Decoding Agency-Related Neural States During Human-AI Interaction in Autonomous Driving Using EEG and Deep Learning

Abstract

The sense of agency, the experience of controlling ones actions and their consequences, is a fundamental component of human interaction with autonomous systems. As artificial intelligence increasingly mediates decision-making in domains such as autonomous driving, understanding and monitoring agency-related processes becomes critical for maintaining user engagement, trust, and appropriate levels of control. However, existing approaches to measuring agency rely primarily on subjective reports or event-based neural markers, which are poorly suited for naturalistic and continuous human-AI interaction. In this study, we investigate whether agency-related neural states can be decoded from ongoing electroencephalographic (EEG) activity using deep learning. Across two experiments, we manipulated agency through (i) decision authority (human vs AI control) and (ii) system explainability (AI with vs without intention-based explanations). Behavioral results confirmed that both manipulations significantly modulated participants perceived control. At the neural level, we trained EEGNet-based models to decode agency-modulating experimental conditions from pre-feedback EEG activity. Decoding performance was robust at the intra-subject level and remained significantly above chance across participants using a leave-one-subject-out framework, demonstrating partial cross-subject generalization of agency-related neural representations. Spectral ablation analyses revealed that low-frequency activity, particularly in the delta and theta bands, made the dominant contribution to decoding performance. Complementary time-frequency analyses showed that these bands exhibited increased power under reduced-agency conditions, specifically during the post-keypress, pre-feedback interval. Together, these findings indicate that agency-related information is embedded in continuous, low-frequency neural dynamics associated with predictive monitoring processes. By demonstrating that such information can be decoded from single-trial EEG in an offline setting, this work provides a foundation for future real-time, non-intrusive monitoring of user states in human-AI interaction. These results open new avenues for the development of neuroadaptive systems capable of dynamically regulating automation and explainability to preserve human agency.