilities across the board. Removing barriers so that People with Disabilities can participate fully as both creators and users of digital technology is a shared responsibility that every Community Builder must take personal ownership of, enacted collectively and embedded into every stage of planning, development, and ongoing improvement, because the vibrant and inclusive community we are all working to build will only become real when People with Disabilities are present and empowered within it.

Read the Full Article: Artificial Intelligence Has One Chance To Get Accessibility Right.
By: Keely Cat-Wells