Detecting Trust Calibration Traits in AI with EEG Signals for Speech Deception

Abstract

This paper investigates the effectiveness of frequency-based electroencephalogram (EEG) measures in capturing self-reported trust and trust mis-calibration in artificial intelligence (AI) systems used as decision support tools. It examines a collaborative human-AI decision-making task where 50 human users interacted with an explainable speech-based AI system to detect deceptive speech. Correlation analysis indicates that Alpha and Theta power values of EEG measured from central, frontal, and parietal regions depict negative correlation with self-reported trust, and Delta power values from frontal regions are negatively correlated with self-reported trust. A machine learning model using EEG power measures to estimate self-reported trust depicted a Spearman’s correlation value ρ = 0.63 (p <0.01) between predicted and actual self-reported trust. Additionally, the Beta-band-power values from left frontal and left central areas are higher during trust calibration compared to under-trust. The Gamma-band-power levels are also higher during trust-calibration compared to over-trust. Machine learning models based on these EEG measures predict different trust dimensions (i.e., over-trust, under-trust, trust-calibration) with moderate macro-F1 scores (i.e., 53-55%). Finally, the most effective trust assessment models leverage data from all considered brain regions – frontal, central, and parietal – achieving similar performance compared to models using central regions only and outperforming models using frontal or parietal regions alone.