When an AI decides what interface to show you instead of you asking for one, who benefits — and who loses the ability to check its work?

This explores what changes about generative UI's value when the AI chooses what interface to build, compared with when the user describes the interface they want and the AI builds it. The corpus has one note that takes this on directly, plus several that help show what the shift costs. The central argument comes from Moran: when the user requests, genUI is mostly a faster way to build something, and it only helps people who already know what a good interface looks like. That's the hidden assumption behind vibe coding. When the AI decides, genUI does something different. It serves the much larger group of people who can't say what they want, because the AI is now acting as the designer (Can generative UI serve users who cannot specify interfaces?).

The catch is that this shift also removes the yardstick. When users spell out their needs, you can at least check whether the tool delivered them. A benchmark of genUI tools found that even with explicit design reasoning written into the prompt, about a quarter of it never made it into the result. For functional requirements the figure was about a third, and the tools recognized only half of the UX principles included in the prompts (Do generative UI tools actually implement their stated design rationales?). If tools drop that much of what they're told, the AI-decides model asks for more trust, not less. There's no spec left to compare the output against, so the user can't easily tell whether the AI made a good design call or quietly skipped something. Socher's account of reward hacking puts the risk in general terms: AI tends to satisfy what was literally said rather than what was meant (Why do AIs keep gaming rewards instead of serving intent?). In the AI-decides model, almost nothing is said, so the whole value depends on the AI guessing intent correctly.

The shift also changes what the user's input is. The IAI model argues that what a generative AI actually receives is never just the prompt. It's an 'augmented instruction' that bundles the prompt with constraints from the interface itself (How do interaction signals reshape what generative AI actually receives?). Once the AI is designing the interface, it is partly writing its own future inputs, because the buttons and fields it chooses limit what the user can ask for next. So the AI's decision isn't a one-off design choice. It shapes the rest of the conversation.

On the human side, the user's job changes from describing what they want to judging what they got. Simkute and colleagues show this pattern is where automation tends to lose productivity: people move into evaluation roles, and the easy tasks get easier while the hard ones stay hard (Does generative AI productivity actually improve or decline in practice?). Krakowski's split between complementary and competitive AI adds a longer-term question (Does passive GenAI use risk causing worker deskilling?). If users passively accept AI-chosen interfaces, they may never build the interface literacy that would let them push back.

The corpus doesn't yet have user studies that compare the two modes head to head, so how this trade-off plays out in practice is still unmeasured. The takeaway is that letting the AI decide doesn't just add convenience. It moves the value from building faster to judging well, and it moves the risk from 'the tool ignored my spec' to 'there was no spec to ignore.'