Prototype Fidelity After Cheap Polish

Sources#

Rethinking Prototype Fidelity in the Age of Generative AI: When High-Fidelity Becomes Cheap, It's Time to Revisit Our Foundational Design Wisdom

Summary#

Christopher Hundhausen (Oregon State) in interactions 33.4 (July 2026, practitioner-opinion — a magazine Features piece and an explicit call for research, containing no new measurement): the decades-old "sketch first, code last" playbook rested on an economic fact that GenAI has removed, and the field is changing practice anyway without the evidence to justify it.

The argument in one move: the classic prototyping wisdom was never primarily about fidelity — it was about a correlation between polish and effort, and that correlation is now broken.

"The signal that such a prototype sends to test users — 'this is polished, therefore significant effort was invested, therefore I should be gentle with my feedback' — no longer correlates with reality. Yet the psychological barrier to providing feedback on a polished product likely remains."

The asymmetry is the whole finding: the cause of the suppression is gone, the suppression probably isn't.

What the old evidence actually said#

Worth stating precisely, because the reasoning depends on which of the three classic results is doing the work:

Schumann et al. (1996) — low-fidelity prototypes elicit significantly more design feedback than polished ones. This is the load-bearing one, and it is a claim about feedback volume, mediated by users' reluctance to criticize apparent effort.
Virzi, Sokolov & Karis (1996) and Walker, Takayama & Landay (2002) — low- and high-fidelity prototypes uncover substantially the same usability problems, independent of medium.

Note what the second pair implies: fidelity was never claimed to affect what you find. The low-fi-first playbook therefore rested on exactly two legs — cheaper to make and change, and more feedback per session. GenAI removes the first outright. Hundhausen's question is whether the second survives without its mechanism.

The inverted economics, and the crossover nobody mentions#

Figure 1 plots development time/cost against visual polish for two regimes: Traditional (2000–2023), rising steeply and S-shaped to the top of the chart, annotated "High cost = cautious commitment to code"; and GenAI-Enabled, nearly flat across the entire fidelity range, annotated "Low cost = rapid iteration possible."

The detail worth carrying — visible in the artifact, unremarked in the prose — is that the two curves cross just to the right of the Low-Fi tick: at the very lowest fidelity, GenAI is marginally more expensive than traditional methods. Paper is still the cheapest way to make something rough. The GenAI advantage is not "cheaper prototypes," it is "flat cost across fidelity," which is a different and more interesting claim: fidelity stops being a budget decision and becomes a purely methodological one.

(Figure 1 is schematic — no units, no data points, no underlying study. It illustrates the argument rather than evidencing it. Both images in this source were viewed directly during compile.)

Boehm's spiral, back from the dead#

The sharpest historical move: Hundhausen reads GenAI as a possible resolution to a 1988 failure. Barry Boehm's spiral model included evolutionary prototyping — prototypes iteratively refined into production systems, eliminating the prototype-to-production rebuild. "This was the holy grail." It never worked, and Boehm himself named why: the "spaghetti code difficulties" of prototypes that lacked proper architecture and grew unmodifiable.

The question is then posed as a clean fork:

"If AI tools generate well-architected, maintainable code even under rapid iteration, evolutionary prototyping might finally be viable. However, if GenAI-generated prototypes prioritize immediate functionality over long-term maintainability, we're simply automating the same mistakes."

This vault can already say which fork the evidence points down, and the article does not know it. Agentic Technical Debt is Boehm's warning restated for agentic coding, with the additional claim that the debt compounds rather than accumulating — each session re-deriving architectural decisions — and surfaces late as a forced rewrite. Security debt in agent-generated code and Agent-Generated Test Quality point the same direction on adjacent axes. The corpus's answer to "does GenAI generate well-architected code under rapid iteration" is: not by default, and the discipline that fixes it (persistent architectural context, Agent Context Files) is exactly the discipline a rapid-prototype workflow skips. Evolutionary prototyping is not obviously more viable than in 1988 — the cost of producing the code fell without the cost of maintaining it falling with it.

The three research questions#

Stated as a research agenda, and each comes with a "what practitioners can do now" that is more actionable than the question:

Does the Schumann effect persist when effort is cheap? Can you effectively signal "this was quick, be brutally honest"? And critically — what kind of feedback is lost, strategic (fundamental workflows) or tactical (polish suggestions)? Now: explicitly tell users how fast the prototype was generated; try deliberately rough aesthetics on functional prototypes.
Can rapid iteration with high-fidelity prototypes finally work? If initial feedback is more guarded but the rebuild takes minutes, does the guardedness matter? Now: track which prototypes evolve into production and which need rebuilds, and what distinguishes them.
Can the dimensions of fidelity be decoupled? Sophisticated interaction logic with rough aesthetics; beautiful interfaces with placeholder function. The precedent is Silk (Landay & Myers, 1995), which achieved this but required learning a specialized visual language — "could GenAI prompting achieve what previously required purpose-built prototyping environments?" Now: build prototypes polished in one dimension and rough in others.

