PaperlayerView exact revision on arXiv

Source-linked AI summary

One Tool, One Taste? How Vibe Coding Trades Collective Diversity for Individual Creativity

Léonard Boussioux, Ziyi Zhao, Kanghyun Cho

arXiv:2609.38183v1cs.HCcs.CY

TL;DR

The paper examines whether websites made with a shared vibe-coding tool converge visually despite distinct business briefs and an originality incentive. It combines cohort-level visual analysis with process and survey evidence, finding substantial visual concentration and a mismatch between subjective experience and distributional distinctiveness.

  • Problem

    Evidence is limited on whether shared generative production homogenizes visual artifacts despite divergent inputs and explicit originality incentives.

    Source: §1 · p. 2

  • Method

    The study analyzes websites from 73 graduate students alongside second-level, time-aligned process coding of recorded build sessions.

    Source: §1 · p. 2, §3.6 · p. 9

  • Results

    Six visual archetypes account for 64 of 73 sites (88%), and the conclusion reports that most builders had similar sites whose similarity was invisible to them.

    Source: §1 · p. 2, §8 · p. 30

  • Takeaways & Limitations

    The findings document within-cohort visual concentration and a mismatch between builders’ subjective experience and their distributional distinctiveness.

    Source: §8 · p. 30

    Survey findings cover 58 of 73 sites, with n=57 for the authorship item.

    Source: §7.4 · p. 29

Abstract

from arXiv · show

Vibe-coding systems turn natural-language instructions into deployed websites, but little is known about the diversity of designs produced within a shared production context. We study 73 promotional websites created by graduate students with Lovable for distinct real businesses under a graded assignment that rewarded original design. We represent each homepage using DINOv3 embeddings and analyze pairwise cosine similarity, effective diversity, and descriptive cluster structure. A representative UMAP--HDBSCAN specification assigns 88\% of sites to six interpretable clusters, while sensitivity analyses yield between three and twelve clusters. In a process subsample of nine recordings, preliminary AI-assisted coding documents recurring generate-and-check behavior, incomplete transmission of verbalized intentions into prompts, and infrequent explicit evaluation of originality. Among survey respondents, perceived human control, satisfaction, and perceived quality show no detectable association with embedding-based cohort atypicality. These results document within-cohort visual concentration and a mismatch between subjective experience and distributional distinctiveness. We call this condition authored ignorance. Alongside early evidence that the collective cost of generative production reaches visual design, we contribute a measurement strategy that joins outcome, process, and perception for interactive artifacts, and design implications running from originality feedback to deliberate friction.

1 Introduction

The introduction asks whether a shared vibe-coding system narrows visual design despite divergent inputs and originality incentives, how its defaults shape shipped sites, and whether builders perceive convergence. It presents a study joining website outcomes, build processes, and builders’ perceptions.

1 Introduction

  • The study analyzes 73 graduate students’ websites for distinct real businesses on Lovable, combining visual embeddings, surveys, and nine recorded build sessions.The assignment explicitly rewarded creative, original design that avoided generic AI aesthetics.Source: §1 · p. 2
  • Six recurring visual archetypes contain 64 of 73 sites (88%), while nearest neighbors share 49% visual similarity versus 23% across all pairs.The authors describe the effective number of distinct designs as roughly a dozen, while noting residual variety.Source: §1 · p. 2
  • Across recorded sessions, builders followed a recurring generate-and-check loop, while some voiced intentions never reached prompts and requests narrowed to generic adjustments.Unspecified design decisions were left to production-stack defaults.Source: §1 · p. 2
  • Originality appeared in only 3.4% of 440 spoken verdicts, and even visibly flagged defects often shipped without verification.The authors characterize acceptance as fast, generic, and frequently absent.Source: §1 · p. 3
  • Authorship, satisfaction, and perceived quality were uncorrelated with measured originality, despite students being instructed to avoid generic AI aesthetics.The authors call this mismatch “authored ignorance”: felt authorship without a way to perceive an artifact’s genericity against the cohort.Source: §1 · p. 3
  • The paper contributes field evidence on visual-design diversity, a joint outcome–process–perception measurement strategy, and design implications tied to documented mechanisms.Proposed implications include originality feedback, divergence-promoting generation, and deliberate friction.Source: §1 · p. 3

