Generative UI

Every app and every website has an interface: buttons, input fields, tables, cards. Normally, developers define this interface in advance, and it looks the same for every user. With a Generative UI, this is different. Here, a program with artificial intelligence decides only at the moment of use which elements are displayed. If someone asks about the weather, a weather card appears. If someone asks about a flight, a booking form appears. The interface is thus generated rather than retrieved, which is why it is called “generative.”

Why chat windows alone are not enough

Most AI applications today respond with text. Text is flexible, but for many tasks it is a poor tool. Anyone who wants to pick an appointment doesn’t need a paragraph with seven dates, but a calendar to tap on. Anyone comparing numbers understands a chart faster than a bulleted list. Generative UI is meant to close this gap.

For companies, this holds an economic promise. Traditional software has to build every conceivable use case in advance as its own page. An online shop has filters for price, brand, and size because someone programmed them. A generative interface could instead invent fitting filters for each search query. This saves development work and feels more personal.

At the same time, this breaks with an old principle of software design. Until now, the rule was: a good interface is predictable, the button is always in the same place. If the interface constantly changes, this reliability is lost. This is exactly what the current discussion around Generative UI is grappling with.

Building blocks instead of freely painted screens

A common misconception is that the AI paints the interface like an image. It does not. In practice, the language model is given a fixed toolkit of building blocks: table, chart, card, slider, dropdown list. Developers have programmed these building blocks in the normal way. The model only chooses which of them to use and with what content to fill them.

You can picture it like a set of building blocks in a box. The shapes are ready and waiting, and the AI reassembles them anew for each request. Technically, the model outputs a structured description for this, usually in a machine-readable format like JSON. The app reads this description and builds the visible interface from it. This detour is important, because it means the AI cannot display anything the developers haven’t allowed.

Often there is a second step as well. The interface is not just a display; it can also trigger actions. If someone clicks “confirm appointment,” the interface reports this back to the model, and the dialogue continues. Interface and conversation thus merge into a single flow.

Where generative interfaces are already appearing

Early examples can be found in major chatbots. When ChatGPT or Google Gemini display a price chart for a stock question or a map for a travel question, that is a simple form of this. Search engines are also experimenting with arranging results differently depending on the query. For developers, there are ready-made toolkits, such as Vercel’s AI SDK, which comes with exactly these kinds of building blocks.

In business news, the term usually comes up in connection with the question of how software will be sold in the future. If interfaces arise automatically, value shifts away from an app’s appearance and toward the data and functions behind it. This particularly affects providers of enterprise software, whose product has so far been strongly defined by its user interface.

Generative UI should be distinguished from two similar terms. “Vibe coding” means that an AI writes the program code of an app, which then remains fixed. Classic personalization, in turn, selects from prefabricated variants, such as a homepage for new customers. Generative UI, by contrast, creates the view anew each time, at the moment of the request. Whether this will catch on remains open, since users must place trust in an interface they may never see again in that exact form.