Sociotechnical systems (STSs) lie at the heart of realizing AI applications for social good. Underlying STSs is the key assumption that users will act prosocially, i.e., to the benefit of others, and thus help achieve societal objectives. How can we persuade users to act prosocially? But prosociality is inherently a matter of users’ values and associated cognitive states. Whereas current approaches for user modeling focus on observable actions, we must additionally consider values and cognitive states to properly model decision-making.
Accordingly, we propose Caper, a cognitive user modeling framework that infers users’ cognitive states from observable behavior via inverse reinforcement learning. Caper learns the cognitive-behavioral mappings unique to each user, thus surpassing traditional models, which emphasize observed actions over users’ thought processes.
We evaluate Caper in two ways. Simulation with 60 ACT-R agents demonstrates that Caper successfully distinguishes between (1) identical actions driven by distinct cognitive processes, and (2) similar cognitive states producing different behavior under contextual constraints. An empirical human study shows that incorporating the hierarchical structure of human decision-making processes statistically significantly improves the interpretability of human behavior. Thus, Caper provides the essential computational foundation for adaptive, context-aware interventions in prosocial sociotechnical systems.