Does. Not. Compute.

If you live in a city, you're familiar with the verbal ramblings of schizophrenics. You may have heard their speeches directed to no one while they're walking down the street toward you or waiting for the train. I used to occasionally share my afternoon commute with a man who liked to stand near the back of the bus and deliver a continuous and incoherent Shakespearean-style monologue, complete with accent and extravagant hand gestures.

One distinguishing feature of schizophrenic speech is disorganization. A sentence might start out normally enough but veer into nonsense. Phrases don't follow one another. Schizophrenia is also characterized by delusions--patients may believe that they're being persecuted, or that they have special powers. The causes of schizophrenia are still mysterious. But researchers at the University of Texas, Austin, have used computers to model several theories of the disease. And one of these models produced computers that talk like a schizophrenic person.

The "computational patients" studied were different iterations of a computer model called DISCERN, designed by professor Risto Miikkulainen. The model is a "neural network" that's meant to simulate how a human learns and recalls a story. Different parts of the network mimic the tasks our brain performs to understand, store, and remember words and sentences.

Miikkulainen and his graduate student Uli Grasemann fed very 28 simple stories into their computer model. Half the stories were in the first person; for example, "I was a doctor. I worked in New York. I liked my job." (I told you they were simple.) The other half were crime stories told in the third person; for example, "Tony was a gangster. Tony worked in Chicago." (Hey now...) It took thousands of repetitions to teach the computer network the stories.

They also taught three slightly less simple stories to actual human subjects, both healthy and schizophrenic. A week later, they asked the subjects to recall those three stories, and recorded the types of errors they made.

To create mentally ill computer patients, the researchers introduced a variety of errors into the DISCERN network. They modeled eight different cognitive problems that have been suggested as factors in schizophrenia. After teaching their impaired computer patients the set of simple stories, it was quiz time. The network was prompted with the first part of a story and asked to complete it. The researchers analyzed the types of errors the computer patients made, then compared them to the human patients.

Out of the eight schizophrenia models, just one had caused the computer to tell stories that sounded like the schizophrenic patients'. The computer got derailed, starting one story and drifting into another story, as did the schizophrenics. The computer had a similar tendency to mix up the characters in the stories, including confusion between first-person and third-person stories. A story about Tony the Chicago mob boss, for example, might become a story about Mary the mob boss.

The underlying error in this pseudo-schizophrenic computer simulation was what the authors called "hyperlearning." Ordinarily, scientists believe, we use a technique called prediction error to help us learn new information. We constantly make predictions, and whenever those predictions don't come true--that is, when reality doesn't match up with what we expected--we take notice and form new associations. This process is thought to involve the neurotransmitter (brain signaling chemical) dopamine. If our brains aren't prudent about this process, though, they assign too much importance to every new piece of information, and we create irrelevant associations. This is why the process is thought to be involved in schizophrenia; every detail may become meaningful to schizophrenics, and connections may appear everywhere in their speech. When the researchers adjusted the computer network's settings so that it relied too heavily on prediction error--it learned too much and ignored too little--the computer's stories sounded the most like the schizophrenic patients'.

In addition to matching the disorganization of the schizophrenic patients' stories, the computer network also provided a tantalizing hint of delusion. Swapping first-person narratives for third-person ones, it put itself at the center of scenarios it didn't belong in: "I was a mob boss."

Computers are, of course, not people. A human brain, no matter how ill, is unspeakably more complicated than a computer model. And while this computer simulation successfully modeled some speech-related symptoms of schizophrenia, it didn't address other symptoms. Still, it gave intriguing hints about prediction error and dopamine that can be used for further study in humans. That information might even lead to new treatments for the people riding your bus.