The neural basis of the automatic activation of words was investigated in an fMRI study. In the study, words were presented briefly (51 or 151 msec) followed by a mask. To prevent attentional processing, subjects attended to the masks and not the words, and were required to make perceptual judgment about the masks. We found that a distributed neural network (including the frontal, temporal, occipital, parietal lobes, and the cerebellum) was activated during non-attentional processing of words in both exposure durations. A significant main effect of presentation duration was found in bilateral cerebellum and the right fusiform gyrus, suggesting their role in the later (151 msec) processing of words. In addition, a significant interaction between presentation duration and word frequency was obtained. When the presentation duration was 151 msec, no significant difference in activation was found between high- and low-frequency words. Alternatively, when the presentation duration was 51 msec, high-frequency words evoked significantly greater activation in bilateral fusiform gyri, cerebellum, right inferior parietal lobe, medial frontal gyrus (BA 45/46/9), and the right temporal-occipital junction (BA 21/37). These results suggest that these regions are sensitive to word frequency, and are related to both the attentional and non-attentional access of lexical representations. Hum. Brain Mapping 18:215–221, 2003. © 2003 Wiley-Liss, Inc.