Predictive model to assess user trust: a psycho-physiological approach

Artificial intelligence (AI) systems are becoming pervasive in modern day society. Nonetheless, AI is not infallible. Therefore as more task and controls get delegated to AI systems, the implications become dire and riskier (e.g. Google assistant falling to make an emergency call hands-free), which impacts users experience and make users unforgiving towards system failure. Trust can be a key element as it can help overcome the fear of loss and can supports users interactions. Therefore making it important to design methods and tools to foster trust between users and these technologies, especially capable of assessing users trust objectively and serving as an interface applicable in real-time during the interaction. Measuring the psycho-physiological signals of the user provides a way for objective assessment of users trust towards the technology, with the potential for real-time trust assessment, provided that we know the neural correlates of trust. This study aims to show that it is indeed possible to objectively detect users trust level in AI technologies, and provides details on what physiological signals are most suitable for real-time trust detection, as well as details on the predictive machine learning model used. The results show that our model achieved a mean accuracy of 77.8% and mean receiver operating characteristics (ROC) for the area under the curve (AUC) was 0.76.