2 Background

The background connects collective creativity, vibe-coding delegation, and computational visual-similarity measurement to frame cohort-level visual concentration and the gap between directing an artifact and judging its distinctiveness. It also clarifies that similarity and atypicality measures describe a selected representation and reference set, not historical novelty, objective originality, or a literal count of designs.

2.1 Individual Creativity, Collective Diversity

  • Individual creativity concerns an artifact’s novelty and usefulness, whereas diversity describes the distinct types a population occupies; AI can raise individual outcomes while narrowing collective outputs.The literature also finds that effects on diversity depend on where AI enters the creative process.Source: §2.1 · pp. 3–4

2.2 Vibe Coding as Delegation

  • Vibe coding turns natural-language intent into working software while allowing inspection and edits; “authored ignorance” describes feeling that one directed an artifact without a comparison set for judging its distributional position.The study measures perceived human control, satisfaction, and quality—not perceived originality or awareness of genericity—so authored ignorance remains an interpretation for future testing.Source: §2.2 · pp. 4–5
  • The process account treats translation from intention to prompt and evaluation of output as possible gaps, but its indicators do not directly measure those gaps or establish causal links.Reported indicators include conditional transmission of verbalized intentions, verdict language, cycles without captured evaluation, and deferral.Source: §2.2 · p. 5

2.3 Measuring Visual Similarity and Cohort Atypicality at Scale

  • DINO-family embeddings support scalable visual-similarity measurement, but the resulting scores are not calibrated human-perceived similarity and cannot be compared with prior production regimes without a shared-pipeline corpus.The method summarizes cohort patterns with nearest-neighbor similarity, cluster structure, occupancy-based diversity, and the order-2 Vendi score.Source: §2.3 · pp. 5–6
  • These measures describe the selected representation and reference set, not historical novelty, objective originality, or a literal number of designs.The measurement strategy is motivated by adjacent work on product aesthetics, design prototypicality, and generative-text distribution comparisons.Source: §2.3 · p. 6

3 Method

This section describes the study’s methods across a 73-site artifact corpus and survey and instrumented-session subsamples, specifying visual-similarity measures, interpretive baselines, and process coding. It details recorded-session measures and coding, limits analyses to cohort- or sub-person-level claims, protects participants’ identities, and notes that some AI-assisted coding estimates are preliminary.

3.1 Setting and Participants

  • Of 84 graduate students, 83 submitted; after exclusions, 73 homepages remained, spanning 31 business types and 21.6 effective domains.Each student selected a different real business, so convergence could not arise from identical business briefs, though the cohort shared an assignment and course context.Source: §3.1 · p. 6

3.2 Task and Tool

  • Students built promotional websites for real businesses, tested them across devices and browsers, and were graded on creative design that avoided generic AI aesthetics.Lovable was recommended but other AI-assisted tools were permitted; students presented sites to a business stakeholder and prospective customer and made changes based on feedback.Source: §3.2 · p. 7

3.3 Data: One Deployment, Three Nested Layers

  • The nested data layers comprise 73 published-site artifacts, 58 complete survey responses, and nine instrumented first-build sessions.The process sample was screened from 24 recordings; seven builders used platform chat throughout, while two were retained as contrast cases.Source: §3.3 · pp. 7–8

3.4 Measuring Similarity, Originality, and Diversity

  • DINOv3 screenshot embeddings support cosine-similarity measures, originality as 100 minus mean similarity to the other 72 sites, and nearest-neighbor similarity.The analysis also estimates effective diversity over archetype occupancy and cross-checks it with clustering-free and intrinsic-dimensionality measures.Source: §3.4 · pp. 8–9
  • A representative UMAP–HDBSCAN partition is treated descriptively because clustering on 73 sites is configuration-sensitive.The study evaluates 60 configurations and reports robustness checks including geometry-preserving variants, page-length invariance, and cross-model probes.Source: §3.4 · p. 9

