We used high-resolution fMRI to investigate claims that learning to read results in greater left occipito-temporal (OT) activation for written words relative to pictures of objects. In the first experiment, 9/16 subjects performing a one-back task showed activation in > or =1 left OT voxel for words relative to pictures (P < 0.05 uncorrected). In a second experiment, another 9/15 subjects performing a semantic decision task activated > or =1 left OT voxel for words relative to pictures. However, at this low statistical threshold false positives need to be excluded. The semantic decision paradigm was therefore repeated, within subject, in two different scanners (1.5 and 3 T). Both scanners consistently localised left OT activation for words relative to fixation and pictures relative to words, but there were no consistent effects for words relative to pictures. Finally, in a third experiment, we minimised the voxel size (1.5 x 1.5 x 1.5 mm(3)) and demonstrated a striking concordance between the voxels activated for words and pictures, irrespective of task (naming vs. one-back) or script (English vs. Hebrew). In summary, although we detected differential activation for words relative to pictures, these effects: (i) do not withstand statistical rigour; (ii) do not replicate within or between subjects; and (iii) are observed in voxels that also respond to pictures of objects. Our findings have implications for the role of left OT activation during reading. More generally, they show that studies using low statistical thresholds in single subject analyses should correct the statistical threshold for the number of comparisons made or replicate effects within subject.