Anticipatory responses can minimize the disturbances that result from the action of one part of the body on another. Such a predictive response is evident in the anticipatory increase in grip force seen when one hand pulls on an object held in the other hand, thereby preventing the object from slipping. It is postulated that such a response depends on predicting the consequences of the descending motor command, as signaled by efference copy, using an internal model of both one's own body and the object. Here we investigate how the internal model learns the temporal consequences of the motor command. We employed two robots to simulate a virtual object held in one hand and acted on by the other. Delays were introduced between the action of one hand on the object and the effects of this action on the other hand. An initial reactive grip force response to the delayed load decayed with the development of appropriate anticipatory grip force modulation. However, no predictive modulation was seen when the object's movement was not generated by the subject, even when the motion was cued by a tone. These results suggest that, when an internal model learns new temporal relationships between actions and their consequences, this learning involves generating a novel response rather than adapting the original predictive response.