Global signal regression (GSR) is a controversial preprocessing method in resting-state fMRI. It has been claimed that the process can introduce artifactual anti-correlations in resting-state connectivity maps. However, a consensus regarding its use has been lacking, due in part to the difficulty in understanding its effects. We show that GSR can be well approximated by a temporal downweighting of the voxel time series, where the weighting factor is a function of the global signal magnitude and is uniform across space. This helps address the concerns about GSR and provides a novel framework for understanding its effects on resting-state data.