Do real AI products let one chat session hop between 'AI designs the screen' and 'you describe the screen you want'?

This explores whether shipped AI products routinely blend two ways of making interfaces in one session: generative UI, where the AI decides what to build, and vibe coding, where you describe what you want and the AI builds it. The short answer is that the corpus doesn't settle it. No note here surveys shipped products to count how often the two modes appear together. What the corpus does have is a sharp account of why mixing them is harder than it sounds, and some evidence that the pieces are built to be combined.

The clearest framing comes from Moran, who separates the two modes by who starts the build. When the AI starts it (genUI), the system answers for its design judgment: did it pick the right interface? When the user starts it (vibe coding), the system answers for execution: did it build what I asked for? Does who initiates a UI element matter for accountability? This is why mixing them in one session isn't a cosmetic choice. Each switch changes what counts as a failure. Moran also argues the two serve different people. GenUI suits users who can't describe the interface they need, while vibe coding assumes a level of interface know-how most people don't have Can generative UI serve users who cannot specify interfaces?. A product that blends both serves both groups, but it also blurs who is accountable for what.

The research on generated interfaces shows the tension you'd expect at that seam. TaskArtisan found that AI-generated analysis widgets were clearer than chat but stiffer. Once a widget appeared, changing it mid-task meant going back to prompting, which pushed non-programmers toward thinking like engineers Do generated analysis UIs really work better than chat?. In practice, that is the moment a genUI session turns into a vibe-coding session, and the study treats it as friction rather than a smooth handoff. Execution also falls short. A benchmark of five generative UI tools found that about a quarter of the design reasons the tools stated never made it into the output, rising to a third for functional requirements Do generative UI tools actually implement their stated design rationales?. So even within one mode, intent gets lost between what the system says it will do and what it actually builds.

Looking sideways at the evidence, the building blocks for mixing exist. A model that mapped 66 generative AI tools found twelve recurring interaction patterns and noted that they combine, so tools already stack several input styles How do interaction signals reshape what generative AI actually receives?. Friedman's design levers point the same way. Capturing intent, shaping output, refining, and canvas editing all assume the user and the AI pass control of the interface back and forth Can task-oriented UI patterns reduce the friction of AI chat interfaces?. Feedforward design, which shows a preview of the likely output before you submit, could act as the bridge. You see what the AI plans to build and can correct it before control changes hands Can previewing AI output before submission reduce prompt iteration?.

Here's the less obvious takeaway. The real question may not be whether products mix the two modes, but whether they show you when they've switched. If the AI picks a layout and you then ask for a change, you've silently moved from 'judge my design' to 'follow my instructions.' Nothing in this corpus shows a shipped product making that switch visible, and that gap looks like a design problem worth naming.