Aarjav Pandya, who works on Firefox UX at Mozilla, named something every usability researcher has watched happen behind the glass: “Test participants tend to recognize when they are interacting with a prototype, and that awareness can change how they behave.” Writing on the Firefox UX blog on 5 August 2026 , he described building a Report Broken Site feature directly inside Firefox Android with Claude rather than faking it in Figma, “so the test reflects native fidelity rather than a simulation of it.”
That instinct — build it real, so nobody can tell it was cheap — is the whole prototyping profession’s problem in miniature. Designers used to show something rough on purpose: a grey box, a paper sketch, anything that visibly took an afternoon, because roughness was proof the thing could still be redrawn. AI has severed that proof from the object, the same collapse that pushed Vercel’s v0 and Figma Make to start opening pull requests against real repositories . A prototype built in three weeks can now look like something that shipped, and the polish that gets people to react honestly to it is also what stops anyone from asking it to change. Roughness was never a design choice; it was a byproduct of cost, and cost is exactly what AI prototyping deletes.
One project produced both halves of the mechanism at once
A July 2025 case study from Purdue researchers, published on arXiv a year ahead of the Firefox and NN/g accounts below, followed one vibe-coded project: an interactive data-analytics tool for highway traffic engineers. Shown sketches and low-fidelity designs first, the engineers “tended to refrain from detailed critiques, considering them preliminary representations rather than concrete implementations,” so early discussions “failed to capture detailed user feedback.” Given a shared link to the interactive, AI-built prototype instead, the same users “actively test the prototypes, uncover potential issues, and collaboratively ideate valuable additional features.”
That is one small, qualitative study of one project — no participant count, no controlled comparison — and it claims nothing beyond what happened on that team. But read next to Pandya’s account of a different team reaching for the same fix on a different product, it points at something neither source set out to measure: that a single change, fidelity no longer signaling cost, produces both effects on two separate audiences at once. Users stop treating the thing as a draft. So does everyone else in the room.
Roughness was never a design choice; it was a byproduct of cost, and cost is exactly what AI prototyping deletes.
Getting honest feedback costs the room its permission to redraw
Nielsen Norman Group’s Megan Chan , citing the Purdue paper on 11 September 2026, tells the same story from the other side. A Ramp senior product designer, Pavan Garidipuri, used Cursor to take an expense-policy editor “from concept to demo in only three weeks,” surfacing a real usability problem — confusing edit tracking — a static mockup likely would have missed. Chan’s warning sits right beside the praise: “The danger is that its polish might cause people to stop working to improve it. AI-generated prototypes may look complete while introducing all sorts of assumptions and inaccuracies.” Stakeholders start asking, reasonably, why something that looks finished isn’t already live.
This is where the reader stake gets concrete. A product designer who used to open a review with a grey-box wireframe and now sends a clickable link gets sharper usability findings by Tuesday — and by Wednesday has lost the unspoken permission to rebuild the thing from scratch, because nobody in the room can see anymore that it was cheap. Chan’s own prescribed fix isn’t to make prototypes rougher again; it’s to tell stakeholders how the thing was made, replacing a visual cue with a verbal one.
The counterargument: this was always about sunk cost, not looks
The strongest objection: the fidelity trap was never about appearance — it was about the week spent in Figma, and AI destroys exactly that. A prototype regenerable in an hour is more disposable than a hand-built file, whatever it looks like — a communication problem, not a perceptual one. The data leans that way: UX Tools’ Spring 2026 survey of 1,478 designers found only 32.8% trust AI output for production even with review, and just 1.4% without — not the profile of a profession mistaking prototypes for finished products. Designer Fund’s AI in Design Report captures the same unease from inside a team: one designer describes AI output as easy to mistake for done — “get an answer back…and think: This is great” — without proper validation. No study yet measures whether teams actually revise AI-generated prototypes less often than hand-built ones; that comparison doesn’t appear to exist. A related worry runs the opposite direction: AI stand-ins for real participants have their own honesty problem, skewing agreeable rather than critical, as research on synthetic users in usability testing has found .
What’s missing isn’t rougher prototypes. It’s a replacement for the cue roughness used to carry — some way of saying “this cost an hour” out loud, since nothing in the artifact says it anymore.
