This is actually a somewhat different issue. The salmon "result" is essentially a multiple comparisons problem. In a typical fMRI analysis, the researchers run

This is actually a somewhat different issue. The salmon "result" is essentially a multiple comparisons problem. In a typical fMRI analysis, the researchers run a huge number of significance tests, one per voxel (i.e., 3d pixel). Even if false positives are relatively rare on a per-test basis, you're going to see a non-trivial amount of them when you run a large number of tests. This phenomenon was well-known in the statistical literature...and in the fMRI literature, but sometimes it takes a cute little parable, like the salmon, to drive something home. In this case, however, there was one honest-to-God bug (in AFNI). The other software modeled something in a way that sounds fairly reasonable, but didn't quite line up with the way that nature (or, actually, the scanner manufacturers) behaved. Unlike the fish issue, I think you could legitimately try to do things correctly and still get bit by this sort of bug.