AI coding could widen dev access for people with disabilities
GitHub’s head of accessibility Ed Summers made a pointed observation at GitHub Universe 2026: AI-assisted development may be the first major technology shift that served people with disabilities at the same time as everyone else, not a decade later.
His reference point was the move to graphical user interfaces, which started at Xerox PARC in the late 1970s and went mainstream through the 1980s. Many developers with visual impairments couldn’t use those interfaces until the 1990s. Natural language and multimodal interaction remove several of those same friction points from the start.
The interface question isn’t settled
Summers is careful not to treat AI tooling as a universal fix. He noted that traditional programming rewards sustained deep focus on a tightly scoped problem. Directing multiple agents, tracking their progress, and fielding their questions is a different cognitive pattern. For some developers, that’s a better fit. For others, it adds friction.
Interface choice compounds this. Summers favours command-line tools personally, while others prefer editors or browser-based environments. Research GitHub conducted with accessibility research firm Fable covered terminal and command-line experiences specifically, running seven one-hour interviews with users of screen readers, magnification, and speech-recognition software. The findings confirmed that a text interface is not automatically an accessible one, and that accessibility conventions are inconsistently applied across terminals and CLI tools.
The output problem: AI generates inaccessible code by default
Summers pointed to the A11y LLM Eval project by Microsoft’s Michael Fairchild, which tested generated interfaces against automated accessibility checks. The study covered eight models, 32 prompt cases, and 1,280 control samples.
The baseline pass rate, with no accessibility-specific instructions, was 12%. The best individual model at baseline reached 25%. Adding a minimal instruction set directing the model to consider accessibility lifted the aggregate pass rate to 37%, roughly 25 percentage points higher. A more detailed basic instruction set pushed that to 60%.
The takeaway is direct: teams cannot assume an AI agent will infer accessibility requirements. Those expectations have to be written into the development instructions explicitly. GitHub publishes guidance on custom agents and accessibility instructions at accessibility.github.com as a starting point.
Why this matters beyond disability access
Summers noted that accessibility improvements tend to become general-purpose features over time. Voice-to-text and text-to-speech are now standard interfaces most people use daily. Both originated in screen reader technology built for users with visual impairments. The same pattern applies here: building for accessibility from the first instruction tends to produce better software for everyone.
GitHub maintains a dedicated accessibility support and feedback channel, with reports triaged alongside product teams. Research from the CLI work has fed directly into Copilot CLI development.
I’m founder of BizStack at Cagri Sarigoz LLC and a passionate advocate for entrepreneurs.
With over 14 years in tech, marketing, and AI, including my role as Head of SEO at CitizenShipper and co-founder of TaleBot at Intale AI, I’m dedicated to sharing genuine, useful product insights and tips.
At BizStack, I aim to cut through the digital noise to provide clear, actionable advice.
And more than all else, I’m a father to a (always) little girl and a husband.
Contact me at [email protected] for assistance.
