the 14th International Conference on Social Computing, Behavioral-Cultural Modeling, & Prediction and Behavior Representation in Modeling and Simulation
Recent literature has shown that racism and implicit racial biases can affect one’s actions in major ways, from the time it takes police to decide whether they shoot an armed suspect, to a decision on whether to trust a stranger. Given that race is a social/power construct, artifacts can also be racialized, and these racialized agents have also been found to be treated differently based on their perceived race. We explored whether people’s decision to cooperate with an AI agent during a task (a modified version of the Stag hunt task) is affected by the knowledge that the AI agent was trained on a population of a particular race (Black, White, or a non-racialized control condition). These data show that White participants performed the best when the agent was racialized as White and not racialized at all, while Black participants achieved the highest score when the agent was racialized as Black. Qualitative data indicated that White participants were less likely to report that they believed that the AI agent was attempting to cooperate during the task and were more likely to report that they doubted the intelligence of the AI agent. This work suggests that racialization of AI agents, even if superficial and not explicitly related to the behavior of that agent, may result in different cooperation behavior with that agent, showing potentially insidious and pervasive effects of racism on the way people interact with AI agents.
Atkins, A. A., Brown, M. S., & Dancy, C. L. (2021). Examining the Effects of Race on Human-AI Cooperation. In proceedings of the 14th International Conference on Social Computing, Behavioral-Cultural Modeling, & Prediction and Behavior Representation in Modeling and Simulation, Virtual, 279-288.