3.5 The Interpretive Baseline

  • Because real small-business sites may already reflect homogenizing production stacks, the authors plan a stratified comparison with bespoke, template-builder, and pre-generative sites.The cohort’s absolute similarity is therefore not treated as evidence about the production stack without an external anchor.Source: §3.5 · p. 9
  • Shared exemplars and rubric language, including rubric text pasted into prompts, and peer visibility during building remain possible sources of convergence.The cohort design cannot exclude these shared-context channels.Source: §3.5 · p. 9

3.6 Process Instrumentation and Coding

  • Process coding aligns second-level behavioral states, think-aloud transcripts, and extracted prompts, extending the workflow of Li et al. [2026].A frontier multimodal model performed conservative first-pass coding with JSON-constrained output; the authors state the validation status of derived quantities.Source: §3.6 · p. 9
  • The 54-cycle set spans eight builders; analyses use cycle-, episode-, or utterance-level units, treat nine sessions as replications, and draw no between-person inferences.Contrast cases are flagged where relevant.Source: §3.6 · p. 10
  • Loop conformance and motif mining assess state sequences, while waiting economics divides generation intervals into dead waiting, productive fill, preview peeking, and idle.Motifs are 3–5-state subsequences tested against a per-person first-order Markov null; loop transitions are compared with a within-person shuffled-sequence null.Source: §3.6 · p. 10
  • Translation loss compares verbalized intentions in the 240 seconds before a prompt with typed requirements, so its transmission rate is conditional on verbalized intent.Think-aloud does not capture all intention.Source: §3.6 · p. 10
  • Breakdown–repair tracing logs subsequent repair actions and verdicts; agency is measured through creation-verb constructions and model attribution of who made the artifact.Source: §3.6 · pp. 10–11
  • Behavioral codes are first-pass estimates pending validation and review; discovery counts are omitted because spot checks found that label triggered too often.The pending work includes a stratified 54-utterance validation sample, review of 92 breakdown–repair ladders, and spot checks of translation-loss pairs.Source: §3.6 · p. 11

3.7 Ethical Considerations

  • Analyses are reported at cohort or sub-person level, identifying labels are anonymized, and no named site is linked to originality scores or process behavior.Quoted recording and reflection fragments are short and unnamed.Source: §3.7 · p. 11

4 Results I: Within-Cohort Visual Concentration

Across distinct business briefs, the websites show locally concentrated visual similarity and a small set of recurring design structures. The cohort also includes highly original outliers, while the artifact patterns point to shared influences upstream of the designs.

4.1 Pairwise and Nearest-Neighbor Similarity

  • Similarity is locally concentrated: each site's closest neighbor among the other 72 has a median score of 49 index points, compared with 22.8 for the average pair.A pair of consultancies built by different students scores 86 index points.Source: §4.1 · p. 11

4.2 A Representative Six-Cluster Description

  • Six visual archetypes absorb 64 of 73 sites (88%) in one specification, but 60 configurations yield three to twelve clusters.The six-cluster count is configuration-specific, so the archetypes are descriptive summaries rather than stable natural kinds.Source: §4.2 · p. 11, Figure 5 · p. 14
  • Card-Grid is the largest archetype at 21 sites, while palette varies across layouts rather than defining a separate style.Card-Grid spans unrelated business types; the Full-Bleed Photo style is the most dark-dominant, with 11 of 12 sites dark.Source: §4.2 · pp. 11–12
  • The effective design count is six to twelve, while the 31 business types have a parameter-free inverse-Simpson count of 21.6.The design estimates include a 12.0 order-2 Vendi score and occupancy-based values of 8.0 Shannon-equivalent and 6.0 inverse-Simpson.Source: §4.2 · pp. 12, 15
  • Sector is associated with archetype assignment, but its adjusted mutual information is 0.20 for nine sector groups and 0.26 for 31 fine-grained types.The associations exceed a permutation null at p< .001, yet knowing sector removes only about a fifth of uncertainty about house style.Source: §4.2 · p. 15

