He asked 14 volunteers to do one of four tasks: sing songs silently to themselves; recall the events of the day; count backwards in threes; or simply relax.
Participants were given a 10-minute period during which they had to do this. For the rest of that time they were free to think about whatever they liked. The participants' brains were scanned for the entire 10 minutes, and the patterns of connectivity associated with each task were teased out by computer algorithms that compared scans from several volunteers doing the same task.
This differs from previous experiments, in which the subjects were required to perform mental activities at specific times and the scans were then compared with brain activity when they were at rest. Greicius reasons his method encourages "natural" brain activity more like that which occurs in normal thought.
Read my mind
Once the algorithms had established the brain activity necessary for each task, Greicius asked 10 new volunteers to think in turn about each of the four tasks. Without knowing beforehand what each volunteer was thinking, the system successfully identified 85 per cent of the tasks they were engaged in. "Out of 40 scans of the new people, we could identify 34 mental states correctly," he says.
It also correctly concluded that subjects were not engaged in any of the four original activities when it analysed scans of people thinking about moving around their homes.
The findings suggest that patterns for thousands of mental states might serve as a reference bank against which people's thoughts could be compared, potentially revealing what someone is thinking or how they are feeling. "In some dystopian future, you might imagine reference patterns for 10,000 mental states, but that would be a woeful application of this technology," says Greicius.
The idea of the system being used by security services or the justice systemto interrogate prisoners or suspects is far-fetched, Greicius says. Thousands of reference patterns would be needed, he points out, and even these might not be enough to tell if someone is lying, for example.
Instead, he hopes it could be used in Alzheimer's and schizophrenia to help identify faults in the connections needed to perform everyday tasks. He also says the system might be useful for gauging emotional reactions to film clips and adverts.
How much detail such brain scans would show remains to be seen. "There would be a pretty coarse limit on what you could distinguish," says John Duncan of the UK Medical Research Council's Cognitive and Brain Sciences Centre in Cambridge. "The distinctiveness of an activity predicts the distinctiveness of brain activity associated with it," he says.
Kay Brodersen of the Swiss Federal Institute of Technology in Zurich, Switzerland, agrees. "You might be able to tell if someone is singing to themselves," he says. "But try to distinguish a Lady Gaga song from another and you would probably fail."
"The most important potential for this is in the clinic where classifying and diagnosing and treating psychiatric disease could be really important," says Brodersen. "At the moment, psychiatry is often just trial and error."