Accessibility meets AI: navigating an evolving landscape
The intersection of accessibility (A11Y) and artificial intelligence (AI) is one interesting topic to work on currently. As both fields rapidly evolve, we find ourselves at a fascinating — and challenging — crossroads. Let’s explore where we are, what’s working, what isn’t, and where we might be headed.
Two Moving Targets
Here’s the reality: we’re trying to solve a dynamic problem with dynamic tools.
Accessibility standards are constantly evolving. What worked for users five years ago may not meet today’s needs. Guidelines adjust as we better understand diverse abilities, as assistive technologies advance, and as new interaction patterns emerge. The WCAG (Web Content Accessibility Guidelines) themselves are living documents, updated to reflect our growing understanding of accessibility.
AI is still being refined. Despite the hype, AI tools are far from perfect. They’re learning, adapting, and yes — sometimes failing. We’re in an era of experimentation, where the technology shows immense promise but hasn’t yet reached full maturity.
When you put these two moving targets together, things get complicated.
The Automation Gap
Here’s a concrete example of this challenge: WCAG guidelines are being updated, but automated accessibility testing tools haven’t caught up yet. Tools on the market currently don’t mention upcoming WCAG 3.0 changes in their documentation or marketing materials, which may create a significant coverage gap for some time.
This creates a window of uncertainty. How long will it take for AI-powered tools to incorporate new standards? During that gap, organisations relying solely on automated testing might miss critical accessibility issues covered by updated guidelines.
It’s a reminder that in accessibility, as in many fields, there’s no “set it and forget it” solution.
Why AI Can’t Replace Human Judgment
The promise of AI-powered accessibility tools is compelling: automatically scan your site, generate alt text for images, identify and fix issues at scale. But here’s what we’re learning in practice:
Context matters immensely. An AI might accurately describe an image as “a vase of flowers on a table,” but it won’t know that the image is purely decorative on an e-commerce site selling perfume. It might generate verbose descriptions where silence would be more appropriate, or miss the cultural significance of certain visual elements entirely.
Testing requires understanding intent. AI may generate accurate descriptions, but might not understand when an image is purely decorative, which could lead to confusion for users. Accessibility isn’t just about technical compliance — it’s about user experience. An AI can flag a missing label, but it takes human judgment to determine if the fix truly improves the experience for people with disabilities.
Real users aren’t average cases. GenAI’s “intelligence” is based on averages and datasets, which can present only the most common scenarios, leading to unintended bias. The disabled community is incredibly diverse. What works for one person with low vision may not work for another. AI trained on “typical” scenarios may overlook edge cases — which, in accessibility, are often the most critical cases to address.
Get Oliver Pereira’s stories in your inbox
Join Medium for free to get updates from this writer.
This doesn’t mean AI tools are useless. Far from it. They can speed up initial scans, catch obvious errors, and help teams prioritise. But combining AI tools with human expertise ensures a comprehensive and accurate approach to complying with accessibility laws and standards and creating an inclusive experience for all users.
Reality Check: Testing an AI-Powered Accessibility Tool
To better understand these limitations, I recently tested an AI-powered accessibility tool — a Chrome extension designed to make websites more accessible on the fly. The results were… illuminating.
What I Tested
I tried the tool across several different sites: Pinterest, Trello, Amazon, and the WCAG documentation itself. The extension offered features like dark mode, text resizing, image removal, dyslexia-friendly fonts, and a “focus mode.” On paper, these sound helpful. In practice? Here’s what happened:
On Pinterest, dark mode rendered images nearly invisible, defeating the entire purpose of the feature. Text resizing worked inconsistently, affecting some text but not others. The “remove images” function only applied to content currently visible on screen, meaning I had to reapply it for each new section as I scrolled. Most confusing was the “focus mode,” which blurred images I hovered over — I genuinely couldn’t figure out what problem this was meant to solve or how it would help anyone. Trello showed similar issues, with dark mode making images difficult to see and the tool clearly struggling to adapt to Trello’s card-based interface.
Amazon presented a mixed bag. Dark mode actually preserved image visibility (finally, a win!), but the overall implementation was poor with noticeable contrast issues. Text resizing became a nightmare, causing text to overlap with surrounding elements and making pages essentially unreadable. Perhaps most frustratingly, all settings reset every time I navigated between pages, forcing me to reconfigure everything repeatedly. The “remove images” feature performed better here, requiring fewer reapplications while scrolling. On the WCAG documentation site itself — the very standards these tools should support — the dyslexia-friendly font was barely perceptible (I honestly couldn’t see a difference), text resizing caused sidebar text to overlap so badly that users had to collapse the sidebar just to read the main content, and focus mode remained as mysterious as ever, randomly blurring some hovered text while leaving other hovered text untouched.
What This Tells Us
This wasn’t a scientific study, but it revealed some critical patterns:
- Inconsistency across sites: Features that worked on one site broke on another, suggesting the AI couldn’t adapt to different design patterns and frameworks.
- Lack of persistence: Settings resetting between pages shows the tool wasn’t considering user journey or state management.
- Unintended consequences: Text resizing and dark mode — features meant to improve accessibility — actually created new barriers by causing overlaps and visibility issues.
- Unclear functionality: If I, someone who works in this field, couldn’t understand how “focus mode” was supposed to help, how would an actual user with cognitive disabilities fare?
- One-size-fits-none: The tool applied the same approach to every website, ignoring that different sites have different structures, different user needs, and different contexts.
Here’s the thing: I don’t share this to shame the tool’s creators. Building accessible solutions is genuinely hard, and they’re trying to solve a real problem. But this experience perfectly illustrates why AI-powered “overlays” and automated fixes can’t replace thoughtful, human-centered accessibility work built into the product from the ground up.
Quick fixes applied on top of inaccessible foundations often create as many problems as they solve.
Final Thoughts
We’re in uncharted territory where two rapidly evolving fields intersect. That’s uncomfortable, but it’s also exciting. The questions we’re grappling with now — about context, judgment, representation, and human-AI collaboration — will shape the future of accessible digital experiences.
The technology will improve. The standards will continue to evolve. Our understanding will deepen. But one thing should remain constant: our commitment to building products that work for everyone, regardless of ability.
The intersection of A11Y and AI isn’t a destination — it’s an ongoing journey. And perhaps that’s exactly how it should be.
What are your thoughts on AI and accessibility? Have you encountered gaps between automated tools and real-world accessibility needs? I’d love to hear your experiences and ideas.
