Cluster-extent based thresholding is the most popular method for correcting multiple comparisons in fMRI data, due to its high sensitivity to weak and diffuse signals. However, this method provides low spatial specificity, as researchers can only infer that there is signal somewhere within a significant cluster, without making specific inferences about the location of the signal. This issue is particularly problematic when a liberal primary threshold (e.g., *p* < .01) is used, which often results in large clusters spanning multiple anatomical regions. A survey of 814 fMRI studies published in 2010 and 2011 revealed that the use of liberal primary thresholds is common, and that the default options in software packages often contribute to this practice. The authors illustrate the problems with liberal primary thresholds using an fMRI dataset and present simulations demonstrating the detrimental effects on false positives, localization, and interpretation of fMRI findings. To avoid these issues, they recommend setting a primary *p* < .001 as a default lower limit, using more stringent primary thresholds or voxel-wise correction methods for highly powered studies, and adopting reporting practices that make the level of spatial precision transparent. They also suggest alternative and supplementary analysis methods, such as threshold-free cluster enhancement (TFCE) and hierarchical false discovery rate (FDR) control on clusters.Cluster-extent based thresholding is the most popular method for correcting multiple comparisons in fMRI data, due to its high sensitivity to weak and diffuse signals. However, this method provides low spatial specificity, as researchers can only infer that there is signal somewhere within a significant cluster, without making specific inferences about the location of the signal. This issue is particularly problematic when a liberal primary threshold (e.g., *p* < .01) is used, which often results in large clusters spanning multiple anatomical regions. A survey of 814 fMRI studies published in 2010 and 2011 revealed that the use of liberal primary thresholds is common, and that the default options in software packages often contribute to this practice. The authors illustrate the problems with liberal primary thresholds using an fMRI dataset and present simulations demonstrating the detrimental effects on false positives, localization, and interpretation of fMRI findings. To avoid these issues, they recommend setting a primary *p* < .001 as a default lower limit, using more stringent primary thresholds or voxel-wise correction methods for highly powered studies, and adopting reporting practices that make the level of spatial precision transparent. They also suggest alternative and supplementary analysis methods, such as threshold-free cluster enhancement (TFCE) and hierarchical false discovery rate (FDR) control on clusters.