Neuroscientific studies suggest that the dorsal premotor area is activated by biological motions, and is also related to the prediction errors of observed and self-induced motions. We hypothesize that biological and nonbiological motions can be discriminated by such prediction errors. We therefore propose a model to verify this hypothesis. A neural network model is constructed that learns to predict the velocity of the self's next body movement from that of the present one and produces a smooth movement. Consequently, a property of the input sequence is represented. The trained network evaluates observed motions based on the prediction errors. If these errors are small, the movements share a representation with the self-motor property, and therefore, are regarded as biological ones. To verify our hypothesis, we examined how the network represents the biological motions. The results show that predictive learning, supported by a recurrent structure, helps to obtain the representation that discriminates between biological and nonbiological motions. Moreover, this recurrent neural network can discriminate the ankle and wrist trajectories of a walking human as biological motion, regardless of the subject's sex, or emotional state.