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

Houdoyer, Eléonore; Le Bars, Solène; Chambon, Valérian · 2026 · Crossref

DOI: 10.64898/2026.07.22.740014

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This study addresses the challenge of monitoring the sense of agency (SoA) during continuous human–AI interaction, specifically in autonomous driving contexts. Existing methods for measuring agency, such as subjective reports or event-related potentials, are ill-suited for naturalistic, ongoing interactions because they rely on discrete events or disrupt task flow. The authors investigate whether agency-related neural states can be decoded from continuous electroencephalographic (EEG) activity using deep learning, aiming to enable non-intrusive, real-time monitoring of user control and engagement. The research involved two experiments with 37 total participants (20 in Experiment 1, 17 in Experiment 2) performing a simulated navigation task. Experiment 1 manipulated decision authority, comparing human-controlled trials against AI-controlled trials. Experiment 2 focused exclusively on AI-controlled trials, manipulating system explainability by providing either no explanation or intention-based distal explanations regarding the AI’s strategy selection. EEG data were recorded during the pre-feedback interval. The authors employed EEGNet, a convolutional neural network architecture, to decode experimental conditions from single-trial EEG data. They assessed performance using intra-subject and leave-one-subject-out cross-validation frameworks to evaluate both individual reliability and cross-subject generalization. Behavioral results confirmed that both control delegation and the presence of explanations significantly modulated participants’ perceived control. Neural decoding models successfully distinguished between agency-modulating conditions. Intra-subject decoding performance was robust, and inter-subject decoding remained significantly above chance, demonstrating partial generalization of agency-related neural representations across users. Spectral ablation analyses revealed that low-frequency activity, particularly in the delta and theta bands, contributed most strongly to decoding accuracy. Complementary time–frequency analyses indicated that these bands exhibited increased power under reduced-agency conditions during the post-keypress, pre-feedback interval. These findings indicate that agency-related information is embedded in continuous, low-frequency neural dynamics associated with predictive monitoring processes. By demonstrating that such states can be decoded from single-trial EEG, the study provides a foundation for developing neuroadaptive systems. These systems could dynamically regulate automation levels and explainability to preserve human agency, offering a scalable solution for maintaining user engagement and trust in safety-critical human–AI interactions without relying on disruptive subjective measures.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-nvidia on 2026-08-10; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• automation

Information type

What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).

• Empirical Findings: physiological data
• Methodological Resource: tool software
• Theoretical Contribution: computational model