4.3 The Originality Spectrum

  • Per-site originality averages 77%, ranging from a site with 4% mean similarity to the field to one with 36%.The most original site shares only 4% mean similarity with the field, while the most generic sits deep inside a house style.Source: §4.3 · p. 15

4.4 From the Artifacts to the Process

  • The recurring archetypes resemble the production stack’s typical styles, suggesting shared influences upstream of the artifacts.Nine instrumented sessions are examined as a candidate mechanism, but the supplied results do not specify their observed behaviors.Source: §4.4 · p. 15

5 Results II: The Mechanism of One Script, Thin Input, and Shallow Acceptance

Together, these results examine how recurring archetypes relate to a one-script generate-and-check workflow, tracing thin or decreasing human input, generic or absent acceptance, and little explicit verification of originality or breakdowns.

5.1 One Script

  • Conformance to the canonical loop was 0.21–0.80 across seven chat-driven builders (z=+2.8 to +6.2); one motif occurred 21 times, 1.89 times more than predicted, in at least six of nine sessions.Sessions shared the loop and recurring motifs, not a common schedule; normalized-timeline agreement was modest.Source: §5.1 · pp. 16–17, Figure 9 · p. 18
  • 61% of generation time was dead waiting; productive fill rose from 9% in builders’ first-half intervals to 17% in the second half.Generation intervals covered 119 minutes, a quarter of recorded time, and builders mostly watched the wait.Source: §5.1 · p. 18, Figure 11 · p. 19

5.2 Thin and Thinning Human Input

  • Across six of nine builders with usable speech windows, 27 of 38 voiced atomic intents were typed (71% transmitted; 29% dropped), while 39% of typed prompt units were never voiced beforehand.The authors treat the percentage as an illustration and draw no population estimate from it.Source: §5.2 · pp. 18, 20
  • Build requests fell from 100% to 11% of prompts as gradient modifications rose from 0% to 56%; later requests increasingly used generic aesthetic adjectives.Reference density rose from 0.79 to 1.23 per 10 words as shared context accumulated, and later prompts increasingly referred to what the model had already made.Source: §5.2 · p. 20

5.3 The Acceptance Test

  • Originality appeared in 15 of 440 evaluative utterances (3.4%); functional and aesthetic criteria were invoked more often than business fit.Verdict language was generic and shared across builders, with 49% of each person’s five most frequent terms appearing in the pooled top ten on average.Source: §5.3 · p. 21
  • 92 verbally flagged breakdowns drew varied first responses: reprompting (23), direct editing (21), researching (18), no action (18), or external AI (12).Source: §5.3 · p. 21
  • 3.4% of 440 spoken verdicts invoked originality, while the acceptance test was generic, fast, often absent, and almost never about originality.Source: Figure 14 · p. 22, §5.3 · p. 23
  • Because the loop is one script summoning the same prior for everyone, convergence takes the form of recurring archetypes; human input is minimal and decreasing.The unspecified space determined by defaults is therefore significant.Source: §5.3 · pp. 22–23
  • 34 of 92 (37%) spoken breakdowns were never followed by a positive verdict; defects can be shipped or silently dropped without verification.The figure reports that 37% were never confirmed fixed.Source: §5.3 · p. 22, Figure 16 · p. 23

6 Results III: The Convergence Is Invisible to Its Creators

Across recordings, surveys, and reflections, builders narrated their own activity while crediting the AI with making; subjective experience did not track measured originality, despite explicit concern about genericity. The authors frame the gap between felt authorship and unseen distributional position as authored ignorance.

