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

Synthesis: Boussioux, Zhao, and Cho (2026) ask what happens to design diversity when a shared generative, Large Language Models (LLMs)-driven tool produces the whole artifact. Seventy-three graduate students each built a website for a different real business on the Vibe Coding platform Lovable, under a rubric that explicitly rewarded "creative and original design that avoids generic AI aesthetics." Embedding every homepage with DINOv3 and comparing 2,628 pairs, they find 73 briefs collapsing into roughly a dozen designs: six archetypes absorb 88% of sites, and a median builder's closest look-alike shares 49 index points against a 22.8-point average. A nine-session think-aloud subsample traces the mechanism: one shared generate-and-check script, 29% of voiced intentions never typed, originality in only 3.4% of 440 verdicts. Most strikingly, felt authorship, satisfaction, and perceived quality show no detectable association with measured originality — builders converged while believing they had not. The authors call this authored ignorance and contribute a joint outcome–process–perception strategy (Mixed-Methods Research).

Key Findings

1. Across 2,628 DINOv3 comparisons of 73 homepages built for distinct businesses, mean visual similarity was 22.8 index points, yet the median nearest neighbour scored 49.
2. A representative UMAP–HDBSCAN partition assigned 88% of sites to six archetypes — Card-Grid, Whitespace, Full-Bleed Photo, Menu, Product Catalog, Editorial — while 60 configurations returned three to twelve clusters.
3. The convergence was not explained by the briefs: the 31 business types spanned an effective diversity of 21.6, roughly twice the 12.0 effective number of distinct designs.
4. Nine recorded think-aloud sessions showed one shared script: every chat-driven builder exceeded a shuffled null on generate-and-check conformance, with prompting peaking early and inspection dominating.
5. Human specification was thin and thinning: 29% of voiced intentions were dropped before typing, build requests fell from 100% of opening prompts to 11%, and 61% of generation time was dead waiting.
6. The acceptance test almost never asked the graded question: originality appeared in 3.4% of 440 verdicts, 26% of cycles got no verdict, and 37% of 92 flagged defects were never confirmed fixed.
7. Felt authorship (67/100), satisfaction (6.4/7), and perceived quality (6.1/7) were internally coherent but uncorrelated with measured originality (r ≤ +0.13, n = 57–58).

What the cohort's artifacts showed

Seventy-three graduate students each chose a different real business and built its promotional site on Lovable as graded work shown to real Stakeholders, an instance of Authentic Assessment. Embedding every full-page screenshot with a self-supervised vision transformer (Multimodal AI DINOv3) and comparing all 2,628 pairs showed local concentration: the average pair shared 22.8 index points, but the median nearest-neighbour score was 49, and the most similar pair reached 86. Six recovered house styles absorbed 88% of sites, while clustering-free indices put the effective number of distinct designs near twelve against 21.6 for the input briefs. The originality spectrum ran from a 64% look-alike to a 96% original, a distributional result motivating the process study.

The mechanism inside the loop

Nine builders recorded their first session with concurrent think-aloud, coded at one-second resolution, a rare process-level view of AI-assisted work. Every chat-driven builder followed the same generate-and-check loop more often than a shuffled null. Human input was thin and thinning: across eight analysable cycles, 38 voiced intentions produced 27 typed ones, so 29% were lost between mind and keyboard. Later prompts collapsed into gradient tweaks and bug reports, increasingly pointing at what the stack had already made. Waiting dominated — 61% of generation time was dead waiting — and the acceptance test was shallow: verdicts were fast and generic ("good", "cool", "cute"). This is the Student-AI Interaction the platform's defaults fill.

Why the convergence stayed invisible

The process data show the decoupling produced in real time: builders narrated their sessions in the first person while attributing the making to the model. The post-task survey confirmed it at cohort scale: felt authorship, satisfaction, and perceived quality hung together, yet none tracked measured originality. Reflections rule out indifference: 48 of 78 posts discussed avoiding generic AI aesthetics, so the missing correlation reflects a missing comparison set, not a lack of intent. The authors name this authored ignorance: originality is relational, and only the platform sees the distribution against which it would be judged. Feedback for 65 sites asked for accurate hours, working links, and usability fixes; nobody asked a site to look less like others. Research participation was separated from grading under IRB approval; only cohort-level statistics are reported (Ethics).

What this means for practice

• Instructors. Grade and teach against the distribution, not the artifact. A rubric saying "avoid generic AI aesthetics" cannot work when no builder sees the cohort; show comparison sets and require a "what should this not look like" specification.
• Faculty developers. Treat Vibe Coding as a delegation skill, not a tool demo. The sessions show the failure is a thin, fast acceptance test, so coach deliberate specification and evaluation over faster prompting — an Educational Development priority.
• Educational technology developers. The diagnosis implies the remedy: surface an originality meter from the platform's own outputs, sample initial candidates from distinct archetypes, and add pre-generation friction funded by the 61% of generation time spent waiting.
• Researchers. Reuse the outcome–process–perception design — vision-transformer embeddings, one-second instrumentation, post-task self-report — as a replicable strategy for studying Human AI Collaboration.

Limitations

• One cohort, one platform. The data cover 84 graduate students in one course, a single recommended platform (Lovable), and one model generation; the archetypes will drift.
• The process layer is nine builders. Process claims sit at the cycle, episode, or utterance level with sessions as replications; the utterance- and prompt-level codes are AI-assisted first passes whose human validation is pending.
• The perception null is bounded, not established. The survey joins 58 of 73 sites (57 for authorship), authorship is a single item, and equivalence holds only for margins of |r| ≥ 0.24–0.29.
• Similarity is not human-calibrated. Embedding scores are not a validated percentage of perceived similarity, and with no pre-generative corpus the absolute level cannot yet be attributed to vibe coding; the think-aloud rate conditions on verbalized intent.

Citation

Boussioux, L., Zhao, Z., & Cho, K. (2026). One Tool, One Taste? How Vibe Coding Trades Collective Diversity for Individual Creativity. arXiv preprint.