AbstractDyadic interactions require dynamic correspondence between one’s own movements and those of the other agent. This mapping is largely viewed as imitative, with the behavioural hallmark being a reaction-time cost for mismatched actions. Yet the complex motor patterns humans enact together extend beyond direct-matching, varying adaptively between imitation, complementary movements, and counter-imitation. Optimal behaviour requires an agent to predict not only what is likely to be observed but also how that observed action will relate to their own motor planning. In 28 healthy adults, we examined imitation and counter-imitation in a task that varied the likelihood of stimulus–response congruence from highly predictable, to moderately predictable, to unpredictable. To gain mechanistic insights into the statistical learning of stimulus–response compatibility, we compared two computational models of behaviour: (1) a classic fixed learning-rate model (Rescorla–Wagner reinforcement [RW]) and (2) a hierarchical model of perceptual-behavioural processes in which the learning rate adapts to the inferred environmental volatility (hierarchical Gaussian filter [HGF]). Though more complex and hence penalized by model selection, the HGF provided a more likely model of the participants’ behaviour. Matching motor responses were only primed (faster) in the most experimentally volatile context. This bias was reversed so that mismatched actions were primed when beliefs about volatility were lower. Inferential statistics indicated that matching responses were only primed in unpredictable contexts when stimuli–response congruence was at 50:50 chance. Outside of these unpredictable blocks the classic stimulus–response compatibility effect was reversed: Incongruent responses were faster than congruent ones. We show that hierarchical Bayesian learning of environmental statistics may underlie response priming during dyadic interactions.