6.1 Invisibility in Production: Narrating the Doing, Crediting the Making

  • Builders used first-person creation language more often than AI-subject language (131 vs. 69; human share 0.66), but credited the AI more often with making (130 vs. 66; human share 0.34).Human credit share fell from 0.38 to 0.28 across sessions.Source: §6.1 · p. 23, Figure 17 · p. 24
  • Builders described activity as their own while attributing design decisions to the model, leaving them little occasion to audit decisions they had not made.The authors connect this production-level pattern to the survey’s lack of detectable association between subjective experience and originality.Source: §6.1 · p. 23

6.2 Invisibility at Scale: Felt Authorship, Satisfaction, and Quality Do Not Track Originality

  • Equivalence tests excluded |r|≥0.27 for felt authorship, 0.24 for satisfaction, and 0.29 for perceived quality, but cannot rule out smaller relationships.No smallest effect size of interest was prespecified.Source: §6.2 · p. 24
  • Felt authorship (r=+.06), AI satisfaction (r=+.00), and perceived quality (r=+.13) showed no significant correlations with measured originality.The three subjective measures nevertheless intercorrelated (r=.47–.56, all p<.001).Source: §6.2 · p. 24, Figure 18 · p. 25

6.3 The Intention Was Present: Evidence from the Written Reflections

  • Forty-eight of 78 reflections discussed originality or generic AI aesthetics, and builders described deliberate tactics to avoid genericity.The reflections indicate that avoiding genericity was an explicit concern, not indifference, yet intention and belief in success did not predict measured originality.Source: §6.3 · p. 26

6.4 Authored Ignorance

  • The authors trace a finer-grained layer of the condition to model-made design decisions that humans narrate and credit away during production.Their account links this process pattern to builders’ inability to audit the design decisions they never made.Source: §6.4 · p. 26
  • Authored ignorance describes felt authorship of a design whose genericity its author cannot perceive because judging originality requires the cohort distribution.The platform can reveal that distribution; inspecting one’s own site cannot repair the missing comparison.Source: §6.4 · p. 26

7 Discussion

The discussion examines visual concentration despite diverse briefs and high reported authorship and satisfaction, treating it as a candidate rather than causally established mechanism. It considers platform responses while qualifying measurement and causal limits and proposing controlled probes and comparisons for future tests.

7.1 What We Learned

  • Seventy-three builders using one tool for distinct briefs produced roughly a dozen designs despite high authorship and satisfaction, extending documented AI-assisted homogenization to visual artifacts.The authors describe the convergence, its candidate mechanism, and its invisibility as the three main findings.Source: §7.1 · p. 27
  • The candidate mechanism combines shared defaults, thin specifications, and shallow acceptance tests; originality appeared in 3.4% of verdicts, and a quarter of generations were accepted in silence.The nine-builder process sample documents these links in the same cohort but does not test whether they cause convergence; causal identification awaits experimental manipulation.Source: §7.1 · p. 27

7.2 Theoretical Contribution

  • Authored ignorance names the gap between feeling that one created an artifact and knowing its position in a distribution visible to the platform.The discussion argues that defaults can enter when intentions go untyped and persist when originality goes unevaluated; rapid, brief inspections and silent acceptance favor fluency judgments.Source: §7.2 · p. 28

7.3 Implications for Platform Design

  • Platform responses include deliberate specification and comparison, originality feedback, divergent candidate generation, and tracking effective design diversity.Suggested prompts include visual references, palette choices, and what a design should not look like; an originality meter could compare locally or sectorally relevant sites.Source: §7.3 · pp. 28–29

