This week, the NPR Morning edition featured a three-part series on lie detection, which included a story about Daniel Langleben, a neuropsychiatrist at the University of Pennsylvannia who uses functional magnetic resonance to try and determine how brain activity differs when one is lying or telling the truth.
I attended a presentation by Langleben at the Wellcome Trust Centre for Neuroimaging a couple of weeks ago. What emerged from the talk was a somewhat fuzzy image of diffuse networks of cortical areas, and the small changes in network activity that correlate with deception, with little reference to what the components of the networks are actually doing.
More interesting for me was Langleben's revelation that the CIA, who funded his more recent research, hoped that his findings would lead not to methods of lie detection, but to methods that secret agents could use to avoid the detection of their lies. However, the funding was withdrawn when the initial findings proved inconclusive.
There is much to be learnt before fMRIs can be used for lie detection. Nevertheless, neuroimaging data, like data from traditional polygraph tests, is admissable as evidence in U.S. courts of law, despite there being little evidence that fMRI is more accurate than the old technology when it comes to detecting lies.
I've not kept up with the imaging studies on lying, but isn't there a massive overfitting problem where limited subjects with huge amounts of data lead to discriminative markers that prove to have minimal predictive value?
It sounds like you know more about it than I do.