We investigated the biophysical mechanism of low-frequency drift in blood-oxygen-level-dependent (BOLD) functional magnetic resonance imaging (fMRI) (0.00-0.01 Hz), by exploring its spatial distribution, dependence on imaging parameters, and relationship with task-induced brain activation. Cardiac and respiratory signals were concurrently recorded during MRI scanning and subsequently removed from MRI data. It was found that the spatial distribution of low-frequency drifts in human brain followed a tissue-specific pattern, with greater drift magnitude in the gray matter than in white matter. In gray matter, the dependence of drift magnitudes on TE was similar to that of task-induced BOLD signal changes, i.e., the absolute drift magnitude reached the maximum when TE approached T(2)* whereas relative drift magnitude increased linearly with TE. By systematically varying the flip angle, it was found that drift magnitudes possessed a positive dependence on image intensity. In phantom experiments, the observed drift was not only much smaller than that of human brain, but also showed different dependence on TE and flip angle. In fMRI studies with visual stimulation, a strong positive correlation between drift effects at baseline and task-induced BOLD signal changes was observed both across subjects and across activated pixels within individual participants. We further demonstrated that intrinsic, physiological drift effects are a major component of the spontaneous fluctuations of BOLD fMRI signal within the frequency range of 0.0-0.1 Hz. Our study supports brain physiology, as opposed to scanner instabilities or cardiac/respiratory pulsations, as the main source of low-frequency drifts in BOLD fMRI.