Can large language models replace human surveys in geospatial studies? A case of vague cognitive regions

Recent advances in artificial intelligence have raised interest in using large language models (LLMs) as cost-efficient proxies for human participants. This paper explores their use in studies requiring place-specific knowledge. Drawing on Daniel Montello’s work on vague cognitive regions, we use Central Ohio as a case study comparing commercial and open-source LLM responses with human surveys. LLMs generally approximate human collective judgments of regional membership, though all models tend to overextend geographic boundaries. Newer commercial models show less response variability compared to humans. Persona prompts have little observable effect. Human involvement remains essential when integrating AI into spatial cognition research.