AI design tools failed WCAG Level A on all 15 test sites, and look finished anyway

AI design tools got good at the part that photographs well. The screens look finished. The spacing is clean, the palette is on brand, the components look like a real product. Then someone tries to use one with a keyboard, or on a slow connection, or when they type the wrong thing, and the whole thing falls apart. Let me catch you up on what the research found and what designers are doing about it.

The deep cut

• Polish hides the questions that matter. The ACM study found 138 AI screens nailed visual convention and failed support heuristics.
• Ask for accessibility and you still do not get it. AudioEye's 15 AI-built sites all failed WCAG Level A despite the brief.
• Put the AI's work where you can inspect it. Quincy Samycia's planner changes nothing until a human accepts the staged change.

The screen that looks done but is not

A 2026 ACM study scored 138 mobile screens from Figma, Banani, and Stitch against Nielsen's ten usability heuristics. The tools copied the look of modern apps well. They were weak on support-oriented heuristics, the parts that help a user when the happy path breaks. The authors call it usability by convention rather than intent.

Here is why that costs you. A finished-looking screen invites a finished-screen review. You end up debating button color while the task hierarchy, the error states, and the empty states never get checked. Nielsen Norman Group saw tools read the word "profile" as a social profile and build a big identity block that pushed the real task down the page. The tool matched a word to a pattern without asking what the page was for.

Accessibility is where it breaks hardest

AudioEye asked five AI tools to build 15 sites to WCAG 2.2 AA and tested them. All 15 failed WCAG Level A, the most basic tier, at an average of 55 issues per page. Worth noting AudioEye sells accessibility products, so check the raw study. But the point holds: asking for accessibility in the brief did not produce it.

The gap is a trust problem. AudioEye's companion survey found 81% of teams using AI tools believed their generated code met accessibility guidelines. It did not. Tab through every element, run a screen reader on the main task, and measure contrast on the rendered page, not the palette in the prompt. Screenshots will lie to you here.

Thinking before the AI builds anything

The designers writing up real workflows share one move: do the thinking first. One designer starts with a skill called /grilling, where the AI asks questions and pulls at the idea before any code gets written. Then /to-spec, /to-tickets, and prototype into four or five directions at once before committing to one. Diverging is cheap now, so decisions come faster.

The catch sits in the NN/g finding. When a brief got specific enough to produce good output, much of the design work had already happened before the AI touched it. That is the real skill. The tool is fast at generating. The scarce step is the judgment that goes in before, and the review that goes in after.

Making the AI show its work

Quincy Samycia replaced Asana, Toggl, HubSpot, Typeform, and Calendly with custom software he built using AI, in under 250 hours. The part worth stealing is not the savings. It is how he designed the AI assistance. His planner suggests priorities and due dates, but a separate staged-changes panel holds every suggestion until a human accepts it. Nothing commits on its own.

That boundary is a design decision. Suggesting a deadline and committing a team to it are different actions, and a good interface keeps them apart instead of burying them in a chat reply. The same logic runs through the Claude memory guides: notes are signs, hooks are locks, and you give the AI a way to check its own work before you trust it to run alone. Visible review is the product.

Three questions for your team

• What are we spending review time on? If it is spacing and color, we are reviewing where the tools are strongest and skipping where they fail.
• Can we prove our AI-generated screens pass WCAG Level A with keyboard and screen reader tests, or are we just confident like the 81% who were wrong?
• Where does our AI commit changes without a human accepting them first, and which of those should move behind a staged-review step like Samycia's?