Lie Against AI: Revealing Private Information through AI in an Economic Experiment

Asymmetric information is a key element in different economic transactions and daily life. The disappearance or major reduction of such asymmetries can largely influence human behavior. A technology that can lead to such a shift is an algorithm that detects lies live based on facial expressions, voice, or head pose. In this article, we show how we produced a data set that can be used to train a lie-detection algorithm, developed such an algorithm, and investigated the economic effects of its application. In doing so, we adapt the Belot & van de Ven (2017) experiment and examine lying behavior in the presence of asymmetric information in a buyerseller game. In our design, sellers have monetary incentives to sometimes misreport their private information. We investigate the ability of buyers to detect lies via video conference, use the obtained video communication to develop a large lie-detection data set, and train a lie-detection algorithm such that it can be applied in a laboratory setting. Results indicate that sellers lie and buyers are not good at detecting such lies. Further, we investigate the willingness of buyers to invest in different mechanisms to reveal the private information of sellers using various methods, including the self-developed lie detection algorithm. The results indicate low application rates of these mechanisms. We consider overconfidence and algorithm aversion as possible explanations.