The Influence of Risk on Interaction Effectiveness and Human Trust in AI-Advised Decision-Making: Experimental Evidence

This paper explores the impact of increased risk and the possibility of loss on the agent's decision to trust the artificial intelligence. It also further evaluates the effectiveness of such collaboration with respect to the chosen incentive. For this purpose, I conducted an experiment in which participants were asked to solve logical problems by themselves or by reusing the answers suggested by the AI. Both groups received additional points for each correct answer. One of them also had an incentive in the form of a penalty for each wrong response. Based on the results obtained, I can conclude that, for a given accuracy of the selected NLP model, agents do not tend to trust the AI more in situations of both higher and lower risk. Meanwhile, this incentive in a situation of limited time leads to a decline in the overall performance: a decrease in the number of attempts and the quality of answers. This kind of effect can be caused by the insufficiently high accuracy of the model, the presence of stress, and many other factors.