Haim Repael
CLOSE

Essay

The Hand We Can't Automate

What AI-generated interfaces reveal about the people we forget to design for, and why that's not a technical problem.

Haim Repael · 3 min read

STAR AT A GLANCE
Situation
AI design tools close the gap from idea to interface fast, but nobody checks what they produce for people who can't see the screen.
Task
Test what a normal prompt actually ships, with no accessibility hints.
Action
Asked Claude Design to build a fictional SaaS landing page, Orbitly, and inspected the structure beneath the visuals.
Result
It looked genuinely good, and was structurally a catalog of the most common accessibility failures. Defaults, not edge cases.

There's a moment I keep coming back to.

A user, blind since birth, is completing a form with a screen reader. The site was built with accessibility in mind. Modern tools. Boxes checked. Contrast passed. The launch went fine. I've been in the room when this ships.

And she's stuck. Not because the form is broken, but because when she misses a field, the error is announced before the label that explains it. Backwards.

The missing piece has a name: aria-describedby, the attribute that declares an error belongs to its field. Visually they're connected, close together, both red. Structurally they're strangers. The fix is one attribute. She's been on this form for eleven minutes.

SituationA genuinely good tool, and a question about its defaults

I want to be fair before I'm critical. Tools like Claude Design matter. For a founder without a design team, or a PM who needs a wireframe by Thursday, the gap between I have an idea and I have something people can react to used to be wide. These tools close it fast. You can kill bad ideas before they're built. More people with design judgment get to participate.

I want this to work. The accessibility field benefits from good tools that think about more people by default. The question is whether the default is doing enough.

TaskWrite the prompt everyone writes

So I ran a test. I asked Claude Design to build a landing page for a fictional SaaS product called Orbitly: a hero with a rotating 3D globe, a pricing table, a competitor comparison with star ratings, a multi-step booking modal with form validation, an interactive calendar, a logo carousel, social icons and a newsletter signup.

A normal prompt. No mention of WCAG. No accessibility hints. Just: build me this.

ActionLook underneath what renders

What came back looked good. Genuinely good. Then I looked at the structure.

  • The globe. The centerpiece, built as a <canvas> with no label and no role. For a screen reader user it doesn't degrade. It's absent. The thing the design spent the most effort on renders as silence.
  • The pricing table. It looks like a table: columns, checkmarks, tier names. It's built as a list. No headers, no relationships, no way to know what's being compared to what. The visual communicates hierarchy. The code communicates nothing.
  • The form. The error appears in red, right below its field, with no declared connection. The screen reader announces the error and the field, and doesn't know they're the same problem. The opening story isn't hypothetical. It's the default output.

ResultNot edge cases. Defaults.

Structurally, Orbitly was a catalog of the most common accessibility failures in modern web design. AI optimizes for what renders, and what rendered looked right. The failure lives underneath, invisible to anyone who evaluates the output by looking at it. Which is everyone.

What travelsIt's a representation problem

Most critiques stay technical: wrong ARIA, unchecked contrast, missing semantic HTML. Those are real, and they're the easier conversation.

The harder one: AI responds to the prompt, and prompts are written by people who describe the users they can picture. The person navigating without sight isn't in the prompt. Neither is the undiagnosed reader who loses their place in dense text, or someone who developed low vision at 68 and is trying to use a banking app.

That's why "better models" won't close this gap alone. It's not a capability problem. It's a representation problem.

And it's two problems, not one. Once a site exists, content changes constantly, and automated tools that scan and remediate at scale solve something no human team can. But creation is different. At the moment someone writes a prompt, the canvas loses its label, the table loses its structure, the form loses its relationship. No remediation layer reaches that moment, because there's nothing yet to remediate.

More things are getting built, faster, by people who've never had to answer for them.

Someone is still stuck on that form. And no one on the team sees it.