“Vibe coding” — generating websites and applications by describing what you want in natural language and letting generative AI produce the code — removes the friction of learning HTML, CSS, and JavaScript. The barrier to creating a website drops to zero. But the designs that emerge converge toward a narrow aesthetic: clean layouts, sans-serif fonts, rounded corners, gradient backgrounds, card-based content organization. The AI reproduces the dominant conventions from its training data.
Shin et al. (arXiv:2603.13036) characterize the lifecycle of vibe-coded design and identify where homogenization enters. The frictionless generation that makes the tool accessible is also the mechanism that eliminates diversity. When creating a website requires no design decisions — the AI makes them all — the defaults become the design. And the defaults reflect the statistical center of the training distribution, which is the most common web aesthetic of the training period.
The risk analysis identifies multiple levels of harm. At the individual level, creators lose the skill-building that manual design provides. At the community level, cultural design traditions are displaced by the AI's default. At the ecosystem level, the feedback loop (AI generates designs → users accept them → new web pages resemble the AI's output → future AI training data is more homogeneous → future AI output is more homogeneous) amplifies convergence over time.
The proposed remedy is “productive friction” — deliberate constraints that force the creator to make design decisions rather than accepting defaults. The friction isn't a bug; it's the mechanism that produces diversity. By adding structured prompts, style constraints, and cultural design templates, the tool can be made to diverge from its defaults without losing the accessibility that makes vibe coding valuable.