Accessibility can’t be a box we trust AI to check

As AI enhances what we’re capable of building, we can’t lose sight of who we’re building for. When AI agents claim to have validated accessibility without actually doing so, we risk excluding the very people we’re designing for.

I recently read an article about the gap between AI agents’ accessibility claims and what they’ve actually verified. The numbers are pretty telling: 88% of engineering leaders express confidence in the accessibility of AI-generated code, yet 64% identify accessibility issues as a leading source of rework after deployment.

I saw an example of this firsthand recently when I was asked to evaluate an AI-generated prototype developed by my business partners. A CTA was embedded inside a tooltip that appeared on hover, but as soon as I moved my cursor toward it, the tooltip disappeared. The interaction was impossible to complete.

To a human designer, that’s an obvious usability and accessibility problem. But the AI generated it without catching the issue.

I use AI to build products every day, and I’m excited about how it’s enabling designers to become builders, taking our ideas beyond prototypes and into production. But as our ability to build expands, so does our responsibility to ensure what we create actually works for everyone. Accessibility can’t be something we assume the agent has accounted for, or a box we trust it to check. It has to be independently verified.

Because progress means very little if the people we’re building for can’t come along with us.