One reason for humans’ advanced social interaction skills is their capacity to continuously make predictions about the state of mind of others. Predictions (or expectations) about one’s social environment can be formed based on the perception of others’ actions, knowledge, and beliefs. A key region for this complex social ability in humans is the temporoparietal junction, and recent evidence suggests that the evolutionary origin of this area was already present in their close primate ancestors. Like humans, dogs are also sensitive to others' actions or intentions, but the neural mechanisms supporting this complex social ability remain unstudied. To address this research gap, we used a series of functional MRI tasks with awake and unrestrained pet dogs (N = 22-28) and humans (N = 40). First results indicate that observing social interactions recruits a complex occipital-temporal-parietal network in both species. It encompasses face- and body-responsive areas and regions sensitive to dynamic aspects of social cues and action features, including the canine suprasylvian and sylvian and human inferior and lateral temporal cortex. Notably, parietal cortices show significantly stronger engagement in humans, whereas the temporal lobe predominates in supporting social cognition in dogs. In addition, our findings show how the dog's temporal lobe responds to observing social interactions that take unexpected turns (e.g., interruption by another event) and if they elicit a social prediction error, analogous to observations in humans and non-human primates. These findings offer new insights into the emergence of social behaviours across species and their neural bases.