7.4 Limitations and Future Work

  • The evidence comes from one class, platform, and model generation; survey null results do not establish equivalence, and process claims draw on nine builders without between-person inference.The process codes are AI-assisted first passes awaiting stratified human validation; the authors also note that the study measures visual, not verbal, originality.Source: §7.4 · p. 29
  • Translation-loss rate conditions on verbalized intent and lacks a cross-medium anchor; comparing builders with a human developer or structured intake could calibrate it.Source: §7.4 · p. 29
  • Hero regions are markedly less convergent than whole pages, suggesting shared structure sits below the fold; its source may be model priors, platform scaffolding, or their interaction.Narrow model priors have independent support as a possible mechanism for preference collapse.Source: §7.4 · p. 29
  • The production stack is the study’s explanans, so raw-model prompts and fixed-platform comparisons across model versions could disentangle model priors from platform scaffolding.The resulting design implications would differ depending on which component drives convergence.Source: §7.4 · p. 29
  • The cohort design removes brief and builder heterogeneity but remains tied to a shared genre, assignment scope, and course context, motivating randomized platform comparisons and mechanism manipulations.The comparison corpus anchors the cohort; proposed manipulations include imposed specification or forced comparison at acceptance.Source: §7.4 · pp. 29–30

8 Conclusion

In this cohort, most builders produced similar sites without seeing where their designs stood in the cohort distribution. Robustness analyses characterize the cluster partition as descriptive and report a clustering-free effective-diversity measure.

8 Conclusion

  • In this cohort, vibe coding made building accessible to everyone, but left most builders with similar sites whose similarity was invisible to them.The authors suggest the pattern may follow wherever the production stack is shared, but say whether it holds for the web at large remains an open question.Source: §8 · p. 30
  • The proposed platform response is to show builders where their designs stand in the distribution and prompt specifications before acceptance..Source: §8 · p. 30

A.1 Clustering Sensitivity

  • Across clustering configurations, archetype counts ranged from 3 to 12 (median 7); the production configuration assigned 88% of sites to six archetypes.The authors treat this configuration-sensitive partition descriptively and base quantitative claims on clustering-free statistics.Source: §A.1 · p. 32

A.2 Effective Number of Distinct Designs

  • The order-2 Vendi score was 12.0, with a jackknife 95% confidence interval of [9.1, 15.0] and a subsampling interval of [11.0, 12.8].This statistic uses the cosine-similarity kernel without clustering or projection.Source: §A.2 · p. 32

A.3 Fixed-Depth Crops and Page Length

Fixed-depth crops show that visual similarity depends on how far down the page is included, while headline statistics remain essentially unchanged at depths covering most pages. A hero-only comparison also shifts all three diversity statistics in the same direction relative to full pages.

A.3 Fixed-Depth Crops and Page Length

  • Headline statistics were essentially unchanged at depths covering most pages, while the length association weakened toward zero under strict geometric equalization.The residual association at intermediate depths is consistent with a long-template explanation for default-accepting builders, but geometric sensitivity cannot be separated and the explanation remains a hypothesis.Source: §A.3 · p. 32
  • Deeper fixed-depth crops were steadily more similar, while per-site originality ordering at the shallowest depth was r=.62.The result is reported in Table 5.Source: §A.3 · p. 32, Table 5 · p. 33

A.4 Hero-Region Variant

  • Hero crops yielded 11.2% mean pairwise similarity, 28% median nearest-neighbor similarity, and a 24.2 order-2 Vendi score, versus 22.8%, 49%, and 12.0 for full pages.The comparison rests on the consistent direction across all three statistics, not any single number.Source: §A.4 · p. 33

A.5 Hero Archetypes: Layout × Palette

The hero embeddings yield six recurring layouts, and the same layout skeletons appear across dark and light palettes. Palette variation therefore does not imply layout variation.

A.5 Hero Archetypes: Layout × Palette

  • 84% of hero regions (61 of 73) fall into six recurring layouts under the favored UMAP–HDBSCAN configuration.The diagnostics favor neighbors 25, minimum distance 0.12, minimum cluster size 5, and seed 42; using minimum cluster size 4 produces a coarser five-cluster split.Source: §A.5 · p. 33
  • 48 heroes (66%) are dark and 25 (34%) light, yet layout skeletons recur across both palette skins.The embedding is described as dominated by layout and composition and largely invariant to global palette.Source: §A.5 · p. 33
  • Table 6 organizes the 73 sites by hero-layout archetype and dark or light palette, with palette defined by hero luminance.Dark is below luminance 160 and light is at or above 160 on a 0–255 scale.Source: Table 6 · p. 33, §A.5 · p. 33

A.6 Color Palettes and Their Convergence

The rendered pages share neutral canvases and a small set of vivid accent hues, while palette distinctions cut across layout archetypes. Effective palette diversity is low under both full-palette and accent-only measures.

