An Experiment on the Impact of Information on the Trust in Artificial Intelligence

Artificial intelligence (AI) has made considerable progress in a variety of fields and is suggested to do as well or better than many experts, creating great expectations about its potential to improve decision-making. While much progress has been made in refining the accuracy of algorithms, much remains to determine on how these algorithms will influence decision-makers, especially in life or death decisions such as in medicine. In such fields, human experts will remain for the foreseeable future the ultimate decision-makers. Literature suggests that reliance on algorithms by decision-makers may be influenced by the accuracy of algorithm and by the information on how the algorithm reached its conclusions. The objective of this paper is to determine the propensity to influence pathologists' decision-making using algorithmic expertise and information on AI algorithm accuracy and model interpretability. To test our hypotheses, we will conduct an online, quasi-experimental survey study with 120 respondent pathologists. Each participant will provided with a series of prostate cancer samples and asked to assess the Gleason grade. Our hypothesis is that increasing the level of information will lead to increased reliance in automated systems. This research will provide insight into trust in AI: first, the extent to which pathologists trust AI advice; second, the extent to which each type of information contributes to trust.