Cover Image for What AI Gets Wrong About Accessibility
Cover Image for What AI Gets Wrong About Accessibility
Avatar for PixelPerfect
Presented by
PixelPerfect
Hosted By

What AI Gets Wrong About Accessibility

Google Meet
Registration
Approval Required
Your registration is subject to host approval.
Welcome! To join the event, please register below.
About Event

​AI build tools ship websites that look finished and fail screen reader and keyboard users in the same handful of ways. Alexandra audited her own AI-built site, client sites and a handful of others, built with different tools, and found the same failures on every one. She built an automated scanner, and it catches only a fraction: the rest a machine cannot see, which is why designers have to learn to see it, or they lock clients out.

​This session shows most common accessibility failures in AI-generated websites, what a screen reader user actually hears, and a six-point checklist with a fix prompt for each item that you can run on your own AI output.

​Key Takeaways

  1. ​The step AI skips. In a designer-led build, heading levels, carousel behaviour and alt text are annotated before handoff. AI builds skip that layer, and that’s where these failures come from.

  2. ​The AI failures that repeat. Styled words that split a sentence, headings animated letter by letter and read as “cap A, cap I”, stats read twice, and gradient text that fails at its midpoint. What each one sounds like to a screen reader user, and how to fix it.

  3. ​Hidden on screen is not hidden. The mobile menu in the desktop tab order, swipe cards still read aloud, slides announced all at once, and the moving strip on every AI site. How carousels, marquees and bottom sheets should be named, navigable and closable.

  4. ​Pictures without context. Live HTML imagery means no alt text to write, and a screen reader reads the product screenshot as loose text with no idea what it is. What to add so it makes sense.

  5. ​The ESCAPE checklist and its prompts. Six checks, ordered by how badly each one locks people out, each with a prompt that makes Lovable, Claude or Cursor fix it.

  6. ​What a machine can decide, and what it can’t. Automated scanners, including Alexandra’s own, can decide only 10 of 55 WCAG 2.2 AA requirements. The other 45 need a person, and that’s where clients get locked out.

​Speaker

​Alexandra Ciobanu is a design lead with 15 years in design, nearly a decade of them at Amazon, where she led design teams across NA, EMEA and APAC. She co-created Basic-Good-Great, a framework that raised visual accessibility from 80% to 95% across three organisations, and helped embed accessibility into Amazon-wide visual and UX standards. Today she runs Alexa C., a studio that builds websites and apps with AI tools and checks each one by hand against WCAG 2.2 AA. She built her own accessibility scanner and writes Basic Good Great, a newsletter about what she learns building with AI: the accessibility errors she finds, what AI gets wrong, and how to fix it.

​----

​More events and educational opportunities for UX and Product Designers at www.pixelperfect.world

Avatar for PixelPerfect
Presented by
PixelPerfect
Hosted By