Drivers and influence of social conformity on decision making in human-AI teams

Abstract Advances in AI are transforming decision making teams. A fundamental open question is whether AI agents elicit social conformity effects comparable to human peers. If so, a critical challenge is determining whether the established drivers of social conformity in human teams—informational influence (the need to gain information from others) and normative influence (the need to gain approval or avoid punishment from others)—also transfer to AI teammates. In two studies, we investigated informational and normative influence in mixed human–AI teams using a medical information cascade paradigm. Participants received private evidence about a medical case and public advice from one or more AI or human advisors. Study 1 explored how participants weighted this public advice compared to their private information in scenarios where private and public information were equally accurate. Study 2 varied the accuracy of the private and public information. Results showed a distinct contrast between the two mechanisms. Humans and AI exerted a similar level of informational influence in both Study 1 and Study 2, and participants systematically under-weighted advice from both human and AI advisors relative to a Bayesian computational benchmark when source accuracies varied in Study 2. However, normative influence differed by advisor identity. In Study 1, participants placed significantly more weight on human advice than on equally reliable AI advice, indicating that AI lacks the normative pull inherent to human peers. Crucially, Study 2 revealed that this preference is fragile. We found that participants did not rely on a single heuristic, but instead attempted to integrate majority consensus with variable advisor accuracy. This complex processing eliminated the normative bias toward humans, and participants systematically under-weighted advice from both humans and AIs relative to the Bayesian computational reference. In addition, we also found an egocentric bias in both studies: participants in both studies weighed their private information significantly more than public advice. Theoretically, these findings suggest that differential reliance on AI versus human advice stems from the absence of normative influence rather than a lack of informational trust. Practically, the results highlight a design risk: presenting heterogeneous accuracy cues may increase cognitive load, leading to suboptimal advice integration.