What four choices actually decide how an AI builds an interface for you on the fly?

This explores how a CHI workshop proposal maps the choices involved in designing interfaces that AI builds on the fly. The four dimensions are **audience, constancy, control, and timeframe** How should generative UI be organized as a design space?. Read plainly, they ask four questions: who the interface is generated for, how stable it stays from one use to the next, who steers it (the user or the system), and how long a generated interface is meant to last, whether a throwaway screen or something that persists. The framework comes out of older work on adaptive and 'malleable' software, where the main tension was always how much the system should adapt on its own versus how much the user should shape it. Generative UI makes that old tension sharper.

The control dimension gets the most attention elsewhere in the corpus, under different vocabulary. Moran separates generative UI from vibe coding by asking who decides an interface should exist. When the AI starts it, the system is responsible for its design judgment. When the user starts it, the system is responsible for faithfully building what was asked Does who initiates a UI element matter for accountability?. Moran also argues that the AI-led mode serves people who can't describe the interface they want, because most people lack the interface literacy that vibe coding assumes Can generative UI serve users who cannot specify interfaces?. So 'control' is about more than user preference. It decides what counts as failure.

Constancy runs into a well-known design rule. Classic usability guidance says interfaces should be consistent. Generative systems produce different output each time, so users end up stating what they want rather than how to get it How should users control systems with unpredictable outputs?. Choosing a level of constancy is partly choosing how much of that unpredictability users have to live with. There is also a gap between what a tool says it is doing and what it delivers. One benchmark found that generative UI tools left about a quarter of their stated design reasons unimplemented, and 34% of functional requirements Do generative UI tools actually implement their stated design rationales?. A neat map of the design space doesn't guarantee that tools respect where you placed them on it.

Other maps in the corpus slice nearby territory differently, and comparing them helps. A review of 291 papers on human–GenAI interaction found six dimensions, including how the model gives feedback, how much control users have and how systems get evaluated How should designers structure human-GenAI interaction systems?. The IAI model looks underneath the interface. It asks what the model actually receives, meaning the prompt plus all the constraints added by clicks and settings, and it finds twelve recurring interaction patterns across 66 tools How do interaction signals reshape what generative AI actually receives?. The four-dimension framework is about what gets generated and for whom. These other maps are about the conversation that produces it. Reading them together gives a fuller picture than any one alone.