Individual differences in autism-like traits are associated with reduced goal emulation in a computational model of observational learning

The ability to infer the goals and intentions of others is crucial for social interactions, and such social capabilities are broadly distributed across individuals. Autism-like traits (that is, traits associated with autism spectrum disorder (ASD)) have been associated with reduced social inference, yet the underlying computational principles and social cognitive processes are not well characterized. Here we tackle this problem by investigating inference during social learning through computational modeling in two large cross-sectional samples of adult participants from the general population (N₁ =&thinsp;943, N₂ =&thinsp;352). Autism-like traits were extracted and isolated from other associated symptom dimensions through a factor analysis of the Social Responsiveness Scale. Participants completed an observational learning task to quantify the tradeoff between two social learning strategies: imitation (repeating the observed partner’s most recent action) and emulation (inferring the observed partner’s goal). Autism-like traits were associated with reduced observational learning specifically through reduced emulation (but not imitation), revealing difficulties in social goal inference (Pearson’s r =&thinsp;−0.124, P < 0.001). This association held, even when controlling for other model parameters (for example, decision noise, heuristics, F_1,925 =&thinsp;15.352, P < 0.001), and was specifically related to social difficulties in autism-like traits (F_1,916 =&thinsp;33.169, P < 0.001) but not social anxiety traits (F_1,916 =&thinsp;0.005, P =&thinsp;0.945). The findings, replicated in an additional sample, provide a powerfully specific mechanistic hypothesis for social learning challenges in ASD, employing a computational psychiatry approach that could be applied to other disorders.