Every creative field touched by generative AI eventually needs the same unglamorous follow-up study: not “can the tool do X,” but “what are people actually doing with it once the novelty wears off.” A new paper does that for prompt-to-design generators (P2dGens) — tools that turn a text prompt into a UI prototype — through semi-structured interviews with 17 professional UI designers.
Ideation accelerators, not autopilot
The study’s central finding cuts against the more breathless version of the “AI replaces designers” narrative: designers report using P2dGens primarily as ideation accelerators rather than end-to-end solutions. Nobody in the sample described prompting their way to a finished, shippable interface. Instead, the tools get reached for at the front of a process — generating a spread of directions quickly, then handing off to the designer’s own judgment for everything that follows. Participants also described a real variety of individual and collaborative prompting strategies, suggesting the field hasn’t converged on a single “right” way to work with these tools yet — it’s still being figured out designer by designer and team by team.
Efficiency gains, and a real anxiety underneath them
The attitudes the researchers surfaced were genuinely mixed, not uniformly positive. Designers reported real gains in efficiency and inspiration — faster exploration of a design space than sketching every direction by hand. But that came paired with explicit concern about loss of personal creativity and design quality, the same tension this site has tracked across other creative fields adopting generative tools: a capability that speeds up exploration can also, if leaned on too heavily, hollow out the part of the craft that was actually generating the designer’s distinctive voice in the first place.
Organizations are improvising rules as they go
Perhaps the most practically useful finding for anyone managing a design team: the study observed emerging governance practices inside organizations adopting these tools, with varying attitudes from one workplace to the next about what’s acceptable — some teams treating P2dGen output as a shared, reviewable draft, others more informal about it. There’s no settled industry norm yet for things like disclosure, review standards, or how much prompt-generated work can ship without a human redesign pass, which tracks with how recently these tools have become genuinely capable rather than novelty-grade.
Why this fits a bigger pattern
This isn’t an isolated result. Google’s own research team has studied a closely related problem with PromptInfuser, finding that how tightly an AI tool couples to a designer’s existing workflow changes how useful — and how trusted — it ends up being; a tool that feels bolted onto the side of a workflow gets used differently than one that’s genuinely integrated into it. The P2dGens study’s interview-based approach adds the missing piece that benchmark papers don’t capture: not whether the tool is capable, but how actual working designers are choosing to use that capability, and what they’re worried about losing while they do. This is an unreviewed preprint.