Head motion remains a major obstacle in fMRI. We have used realistic human digital motion phantoms with empirically-derived head movements and known BOLD signals to address two unresolved questions: 1) how effective are motion correction algorithms? and 2) how much motion is too much when assessing scan quality? Our analysis evaluated different motion metrics and motion correction methods using both block-designed and event-related fMRI task data. The results show that head motion metrics need to distinguish between positional offsets versus active movement, that combining image realignment plus motion-censoring is most effective, and that residual motion after corrections determines acceptability thresholds.