Proceedings of the Annual Meeting of the Cognitive Science Society

Pain is a subjective experience that people sometimes communicate through words. How do observers, from family members to medical professionals, interpret these reports? We propose that people use causal inference to infer someone else's subjective pain experience, treating the pain experience as an unobservable latent variable and the report as one of several sources of observations. We formalized this as Bayesian inference over a causal model of bodily pain and tested it against human judgments on naturalistic stimuli drawn from Reddit posts. Critically, the model is stimulus-computable, taking as input exactly what participants see. The model quantitatively predicted human judgments about pain intensity and appropriate painkiller dosage and generalized across different kinds of pain (e.g., from headaches to finger burns). Furthermore, the model captured individual differences correlated with people's explicit beliefs about how best to assess pain. This work takes a first step toward a scientific understanding of folk intuitive theories of pain.