Question 3 is the one the vault can partly answer, and the answer is yes. Design by Selection records a working designer at Anthropic already treating fidelity as a control knob — wireframing first specifically when visuals would distract from structural feedback — reached by practice rather than by tooling, through prompting rather than a specialized environment. That is Hundhausen's open question resolved in the affirmative as a capability; what remains open is his actual question, which is whether it changes feedback quality.

Table 1, and a discrepancy worth preserving#

Table 1 lays out the design space across visual polish, speed to create, data gathered, and best-for. Its contribution is locating GenAI Low-Fidelity as the unexplored cell: low polish, fast (hours to days), and — uniquely — able to gather behavioral data and performance issues that paper cannot, while keeping the unpolished look that protects feedback volume. That combination is the article's actual design-space proposal, and it is what nothing in the traditional 2×2 offered.

The table also predicts its own obsolescence for one row: Traditional High-Fidelity "may be largely replaced by GenAI high-fidelity prototyping."

Discrepancy, verified by viewing the image: the prose states Table 1 "compares five prototyping approaches." The published table has four rows — Paper, GenAI Low-Fidelity, GenAI High-Fidelity, Traditional High-Fidelity. The artifact is what shipped and the artifact has four. Recorded rather than smoothed, because a future reader counting cells against the prose will otherwise assume a row was lost in extraction.

Practice has already moved without the evidence#

The closing empirical note, and the most checkable claim in the piece: Chen & Yoon (CHI 2024) found designers in the software industry already moving away from low-fidelity paper prototypes toward digital-first, high-fidelity approaches — driven by better tooling, hybrid work, and corporate preference for polished presentations. Note the date: this predates the GenAI economics Hundhausen describes, so the abandonment of the low-fi-first playbook was already underway for unrelated reasons and GenAI is accelerating a shift it did not start.

That reframes the piece's urgency argument. The risk is not that practitioners will follow decades-old research off a cliff; it is that they have already stopped following it, and no one has measured what replaced it.

Connections#

Prototype Over PRD — the adjacent move, one level up: that page removes the spec in favor of the prototype, this one asks what fidelity that prototype should be built at and what feedback it will elicit
Design by Selection — the practitioner answer to research Question 3: fidelity already used as a control knob, wireframe-first when visuals would distract, achieved through prompting rather than a Silk-style specialized environment
Implementation Abundance Inverts Product Work — the same economic shift read from the product side: when implementation is free the expensive step moves elsewhere. Hundhausen's version locates the new cost in methodology (which fidelity, what feedback), Ambrosino's in curation and taste
Agentic Technical Debt — the corpus's answer to the evolutionary-prototyping fork: Boehm's 1988 spaghetti-code warning restated for agentic coding, with the debt compounding rather than accumulating
Vibe Coding vs. Agentic Engineering — the practice this article calls "vibe coding" and proposes teaching as a distinct professional skill; that page's distinction is precisely what the education section is missing
Why AI Lags at Design — the capability-side counterweight: cheap polish is not cheap design, and the gap this article treats as an economics story has a model-capability half

Open Questions#

Does the Schumann effect survive the loss of its mechanism — do users still soften feedback on polished artifacts once told the artifact took an hour? The article's Question 1, and the field's key unknown.
Which feedback is lost to polish: strategic (workflow, information architecture) or tactical (visual polish)? The distinction determines whether the low-fi-first playbook mattered for the reasons its advocates claimed.
Does GenAI-generated prototype code actually evolve into production, or rebuild? Hundhausen poses it as open; this corpus's debt evidence suggests rebuild, but no source measures prototype-to-production survival directly.

Sources#

Rethinking Prototype Fidelity in the Age of Generative AI: When High-Fidelity Becomes Cheap, It's Time to Revisit Our Foundational Design Wisdom — Christopher Hundhausen, interactions 33.4 (2026), pp. 42–45, DOI 10.1145/3815553 (practitioner-opinion, CC-BY 4.0). A Features essay and research call, no new measurement; the empirical citations are to Schumann 1996, Virzi 1996, Walker 2002, Boehm 1988, Landay & Myers 1995, and Chen & Yoon 2024. Figure 1 and Table 1 were served as GIFs and both were viewed directly during compile (the two-pass image rule): the figure's Low-Fi crossover and the table's four-row count are confirmed against the artifacts, and the prose's "five prototyping approaches" is a prose/artifact discrepancy, not an extraction loss. Fetched via headless browser after dl.acm.org returned Cloudflare 403 to WebFetch — see the Source Notes entry