What artificial intelligence can't do

This third session of the seminar organised by GRETS, "What artificial intelligence can't do", will be attended by Étienne Ollion - École Polytechnique, CNRS. The session will be introduced by Mathieu Brigidou - EDF R&D - GRETS / PACTE.

Does AI enable us to increase our knowledge of society? What to do (and what not to do) in the field of research with Large Language Models (LLMs), artificial intelligence models designed to understand and generate text in natural language?

Among artificial intelligence technologies, giga-language models hold the greatest promise. They can be used to carry out polls based on synthetic data, in silico experiments (using computer simulations) and surveys of past eras. In part, the promises made by the promoters of these technologies are probably untenable, despite the intense publicity surrounding their use. To what extent are they plausible, and indeed already realized?

This presentation looks back at the discourse surrounding artificial intelligence in research. Based on a series of experiments, it seeks to assess the plausibility of the proposed solutions. It shows their interests, limits and possible ways forward. After an overview of the uses of LLMs in the social sciences, the aim of this presentation is to show how artificial intelligence can be used to equip knowledge of society.

• Étienne Ollion is a CNRS research director in computational sociology. His work focuses on, but also mobilizes, artificial intelligence for research.

References

• Julien Boelaert, Samuel Coavoux, Etienne Ollion, Ivaylo Petev, Patrick Präg, "Machine Bias. How do Generative Language Models Answer Opinion Polls?", working paper
• Salomé Do, Etienne Ollion, Rubing Shen, "The Augmented Social Scientist. Using Sequential Transfer Learning to Annotate Millions of Text", in Sociological Methods and Research, 2022
• Francois Levin et Etienne Ollion (dir.), Ce qui échappe à l'intelligence artificielle, éditions Hermann.