A.6 Color Palettes and Their Convergence

  • Palettes were extracted from rendered full-page screenshots, so photographic imagery contributes to the measured colors and blends with brand chrome.The extraction uses k-means with k=5; neutral colors are defined by HSV saturation below 0.15, and the brand accent is the highest-area vivid mid-tone.Source: §A.6 · pp. 33–34
  • On average, 63% of screen area is neutral, and 54 of 73 sites have a vivid accent.Among those accents, warm/earth tones number 28 and trust-blue 11; green, teal, red, and yellow number 5, 3, 6, and 1, respectively.Source: §A.6 · p. 34
  • The full-palette order-2 Vendi score is 2.2, while accent-only diversity is 3.0 by inverse-Simpson and 2.9 by Vendi.The accent-only Vendi value uses a circular-hue kernel over 54 accents; color converges more sharply than layout (≈12 distinct designs) or input briefs (21.6 business domains).Source: §A.6 · pp. 34–35
  • Accent hue tracks genre only weakly and cuts across layout archetypes; dark and light skins recur on the same layout skeletons.Food and hospitality skew warm, while technology skews blue; the palette is a skin applied across layouts rather than an archetype of its own.Source: Figure 19 · p. 34, §A.6 · pp. 35–36

A.7 Typography and Its Convergence

Typography converges around a small set of heading and body fonts. A high-contrast serif headline paired with an Inter or DM Sans body is the recurring type-system signature.

A.7 Typography and Its Convergence

  • Computed typography was read from 72 of 73 live sites; one site was unreachable, and some sites specify only generic font stacks.Because all but one site loaded, the still-live subset is effectively the whole cohort.Source: §A.7 · p. 36
  • Playfair Display headings appear on 24 sites (35%) and Cormorant Garamond on 12 (17%), while Inter is the body font on 36 sites (50%).DM Sans is the second most common body font, appearing on 9 sites (12%).Source: §A.7 · p. 36
  • The three most common heading fonts cover 62% of sites, and the three most common body fonts cover 69%.The families are drawn from about twenty distinct heading and body fonts each, with 40 across the cohort.Source: §A.7 · p. 36
  • A high-contrast serif display headline over an Inter or DM Sans body is the recurring “AI-generated” type-system signature.Figure 25 renders the dominant faces set in themselves.Source: §A.7 · p. 36

A.8 Cross-Model Probes

Two non-DINO probes provide limited context for what the embedding measure captures: SigLIP 2 agrees moderately, while ViT-MAE is uninformative for this purpose. The authors also note limits to like-for-like model comparison and plan a human pairwise validation.

A.8 Cross-Model Probes

  • SigLIP 2 per-site originality agrees moderately with DINOv3 (r= .44), while its mean pairwise similarity saturates at 55.9%.The authors interpret the pattern as consistent with DINOv3 capturing visual style beyond the pages’ shared topic.Source: §A.8 · pp. 36–37
  • ViT-MAE shows near-zero agreement with DINOv3 (r= .04) and is uninformative for this purpose.The passage characterizes its unfinetuned embeddings as weak global descriptors.Source: §A.8 · p. 37
  • A human pairwise-similarity validation on a stratified sample of pairs is planned for the archival version.The authors say no open non-DINO model currently provides a comparable style-focused global descriptor for this image geometry, which lies outside the natural-image settings of DINO evaluations [Oquab et al. 2024; Siméoni et al. 2025].Source: §A.8 · pp. 36–37

A.9 Environment

  • The analysis pins dependency versions because exact density-clustering partitions can vary across library versions even at a fixed random seed.The pinned versions are recorded in the archival package.Source: §A.9 · p. 37

B Stakeholder Feedback and Early Adoption

This section examines stakeholder and customer feedback in the real-business website assignment, including testing, iteration, and reflection, and describes students’ reflection blogs and the grading rubric. It also presents visual summaries of cohort palettes and type systems, alongside feedback on operational and usability concerns and the infrequent attention to resemblance among websites.

