It's interesting research, but it's far from reading thoughts, so everyone can put their tinfoil hats away. The researchers showed random letters to subjects and recorded the fMRI signal that occured after each presentation. They trained a neural network with the data, then fed it results from other random trials. The part of the brain they focused on, primary visual cortex (V1), works well for this because there is a direct mapping between where light hits on the retina and which area of cortex is activated (check out the retinotopic map). Thus, it makes sense that you could easily go from V1 activation to retinal activation to original image.

Other areas of the brain, even other areas of the visual cortex (such as those that process motion, color, and other aspects of an image), are not laid out as nicely, so dreams and other imagined imagery are safe from this technique. The real news seems to be that they were able to get enough of a signal difference via fMRI in order to decode V1 information.