This study evaluates the ability of large-language models to approximate human cultural tastes by generating silicon surrogates from the Survey of Public Participation in the Arts. Using models from OpenAI, Anthropic, and DeepSeek, the authors analyze 277,470 synthetic respondents to determine if LLMs can faithfully replicate real-world survey data.
- Silicon samples exhibit a systematic positive bias for liking, leading to inflated ecological estimates of tastes that are not explained by WEIRD-bias.
- The complex relationality found in real taste structures is completely lost within the silicon samples.
- Known cultural alignments between tastes and social space are poorly preserved, with age-taste associations attenuated and class-taste associations resurrected anachronistically.
- Gender- and race-taste associations are caricatured rather than accurately represented in the synthetic data.
The findings indicate that LLM-generated survey panels produce highly stylized facsimiles of human tastes, raising concerns about their validity for market research and social science applications.