B Stakeholder Feedback and Early Adoption

  • Business stakeholders emphasized accuracy and content, while customers surfaced usability defects; students reported acting on this concrete feedback and shipping requested changes.Examples included correcting business hours and ordering links, fixing cart or scheduling issues, and revising a page judged unprofessional.Source: B Stakeholder Feedback and Early… · p. 38
  • Some deliverables entered real business use: an owner shared a site as the official website, and other owners planned domain purchases or launches.One consultancy founder requested two additional service pages, which the student generated and shipped in the same session.Source: B Stakeholder Feedback and Early… · p. 38
  • Across 65 accounts, feedback addressed content, trust, and function but never asked sites to look less like other websites.The clearest aesthetic request was for a site to feel “more professional, modern, and inviting,” echoing generic prompt adjectives; the authors argue viewers assess one artifact, not its distribution.Source: B Stakeholder Feedback and Early… · p. 38
  • One student used structured AI personas for owners, executives, and customer types, evaluating a model-generated artifact with simulations run on the same model class.This made the feedback loop circular.Source: B Stakeholder Feedback and Early… · p. 38

Overview

  • The assignment asked students to create a professional website for a legitimate real business, with functional, responsive design aligned to its brand and audience.Students were encouraged to choose businesses that could benefit from a website and for which they could reach a stakeholder and potential customers; evaluation criteria included usability, creativity, originality, and alignment.Source: Overview · p. 38, Part 1: Business Selection · pp. 38–39, Part 2: Website Creation · p. 39

Part 3: Testing & Feedback (Critical)

  • The assignment required documented feedback from at least one business stakeholder and one potential customer, at least one resulting iteration, and cross-platform testing.Testing covered at least two devices and two browsers, including features, images, readability, navigation, and small-screen presentation.Source: Part 3: Testing & Feedback (Critical) · pp. 39–40

Part 4: Recording Your Process (Optional Research Component)

  • Process recording was optional and could be omitted without grade penalty; participants were asked to record the start of a building session while talking through their process.The recordings were intended to illuminate evaluation, testing, iteration, and correction during human-AI collaboration.Source: Part 4: Recording Your Process… · p. 40

Part 5: Reflection Blog Post

  • The reflection assignment required a 500+ word personal blog post about creating a real-business website with AI tools, including feedback received and changes made in response.Students were asked to describe the business, their approach, goals, stakeholder and customer feedback, and iteration.Source: Part 5: Reflection Blog Post · p. 40
  • The reflection blog asked students to describe three important design decisions and why, estimate the AI-generated share of the final site, and explain challenges and solutions.Source: Part 5: Reflection Blog Post · p. 40
  • Students were also asked to assess how well their site served the business’s goals and reflect on learning about AI tools, web and mobile-first design, and stakeholder feedback.Source: Part 5: Reflection Blog Post · p. 41

Grading Rubric (15 Points Total)

  • Table 7 reproduces the assignment’s grading rubric in full.Source: Table 7 · p. 41
  • The palette wall shows a shared neutral base and a narrow warm/blue accent range across all 73 sites.Sites are ordered neutral-dominant first and then by accent hue; each site’s brand accent is marked at the right edge.Source: Figure 24 · p. 42
  • Playfair Display or Cormorant Garamond headings over Inter body text recur across the cohort’s 73 sites.The figure shows the share of sites using each dominant heading/display and body font.Source: Figure 25 · p. 43
Loading 2609.38183v1…