Abstract Several theories suggest that actions are coded for imitation in terms of mentalistic goals, or inferences about the actor’s intentions, and that these goals solve the correspondence problem by allowing sensory input to be translated into matching motor output. We tested this intention reading hypothesis against general process accounts of imitation using the pen-and-cups task. The task has three components: participants place a pen in one of two cups, using their right or left hand, and one of two grips. Previous research has revealed a colour minimum error pattern; when one of the components is differentially coloured (e.g., one cup is red and the other blue), accuracy is greatest on the coloured dimension. We found the colour minimum error pattern, not only in the standard version of the task, where participants imitate the actions of a human model, but also in three novel variants of the task, in which participants responded on the basis of spatial or arbitrary stimulus–response mappings to ‘geometric’, non-biological stimuli. These stimuli do not afford the attribution of intentions, and therefore our results support generalist theories of imitation by showing that the colour minimum error pattern is due, not to intention reading, but to the operation of task-general processes of perception, attention and motor control.