Project Lead: Jing Xu, UW Center for Globally Beneficial AI and UW Anthropology

Data Scientists: June Yang and Curtis Atkisson

LLMs are increasingly used to analyze and explain cultural-historical materials, yet their interpretation is neither neutral nor objective. Rather, it is an encounter between texts and interpretive devices, both with their own positionalities shaped by complex knowledge production and datafication processes. This project will combine human close-reading and LLM-empowered computational analysis to interpret a rare archive of naturalistic fieldnotes on everyday social interactions among children and adults in mid-20th-century rural Taiwan. By treating AI as both an analytical tool and an object of critical inquiry, the project asks two interconnected questions: How do AI-generated moral interpretations compare with culturally situated human morality? And how do the implicit moral values and outlooks of U.S.- and China-based LLMs differ? The project adopts a two-pronged methodology that integrates computational analysis with human interpretation. First, it will systematically analyze thousands of pages of observations and interview texts, using LLM-assisted topic modeling and knowledge graphs to identify broad cultural themes and map relationships among people, practices, and moral values. Second, it will conduct a systematic comparison of moral interpretations and evaluations generated by U.S.- and China-based LLMs. The project’s central innovation lies in developing a bidirectional framework for studying human culture and artificial intelligence, or what we call an anthropological approach to LLM: using LLMs to uncover patterns of moral life in cultural records of everyday life, while using these culturally situated records to critically examine the moral assumptions embedded in LLMs.