*New* CSSS Seminar: What Do Language Models Know About the Past? Recovering Historical Perspectives Using Large-Scale Language Models (10/28/26)
Posted: 2026-09-25 11:18:15 (Local Events)
Historical archives preserve some of the written record, yet social scientists are interested in more than just facts. We are also interested in taken for granted cultural assumptions, such as how much prestige was accorded to certain occupations in different historical periods. In this talk I ask whether large language models, trained on vast historical text, can recover these lost “collective mentalities,” and whether we can trust that recovery as faithful. Using occupational distribution and prestige as a case, I report on experiments probing LLM token probabilities in two stages. First, I validate whether base model probabilties recover historical U.S. occupational distributions against census data. Second, I probe whether these probabilities recover diffuse occupational prestige hierarchies. Asking models to generate a ranking directly fails and degenerates, but probing their probability distributions over a fixed occupation list yields rankings that match known historical classifications. I also find some evidence that we can predictably shift this ranking with the perspective a prompt implies (gendered, regional, temporal). I offer this as a preliminary pipeline for asking whether LLM outputs can serve as evidence of social phenomena otherwise inaccessible to historical and comparative research. Meeting details can be found at this link.
Laura K. Nelson is Associate Professor of Sociology and Director of the Centre for Computational and Digital Social Science, Humanities, and Creative Arts (CODI Centre) at the University of British Columbia. Her research uses computational text analysis to study social movements, gender, and culture, with a focus on developing rigorous, interpretive methods for working with large-scale textual data.
Date: 10/28/2026
Time: 12:30-1:30 PM
Deadline: 10/28/2026
Location: Savery 409