Forty-seven normal subjects performed two-dimensional arm movements on a digitizer board using a mouse device. The movements were projected on a computer monitor. Subjects were instructed to move the mouse using the whole arm from a center position to a peripheral target so that the projected movement would pass over the target without stopping on the target. A large number of targets (360) were used to cover the entire directional continuum. The direction of the arm movement was the parameter of interest, which was measured at an initial position, at one third of the distance towards the target, and at the vicinity of the target. Four conditions of delay between target presentation and movement execution were used (0, 2, 4, 6 s). A systematic directional error was observed at the initial portion of the trajectory. This error resulted from a clustering of movement directions on an axis that was perpendicular to the axis of the resting forearm before movement onset. This pattern of errors can be explained by the initial inertial anisotropy of the arm. As the trajectory evolved, a different directional error emerged, resulting from a clustering of movement directions in two orthogonal axes. This pattern of directional error increased in amplitude as the delay increased, in contrast to the error at the initial portion of the trajectory which remained invariant with increasing delay. Finally, the information transmitted by the movement direction was shown to increase with the evolution of the trajectory. The increase in delay resulted in a decrease in directional-information transmission. It is proposed that the directional bias towards the end of the movement trajectory might reflect the action of "movement primitives", that is patterns of muscle activation resulting from spinal interneuronal activation. It is further proposed that the directional bias observed at the vicinity of the target might reflect a loss of cortical directional information with increasing delay between target presentation and movement onset.