The duration of the prompt emission of long gamma-ray bursts is generally considered to be fairly similar to the duration of the activity of the engine in the center of the progenitor star. Here, we investigate the relation between the duration of the engine activity and that of the observed light curve, using inputs from both numerical simulations and observations. We find that the observed burst duration is a good proxy for the engine duration after the time necessary for the jet to break out the star's surface is subtracted. However, the observed duration is a function of the viewing angle and can be significantly shorter than the duration of the engine activity. We also show that the observed, redshift-corrected burst duration evolves only moderately with redshift for both observations and synthetic light curves. We conclude that the broad distribution of the observed duration of long BATSE gamma-ray bursts is mostly accounted for by an engine lasting ~20 s, the dispersion being due to viewing and redshift effects. Our results do not rule out the existence of engines with very long duration. However, we find that they are constrained to be a small minority of the BATSE detected bursts.