There are well-known problems with fMRI studies and particulary with how they are interpreted, see e.g. http://www.edvul.com/pdf/VulHarrisWinkielmanPashler-PPS-2009... (PDF) for a good critique. It's an intrinsically low-resolution tool as well, "there seems to be a lot of blood flow in vaguely this area of the brain", that is sold as if it were high tech and the only/best way to learn about human cognition.

I think it's important to point out that papers like that usually are pointing out statistical pitfalls (and often provide solutions), but they don't really cast doubt on fMRI as a tool for measuring brain function.

That's true and I don't think anybody disputes that fMRI is a good way to measure brain function, but (1) the improved statistical methods you talk about are often not adopted and (2) brain function is at a different level of abstraction than cognition and emotion, and making inferences about the latter based on the former is really hard, and would still be hard even with better statistics.

Agreed, I think those factors detract attention from some of the better designed experiments out there.

I became a total cynic toward the entire field of fMRI studies the day I found out that they found a dead salmon [0] could tell what emotions humans were depicting in photos.

[0] http://prefrontal.org/files/posters/Bennett-Salmon-2009.pdf

This is a prime example of what I described. The problem is not with fMRI as a tool, but with what statistical modeling assumptions are made when making inferences. In this case, it's about statistical thresholds and how to avoid false positives--something that is a concern for any experiment, not just one that using an fMRI. To make this clear, here's the discussion:

"Can we conclude from this data that the salmon is engaging in the perspective-taking task? Certainly not. What we can determine is that random noise in the EPI timeseries may yield spurious results if multiple comparisons are not controlled for. Adaptive methods for controlling the FDR and FWER are excellent options and are widely available in all major fMRI analysis packages. We argue that relying on standard statistical thresholds (p < 0.001) and low minimum cluster sizes (k > 8) is an ineffective control for multiple comparisons. We further argue that the vast majority of fMRI studies should be utilizing multiple comparisons correction as standard practice in the computation of their statistics."

The authors don't cast doubt on the entire field of fMRI, just in what conclusions we can draw from certain statistic tests of high dimensional data in the presence of noise...