Computational Cognitive Modeling to understand the effects of Racializing AI on Human-AI cooperation with PigChase Task
Swapnika Dulam, Christopher L Dancy
TL;DR
This work investigates how the perceived racialization of AI training data affects human-AI cooperation using the Pig Chase task. It combines a large online experiment with seven race-based treatment conditions and an ACT-R cognitive modeling approach to reveal how sociocultural perspectives modulate decision-making in cooperative scenarios. Quantitative analyses show demographic01 and treatment×demographic interactions, while qualitative coding and ACT-R modeling reveal distinct strategies and cognitive processes across groups, highlighting implicit and explicit biases toward racially White AI cues. The findings underscore the importance of sociocultural context in designing equitable and effective human-AI teams, and point to cognitive-modeling as a tool to understand and mitigate racialized effects in AI collaboration.
Abstract
Despite the continued anthropomorphization of AI systems, the potential impact of racialization during human-AI interaction is understudied. This study explores how human-AI cooperation may be impacted by the belief that data used to train an AI system is racialized, that is, it was trained on data from a specific group of people. During this study, participants completed a human-AI cooperation task using the Pig Chase game. Participants of different self-identified demographics interacted with AI agents whose perceived racial identities were manipulated, allowing us to assess how sociocultural perspectives influence the decision-making of participants in the game. After the game, participants completed a survey questionnaire to explain the strategies they used while playing the game and to understand the perceived intelligence of their AI teammates. Statistical analysis of task behavior data revealed a statistically significant effect of the participant's demographic, as well as the interaction between this self-identified demographic and the treatment condition (i.e., the perceived demographic of the agent). The results indicated that Non-White participants viewed AI agents racialized as White in a positive way compared to AI agents racialized as Black. Both Black and White participants viewed the AI agent in the control treatment in a negative way. A baseline cognitive model of the task using ACT-R cognitive architecture was used to understand a cognitive-level, process-based explanation of the participants' perspectives based on results found from the study. This model helps us better understand the factors affecting the decision-making strategies of the game participants. Results from analysis of these data, as well as cognitive modeling, indicate a need to expand understanding of the ways racialization (whether implicit or explicit) impacts interaction with AI systems.