Asymmetric Algorithm Aversion

Artificial intelligence (AI) is increasingly embedded in decision support systems, yet people do not treat all AI recommendations equally. Drawing on two randomized controlled experiments, a field study using data from a real-world Fintech investment platform, and interviews with industry experts, we identify a phenomenon we call asymmetric algorithm aversion: Although AI advice significantly influenced experts’ evaluations when it recommended against investing, it had no significant effect on experts’ evaluations when it recommended investing. This pattern arises because decision makers perceive AI as effective at analyzing explicit, codified information but less capable of evaluating tacit knowledge, such as leadership quality. As a result, negative AI recommendations are readily accepted, whereas positive recommendations often trigger additional human scrutiny and are frequently discounted. These findings have important implications for organizations deploying AI in finance, hiring, healthcare, and other high-stakes domains where both explicit and tacit knowledge matter. Policymakers and system designers should recognize that AI advice can create systematic biases in human decision making. Effective governance and design strategies should account for users’ differing responses to positive and negative AI recommendations, including the use of explainability features, confidence indicators, and decision processes that promote intentionality in the way in which users incorporate AI advice into their decisions.