A new research paper explores the use of large language models (LLMs) as surrogates for human survey respondents, finding their generated tastes to be stylized and biased facsimiles of human preferences. The study utilized models from OpenAI, Anthropic, and DeepSeek to create nearly 277,500 "silicon surrogates" for the Survey of Public Participation in the Arts (SPPA). Key findings indicate that these LLM-generated responses exhibit a positive bias towards liking, lose complex relational structures found in human tastes, and distort known associations between culture, age, class, gender, and race. AI
IMPACT Highlights potential biases and inaccuracies when using LLMs for survey data, cautioning against over-reliance on synthetic respondents.
RANK_REASON Research paper published on arXiv detailing findings about LLM behavior.
- alphaXiv
- Anthropic
- arXiv
- CatalyzeX Code Finder for Papers
- CORE Recommender
- DagsHub
- DeepSeek
- Gotit.pub
- Hugging Face
- Influence Flower
- OpenAI
- ScienceCast
- Survey of Public Participation in the Arts (SPPA)
- WEIRD
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
