We have evaluated the ability of a time-delayed artificial neural network (TDANN) to predict shoulder and elbow motions using only electromyographic (EMG) signals recorded from six shoulder and elbow muscles as inputs, both in able-bodied subjects and in subjects with tetraplegia arising from C5 spinal cord injury. For able-bodied subjects, all four joint angles (elbow flexion-extension and shoulder horizontal flexion-extension, elevation-depression, and internal-external rotation) were predicted with average root-mean-square (rms) errors of less than 20 degrees during movements of widely different complexities performed at different speeds and with different hand loads. The corresponding angular velocities and angular accelerations were predicted with even lower relative errors. For individuals with C5 tetraplegia, the absolute rms errors of the joint angles, velocities, and accelerations were actually smaller than for able-bodied subjects, but the relative errors were similar when the smaller movement ranges of the C5 subjects were taken into account. These results indicate that the EMG signals from shoulder and elbow muscles contain a significant amount of information about arm moVement kinematics that could be exploited to develop advanced control systems for augmenting or restoring shoulder and elbow movements to individuals with tetraplegia using functional neuromuscular stimulation of paralyzed muscles.