Projet Pref-AIdvisor

Experimental Elicitation of LLM preferences in AI-human advice

About the project

Large Language Models are attracting considerable interest from the general public, and their democratisation is leading a growing number of users to employ them as advisory tools in areas for which they were not specifically designed (e.g. seeking financial investment, health or educational advice from ChatGPT). This project aims to identify the underlying 'preferences' and 'beliefs' of popular LLMs to understand the types of advice they provide. It also aims to assess how sensitive these models are to framing effects based on contextual elements in prompts. The methodology involves replicating classic experimental economics protocols and submitting them to LLMs to identify their preferences, beliefs and potential biases.

Principal investigator
Guilhem Lecouteux (Université Côte d’Azur, CNRS, GREDEG, France)
Duration

March 2025 – March 2027

Total amount
35k€