This preprint examined how autonomous systems affect human experience of control and which neural signatures track loss and recovery of agency. Using an autonomous-driving paradigm across three EEG experiments, the authors tested the behavioural and early electrophysiological consequences of automation and of different forms of AI explanations. The study focused on explicit agency judgements and early auditory evoked components implicated in predictive processing.
In the first experiment, introduction of automation reduced participants' explicit feelings of control, demonstrating an operational loss of the sense of agency (SoA) when actions and outcomes are governed by an autonomous system. Concomitant EEG changes reflected reduced sensory attenuation: the study reported increased P1-N1 amplitudes and decreased N1-P2 amplitudes under automation. These component-specific shifts were taken as markers that automation disrupts the predictive link between intention and effect, altering early sensory processing associated with agency.
Experiment 2 assessed whether providing distal, goal-level explanations from the AI could restore agency. Distal explanations are high-level, goal-oriented information about why the system acted as it did. Behaviourally, these explanations partially restored explicit agency ratings compared with the automated condition without explanations. Neurally, distal explanations selectively modulated early auditory responses: they decreased P1-N1 amplitudes and increased N1-P2 amplitudes relative to the automated baseline. This pattern indicates partial recovery of predictive engagement when the AI shares goal-level intentions.
The third experiment combined distal (goal-level) and proximal (trajectory-level) explanations to test whether multi-level intention sharing yields stronger restoration. Behavioural effects were strongest in this combined-explanation condition: explicit control experience improved more than with distal explanations alone. Correspondingly, EEG markers showed the most pronounced neural restoration: a graded attenuation of P1-N1 amplitudes and enhanced N1-P2 amplitudes, consistent with re-engagement of predictive sensory processing. The combination of explanation types produced the most robust recovery across both subjective and neural measures.
Across all three experiments, mismatch negativity (MMN) measures remained unaffected by automation or by the presence of explanations. The authors interpret this as evidence that pre-attentive deviance detection—an early automatic response to auditory change—remains intact regardless of changes in agency or explainability. In other words, while early sensory components tied to predictive sensory attenuation (P1-N1, N1-P2) were sensitive to agency modulation, MMN was not.
The study identifies component-specific EEG markers that track fluctuations in the sense of agency and demonstrates that explainability can restore both subjective control and neural signatures of predictive processing. Specifically, distal explanations partially restore agency and modulate early auditory responses, while combining distal and proximal explanations yields the strongest behavioural and neural recovery. The preserved MMN suggests that some automatic sensory processes remain stable even when agency is disrupted.
These findings provide a neurocognitive basis for designing explainable autonomous systems: sharing intentions at multiple levels (goal and trajectory) appears to re-engage users' predictive mechanisms and improve explicit control experience. The source frames these results as evidence that multi-level intention sharing by AI systems can maintain or restore user agency, with measurable EEG markers useful for assessing design choices.
Limitations and reporting details
The source preprint reports the experiments and their main outcomes but does not include further methodological details or statistical values in this summary. Specifics on sample sizes, exact task parameters, or statistical effect sizes were not reported in the extracted source text. The authors declared no competing interests in the preprint.