This perhaps overly excited piece in the Economist got my attention. At it’s root are a series of studies that have come out over the past several years where machine learning has been used to “recognize” the signatures of concepts (like nouns) from many fMRI scans. Tom Mitchell’s work at Carnegie Mellon comes to mind.
While these are indeed exciting studies, the notion that we’re somehow at the threshold of ubiquitous “mind reading” and deception detection strikes me as far fetched. As an example, the concept of “banana” can surely be found in either an individual’s ground truth or a lie. While we might be able to pick out the brain activity signature of banana in both, we’d really have a great deal of difficulty figuring out which context was the lie.