Automation Disrupts, Explanations Restore: The Neural Signatures of Agency Loss and Recovery in Human–AI Interaction

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

DOI: 10.64898/2026.07.22.740020

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 investigates the neurocognitive mechanisms underlying the sense of agency (SoA) in human–AI interaction, specifically examining how automation disrupts agency and whether explainable AI (XAI) can restore it. The research addresses the problem that automation often weakens users' subjective experience of control by disrupting the predictive link between intention and outcome. While XAI is proposed as a solution, the specific neural processes through which explanations restore agency remain unclear. The authors hypothesized that providing AI systems with hierarchical intention sharing—disclosing both distal (goal-level) and proximal (trajectory-level) intentions—would enhance predictive processing and restore the SoA. The researchers conducted three EEG experiments using a simulated autonomous driving paradigm with 57 participants (after exclusions). In Experiment 1, participants compared self-controlled driving against AI-controlled driving to establish the neural signature of agency loss. Experiment 2 tested whether distal explanations (revealing the AI’s strategic goal) partially restored agency. Experiment 3 examined the effect of combining distal explanations with proximal explanations (revealing the specific trajectory). Across all experiments, participants monitored an AI or controlled a car to reach targets, receiving auditory feedback. The study measured explicit agency ratings and early auditory event-related potentials (ERPs), specifically the P1–N1 and N1–P2 complexes, which index sensory attenuation and predictive processing, as well as the mismatch negativity (MMN) for pre-attentive deviance detection. The results demonstrated that automation significantly reduced explicit feelings of control and disrupted sensory attenuation, evidenced by increased P1–N1 amplitudes, decreased N1–P2 amplitudes, and delayed N1 latencies compared to manual control. Distal explanations partially restored agency, selectively modulating early auditory responses by decreasing P1–N1 and increasing N1–P2 amplitudes. The strongest restoration of both behavioral and neural markers of agency occurred when distal and proximal explanations were combined, yielding graded attenuation of P1–N1, enhanced N1–P2 responses, and accelerated N1 latencies. Notably, the MMN remained unaffected across all conditions, indicating that pre-attentive deviance detection is preserved regardless of agency levels or explainability. These findings identify component-specific EEG markers that track fluctuations in the sense of agency, demonstrating that multi-level intention sharing by AI systems enhances predictive engagement and the explicit experience of control. The study provides a neurocognitive foundation for designing explainable autonomous systems, suggesting that transparency regarding both high-level goals and low-level actions is critical for maintaining user agency and trust in human–AI interactions.

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
• automation surprise