Peter Watts on no-kidding mind-reading technology:
Now here comes Kendrick Kay and his buddies in Nature with "Identifying natural images from human brain activity", and if they haven't actually vindicated all those cheesy narrative gimmicks, they've made a damn good first pass at it. They used fMRI scans to infer which one of 120 possible novel images a subject was looking at. "Novel" is important: the system trained up front on a set of nearly 2,000 images to localize the receptive fields, but none of those were used in the actual mind-reading test. So we're not talking about simply recognizing a simple replay of a previously-recorded pattern here. Also, the images were natural -- landscapes and still-lifes and snuff porn, none of this simplified star/circle/wavey-lines bullshit.
The system looked into the minds of its subjects, and figured out what they were looking at with accuracies ranging from 32% to 92%. While the lower end of that range may not look especially impressive, remember that random chance would yield an accuracy of 0.8%. These guys are on to something.