Abstract

Generative AI is known for its tendency to homogenize, reproducing dominant style conventions in training data. Less studied is whether and how this extends to complex creative tasks like website design. As more lay creators turn to LLMs to ‘vibe-code’ websites—prompting for aesthetic and functional goals rather than writing code—they risk widespread design uniformity, making it crucial to understand homogenization risks. We first characterize the lifecycle of vibe coding, identifying where homogenization might emerge. We then analyze one high-stakes area for homogenization: website localization. Through large-scale analysis of LLM-generated vs. real-world websites (N=1,800), we show LLMs stereotype when localizing designs across different countries, enabling homogenization at scale. Next, we conduct risk analysis showing how design homogenization compounds known sociotechnical risks, across individual and societal levels. We argue that designers should emphasize friction in vibe coding tools, and outline future directions for preserving creative expression in AI-mediated design.

  @inproceedings{shin2025unpacking,
    title = {Unpacking Design Homogenization in Vibe Coding: A Case Study of Localization in LLM-generated Websites},
    author = {Shin, Donghoon and Gao, Alice and Pang, Rock Yuren and Lee, Jaewook and Reinecke, Katharina and Tseng, Emily},
    booktitle = {},
    year = {2025}
  }