Generative AI can now assemble user interfaces from the patterns already stored in a company's design system. The tools can produce a screen that matches the library, with the right buttons, spacing, and components. That ability, however, answers only part of the question. Matching a system is necessary for consistency, but it is not enough to make an experience good, and the decisions that remain are the ones that still rest with human designers.
The central problem is that the same visual response can suit one moment and irritate another. A confetti animation may feel rewarding after a completed task and feel intrusive during a routine action. Designers therefore have to decide when an experience should celebrate, when it should surprise, and when it should stay quiet. Expressive choices, the argument runs, need a justification that goes beyond the idea that they are more delightful.
Several design frameworks help explain why these choices matter. Don Norman's work on emotional design describes three levels: appearance, use, and meaning. Writer Fessenden draws a distinction between surface delight and deeper satisfaction that comes from a smooth, well-designed workflow. Figma's guidance for teams using AI adds a practical point: the context supplied to a generative tool shapes the output it produces, so the quality of the instructions and background material affects what the tool builds.
Real product examples show how these trade-offs play out. Duolingo varied its illustrated headers by purpose after consistency tests, rather than applying one decorative style everywhere. Mailchimp's voice guide keeps the brand voice fixed while letting the tone adapt to the reader's situation, a distinction that applies just as well to visual feedback. Google's Material 3 Expressive work included a larger, closer Send button that sped up how quickly people located it, a case where expressive styling and usability pointed the same way.
Other examples cut the other way. In a Google test, scattered album artwork looked exciting to reviewers but hurt usability. Asana's animated celebrations drew at least one user request to disable them. These cases suggest that an interface can look lively in a design review and still make everyday tasks harder, which is why the source material argues that generated screens need evaluation beyond whether they match the library.
Accessibility adds a further constraint. W3C guidance at the AAA level says users need ways to disable animation triggered by interaction. Removing motion, however, must not strip away clear feedback or visual hierarchy. A designer who turns off an animation still has to show users that an action succeeded and where their attention should go next.
Several questions remain open. Designers still need to determine which events merit a fuller acknowledgment and which call for only brief feedback. It is unclear whether AI-generated alternatives can judge the occasion, as opposed to simply reproducing appearance. The framework also leaves open how teams should evaluate generated interfaces against goals beyond conformity to the library. The source material's claims about how users reacted to Zillow and Asana remain unverified beyond the examples cited, so those specific reactions should be read with that caveat in mind.