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

Research into "vibe-coding" systems, which convert natural language instructions into deployed websites, indicates a significant degree of design homogenization despite efforts to encourage originality. A study of 73 promotional websites created by graduate students using a shared tool revealed that 88% clustered into six interpretable design categories, suggesting that while individual creativity is fostered, collective design diversity may be constrained by the underlying system.

Why it matters

The findings highlight a potential inherent bias or limitation in generative AI tools for design, where underlying algorithms may lead to predictable patterns despite users' intentions for originality. This has implications for market differentiation and brand identity, particularly in competitive digital environments where unique visual communication is crucial for standing out.

Key insights

• Vibe-coding systems can generate deployed websites from natural language instructions.
• A study of 73 websites created with a specific vibe-coding tool (Lovable) for distinct businesses revealed a notable lack of collective design diversity.
• Using DINOv3 embeddings and cosine similarity analysis, 88% of the websites clustered into six primary design categories.
• Sensitivity analyses indicated a range of three to twelve clusters, consistently pointing to design patterns.
• The findings suggest that these systems might trade collective diversity for individual creativity within a shared production context.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2609.38183