The ability to predict the actions of other agents is vital for joint action tasks. Recent theory suggests that action prediction relies on an emulator system that permits observers to use a model of their own movement kinematics to predict the actions of other agents. If this is the case, then people should be more accurate at generating predictions about actions that are similar to their own. We tested this hypothesis in two experiments in which participants were required to predict the occurrence and timing of particular critical points in an observed action. In Experiment 1, we employed a self/other prediction paradigm in which prediction accuracy for recordings of self-generated movements was compared with prediction accuracy for recordings of other-generated movements. As expected, prediction was more accurate for recordings of self-generated actions because in this case the movement kinematics of the observer and observed stimuli are maximally similar. In Experiment 1, people were able to produce actions at their own tempo and, therefore, the results might be explained in terms of self-similarity in action production tempo rather than in terms of movement kinematics. To control for this possibility in Experiment 2, we compared prediction accuracy for stimuli that were matched in tempo but differed only in terms of kinematics. The results showed that participants were more accurate when predicting actions with a human kinematic profile than tempo-matched stimuli that moved with non-human kinematics. Finally, in Experiment 3, we confirmed that the results of Experiment 2 cannot be explained by human-like stimuli containing a slowing down phase before the critical points. Taken together, these findings provide further support for the role of motor emulation in action prediction, and they suggest that the action prediction mechanism produces output that is available rapidly and available to drive action control suggesting that it can plausibly support joint action coordination.