Asymmetric Algorithm Aversion

Abstract

We examine human decision-making in the presence of Artificial Intelligence (AI) advice in the context of Fintech. Through four studies (two randomized controlled experiments, a field study, and a qualitative interview study), we show that human acceptance of AI advice exhibits asymmetric algorithm aversion that depends on the advice the AI provides. In our Fintech setting, investments are commonly assessed using criteria based on both explicit knowledge and tacit knowledge. Human experts believed that AI had the ability to effectively use explicit knowledge but were less confident in its ability to effectively use tacit knowledge (a common perception regarding AI). As a result, AI advice significantly influenced experts’ evaluations when it recommended against investing, as they believed that this advice could have been correctly generated using explicit knowledge criteria alone. However, AI advice had no significant effect on experts’ evaluations when it recommended investing, as they believed that this advice could not have been correctly generated without using both types of knowledge and that AI could not effectively use tacit knowledge. Instead, AI advice to invest served as a trigger for them to examine the investment with a focus on tacit knowledge criteria, which often meant that the human expert’s evaluation did not match AI’s advice. We did not find this same asymmetric pattern for accepting human advice in this context, indicating that the results are not due to loss aversion. Recognizing asymmetric algorithm aversion—that human experts are more likely to accept AI’s advice when it produces one recommendation and less likely when it produces a different recommendation —is important because an inconsistent attitude towards accepting AI advice may lead to systematic biases in human-AI collaboration that may defeat the very purpose of using AI.