AI can now build your app's screen on the fly instead of a designer fixing it in advance — what choices actually shape that?

This explores how researchers map the possible ways to design generative UI, meaning interfaces that AI assembles at runtime instead of designers locking them down ahead of time. The most direct answer in the collection is a four-part framework: audience (who the interface is generated for), constancy (how stable or changeable it stays), control (who steers it, the user or the system), and timeframe (when and for how long generation happens) How should generative UI be organized as a design space?. These dimensions didn't come out of nowhere. They grew from older work on adaptive and 'malleable' software. Underneath all four sits one long-running tension: how much should the system adapt on its own, and how much should the user stay in charge?

A less obvious dimension is origin, meaning who decides that an interface should exist in the first place. Moran draws a sharp line between generative UI, where the AI chooses what to build, and vibe coding, where the user asks for something specific Does who initiates a UI element matter for accountability?. The difference matters because it changes what counts as failure. If the AI started it, it is responsible for good design judgment. If the user started it, it is responsible for building what was asked. That is also why generative UI can help people who couldn't describe the interface they need: most users lack the interface vocabulary that vibe coding assumes Can generative UI serve users who cannot specify interfaces?.

If you step back from UI specifically, broader frameworks for human-AI interaction cover overlapping ground under different names. A review of 291 papers found six recurring dimensions: purpose, how the model gives feedback to users, user control, engagement level, application domain, and evaluation How should designers structure human-GenAI interaction systems?. 'Control' shows up in both frameworks, which suggests it is the dimension the field keeps returning to. Another approach maps the plumbing instead: it treats the interface as the thing that packages your prompt plus your clicks, sliders, and constraints into what the model actually receives. Applied to 66 tools, this produced twelve reusable interaction patterns How do interaction signals reshape what generative AI actually receives?. There is also a principles-based view, which starts from the fact that generative interfaces are unpredictable. Users describe what they want, not how to get it, and that breaks classic rules like 'keep things consistent' How should users control systems with unpredictable outputs?.

Here's the part you might not expect. A design space describes what could be built, not what the tools actually deliver. One benchmark found that generative UI tools left about a quarter of their stated design rationales unimplemented, rising to a third for functional requirements, and they recognized only half the UX principles written into the prompts Do generative UI tools actually implement their stated design rationales?. So choosing a point in the design space, such as 'high system control, adapts per user,' doesn't guarantee the system will carry it out. On the governance side, AWS's Cloudscape design system adds obligations that sit upstream of all these dimensions: disclose when content is AI-generated, get consent before using customer data, and let users override the system How should generative AI design embed responsible AI principles?.

The collection has only one framework built specifically for generative UI. Everything else comes from neighboring human-AI interaction research. If you want to dig in, start with the four-dimension paper. Then read Moran's origin distinction, which works as a fifth axis the framework doesn't name. Finally, look at the benchmark to see where the map and the actual products don't match.