Unraveling trust in collaborative human-machine intelligence from neurophysiological perspective: a review of EEG and fNIRS features

Abstract

With the advancement of artificial intelligence and intelligent technologies, machines have evolved beyond mere mechanical tools and now possess human-like capabilities, including sensing and decision-making. This has led to the emergence of Collaboration in Human-Machine Intelligence (CHMI) as machine intelligence becomes more integrated into collaborative work, with trust serving as the foundational mechanism. By comprehending the concept of trust and its influencing factors, machines have the potential to adapt to human trust states, thereby enhancing the overall CHMI experience. A prerequisite for achieving this is the precise real-time measurement of trust. Neuroimaging technologies, as means to measure neurophysiological responses, have shown the potential to map brain functions and cognitive processes involving trust directly. However, methodological and context variations have led to dispersed neurophysiological evidence related to trust. Therefore, this paper comprehensively summarizes neuroimaging-based trust measurement in CHMI, employing electroencephalography (EEG) and functional Near-Infrared Spectroscopy (fNIRS) as representative neuroimaging tools. A systematic literature review was conducted on Scopus and the Web of Science, and 56 relevant articles were identified. Quantitative and qualitative analysis of these articles revealed a diverse range of neuroimaging tools being used in different CHMI contexts. The neurophysiological-trust relationship has been studied using various methodological perspectives, including time–frequency analysis, event-related potential (ERP), and connectivity analysis. These investigations have identified multiple brain features associated with trust, some of which demonstrate consistency while others show inconsistency. These findings suggest a lack of coherence in trust-related neurophysiological features. Accordingly, three main limitations are identified: a lack of standard methodologies, insufficient focus on connectivity analysis, and a dearth of multimodal and machine learning analysis. Five future directions are proposed to address these limitations and facilitate real-time trust estimation using neuroimaging tools. These directions aim to overcome the identified limitations and pave the way for developing trustworthy CHMI systems. The practical implications of this study lie in informing the design and implementation of such systems, ultimately leading to improved team performance and the establishment of more efficient and reliable collaborative systems.