Generative Interface

A generative interface is a user interface whose layout, components, and interaction flow are produced at runtime by a generative model rather than being hard-coded by a developer in advance. In a traditional interface, the design team specifies every screen, button, and transition; in a generative interface, the model decides what to render in response to the user's task, context, and prior turns. The result is an interface that adapts in shape as well as content — the same product might present a chart, a form, a step-by-step wizard, or a free-form conversation depending on what the model judges most useful for the current moment.

Mechanically, a generative interface system has three parts: a model that decides what to render (often an LLM with a constrained output schema that maps to UI components), a renderer that turns the model's structured output into pixels (React, SwiftUI, a custom DSL, or a wire protocol like A2UI), and a feedback loop that lets the model observe the user's subsequent actions and adjust. The model output is typically a tree of components with properties — a chart with a data array, a form with field types, a list with action callbacks — rather than free-form HTML or pixel coordinates. Some systems also let the user steer the regeneration by asking for a different layout ("show this as a timeline instead") or by editing the rendered result directly.

The advantage of generative interfaces is personalization and task-fit at a granularity that hand-coded UI cannot match: the same product can present a financial dashboard to an analyst and a plain-language summary to a novice without two separate code paths. The cost is loss of designer control, predictability, and accessibility — when the model chooses the layout, brand consistency and screen-reader semantics become probabilistic rather than guaranteed, and a model that produces a clever but unusable component creates a worse experience than a deliberately minimal hand-coded one. Latency is also higher, since the model has to run before the first pixel is drawn, and the cold-start experience is often worse than a static page. There is also a real risk of the model regenerating state mid-interaction in ways that confuse the user.

Open questions include how to evaluate the quality of a generative interface — traditional usability metrics assume a fixed design — and how to enforce accessibility, privacy, and brand constraints when the model is choosing what to render. The relationship between generative interfaces and traditional design systems is also unsettled: some teams treat the design system as a constraint set the model must respect, others treat it as a library the model is free to ignore. The longer-term question is whether generative interfaces will replace hand-coded UI for most products, or remain a layer used in specific high-variance contexts (data exploration, agentic tools, dashboards) where the cost of pre-building every screen is prohibitive.