1. Reflex activity in human arm muscles has been measured in response to torque perturbations around the elbow joint in the flexion-extension direction and/or the supination-pronation direction. Intramuscular fine-wire electrodes were used to record electromyographic (EMG) activity in the muscles. A pre-load was applied in the same direction as or in a different direction to the perturbation. The subjects were instructed to 'hold on', which means that they had to actively resist the perturbation without unduly co-activating their muscles. 2. The EMG activity showed segmented reflex responses with short-latency (25-50 ms) and long-latency (50-75 ms) components, followed by other bursts of activity which probably originated from the subject's voluntary or triggered reactions. 3. Motor units in m. triceps gave short-latency and long-latency responses to imposed elbow extension but only long-latency responses were seen to imposed pronation, which does not stretch m. triceps. Motor units in m. brachialis gave short-latency and long-latency excitatory (inhibitory) responses to imposed extension (flexion) torques. However, only long-latency inhibitory responses were observed to imposed pronation torques. Motor units in m. biceps gave short-latency and long-latency excitatory responses to imposed pronation torques. 4. These results show that the long-latency reflex cannot be the result of a simple feed-back mechanism that controls muscle length only. We argue that the long-latency reflex activity reflects the co-ordinated activation of muscles which is necessary for an adequate response. This hypothesis is supported by the fact that the distribution of long-latency reflex activity over muscles was found to be similar to the distribution of activation found during a voluntary exerted torque in the direction opposite to the applied perturbation.