Generative AI promises speed: copy, images, alt text, translations in seconds. But speed without validation can produce technically inaccessible or experientially exclusive interfaces — even while superficially ticking checklists.
Automatically generated alt text is often generic ("image of a person") or wrong. For screen reader users, this is worse than no alt text — it gives false confidence. Every meaningful image deserves human review or, at minimum, the ability for the author to correct it.
Semantic structure (heading hierarchy, landmarks, labels) is often ignored when AI generates entire pages. A wall of text without H2/H3, buttons without labels, forms without label-input association — classic problems amplified by generation speed.
Contrast and readability: AI-"suggested" palettes can fail WCAG. Decorative fonts generated for hero sections can be illegible. The design system exists precisely to prevent each generation starting from scratch ignoring standards.
Inclusive language and plain language are not automatic. Models trained on generic web produce gendered copy, exclusive idioms, unnecessary complexity. For public services, healthcare, welfare — where cognitive accessibility is critical — human review is mandatory.
Recommended workflow: AI generates draft → designer/reviewer verifies accessibility (automated tools + manual test) → test with disabled users or a11y experts → publish. Skipping central steps for "velocity" is ethical debt paid in exclusion.