Soutenance de thèse - Éléonore Houdoyer, "Aligning Minds and Machines : Behavioral, Neural, and Deep Learning Markers of Agency in Human-AI Interaction"

Date : Mercredi 29 Mai 2026 de 13h à 17h

Lieu : Salle des Actes | École normale supérieure | 45, rue d’Ulm 75005 Paris

Titre : "Aligning Minds and Machines : Behavioral, Neural, and Deep Learning Markers of Agency in Human-AI Interaction"

Résumé :
As artificial intelligence (AI) systems increasingly mediate human action and decision-making, understanding their impact on the human sense of agency (SoA), the subjective experience of controlling one’s actions and their consequences, has become a critical scientific and ethical challenge. This thesis investigates how automation and explainable AI shape agency at behavioural, neurocognitive, and computational levels, with a particular focus on human-AI interaction in autonomous driving contexts. The first part examines how varying degrees of automation, explanatory transparency, user engagement, and decision conflict modulate explicit and implicit components of agency. The second part investigates the neurocognitive mechanisms underlying these effects using EEG across three experiments. The final part addresses agency during continuous human-AI interaction, where traditional ERP-based approaches are limited by the overlap of successive events. To capture agency-related neural information without interrupting ongoing behaviour, deep learning methods were applied directly to continuous EEG signals. Convolutional neural networks, based on the EEGNet architecture, were trained to discriminate levels of automation and AI explanation from ongoing neural activity. This approach identifies distributed neural patterns associated with agency and provides a computational framework for decoding agency-related neural dynamics beyond discrete events. Overall, this thesis provides evidence that automation and AI opacity can diminish the sense of agency in human-AI interactions, while explainable AI can partly restore agency by making system intentions more transparent to users. By combining behavioural measures, neurophysiological markers, and computational decoding methods, the thesis offers a coherent framework for characterizing, decoding and monitoring agency in human-AI interaction. It also lays methodological foundations for the development of adaptive and explainable AI systems designed to preserve human agency in everyday use.

Directeurs :
• Valérian Chambon (IJN ENS)
• Solène Le Bars (DEC ENS)