Adaptative action selection in threatening social contexts, for example when facing aggressive individuals, is core to social behavior. It is debated whether, when facing threat, initial action opportunities are mostly determined by automatic action tendencies – i.e., reactive stimulus-response (SR) associations – or by rapid and implicit goal-directed (GD) processes, which depend on the predicted consequences of available actions. To investigate the balance between SR and GD, we conducted three experiments manipulating the predictability of action-outcomes associations, in a novel approach-avoidance decision task in virtual reality. Participants presented a greater avoidance rate from angry individuals when the action-outcome association was predictable. Computationally, this was subtended by more efficient evidence accumulation, as indicated by an increased drift rate. Furthermore, cardiac deceleration around the time of choice was greater in the predictable condition and allowed the value attributed to the outcome to be better integrated into the decision. Finally, while most participants avoided angry individuals only under predictability, supporting a predominant role of GD processes, a minority of them avoided regardless of predictability. Overall, our results shed light on the computational architecture of social avoidance, opening interesting new research and clinical perspectives.