High reproducibility of human FMRI studies is imperative for potential clinical applications of this new method for mapping human brain functions. So far, published data are not comparable quantitatively (even at the same field strength) as differences in sequence design and parameters as well as statistical methods applied to enhance function related image contrast, in particular, to extract the size of the "activated areas," are manifold. We present a study on reproducibility of gradient-echo FMRI in the human visual cortex using the different threshold strategies for correlation analysis that shows that, (a) applying adaptive correlation thresholds results in higher reproducibility compared to a fixed (0.5) threshold; (b) highly reproducible data can be obtained on a clinical 1.5 T MRI system, at least for repeated single subject studies (i.e., standard deviation of 2-30% for signal enhancement in 72-94% of the studies and 5-50% for activated area size in 63-88% of the studies, respectively, depending on threshold strategies); however, depending also on subject cooperation; (c) reproducibility across groups (alpha = const.) is worse, i.e., standard deviations are within 33-45% for signal enhancement and 41-74% for activated area size, respectively; (d) SNR is maximum at about 30 degrees flip angle, suggesting significant contributions from T1-effects for larger flip angles. Various technical, methodological, and physiological factors are influencing variability of signal enhancement and apparently activated area size, which should be taken into account if interpreting FMRI data quantitatively.