That's because the reconstructions are based not on the actual image, but on how the image is perceived by a subject's brain. If an autistic person sees a face differently, the difference will show up in the brain scan reconstruction.
Images from dreams are also detectable.
"And you can even imagine," Cowen said, "way down the road, a witness to a crime might want to come in and reconstruct a suspect's face."
How soon could that happen?
"It really depends on advances in brain imaging technology, more so than the mathematical analysis. It could be 10, 20 years away."
One challenge is that different brains show different activity for the same image. The blurry images pictured here are actually averages of the thoughts of six lab volunteers. If one were to look at any individual's reading, the image would be less consistent.
"There's a wide variation in how people's brains work under a scanner -- some people have better brains for fMRI -- and so if you were to pick a participant at random it might be that their reconstructions are really good, or it might be that their reconstructions are really poor, which is why we averaged across all the participants," Cowen said.
For now, he added, you shouldn't worry about others snooping on your memories or forcibly extracting information.
"This sort of technology can only read active parts of the brain. So you couldn't read passive memories -- you would have to get the person to imagine the memory to read it," Cowen said.
"It's a matter of time, and eventually -- maybe 200 years from now -- we'll have some way of reading inactive parts of the brain. But that's a much harder problem, as it involves measuring very fine details of brain structure that we don't even really understand."