Automatiser l'interprétation ? Établir la validité des LLM en sciences sociales

Presentation

The use of large language models (LLMs) is becoming increasingly widespread in the scientific community, particularly in the social sciences, where they enable the automation of text interpretation tasks that were previously carried out by researchers or research assistants. But how can we establish the scientific validity of results produced by probabilistic models, whose outputs may vary and which sometimes produce unjustified inferences? Drawing on a research project dedicated to the semi-automated analysis of protest events, this presentation proposes practical methods for monitoring the quality of data produced using LLMs. It also prompts a broader reflection on the objectives we pursue when using them: are we seeking to eliminate uncertainty in interpretation, or to acknowledge it and account for it transparently?

The speaker

Guillaume Sauvé is a political scientist, researcher and lecturer in political science at the Université du Québec à Montréal, and a lecturer in international studies at the Université de Montréal. Since September 2026, he has been in residence for ten months at the Paris IAS as part of the FIAS (French Institutes for Advanced Study) programme, co-funded by the European Union’s Horizon Europe framework programme – Marie Skłodowska-Curie Action (GA-101217263).

This lecture forms part of the Midis BIN-PHuN series, organised by the Inter-University Centre for Research on Science and Technology (CIRST), a Quebec-based inter-university platform dedicated to digital and computational methods in the social sciences.

Practical details

25 November 2026, 12.30–1.30 pm (Quebec time), i.e. 6.30–7.30 pm (Paris time)
Online, open to the public, via Zoom using the following link: https://uqam.zoom.us/j/84878689005#success