AI Can Now Test Brain Treatments on a Computer First

In psychiatry, the goal is often not just to read what a brain is doing, but to change it — to soften a negative emotional bias or an overly vivid bad memory. Testing those changes on real people is slow, expensive and sometimes risky. So a research team built an AI "neural surrogate": a computer model that mimics one person's brain responses, based on brain scans taken while they looked at pictures.

The raw material was huge: more than 36,000 image-and-brain-scan pairs from four deeply studied volunteers. The AI learned to turn brain activity back into images, then "steered" those patterns toward two targets — feeling more positive, and being more memorable. Finally, the team showed the AI-generated images to 18 real people across 7,200 rating trials to see whether the predicted emotional effect actually showed up.

The results were encouraging but early. For emotional tone, ratings moved in the direction the AI predicted in 16 of the 18 volunteers — a real but small shift, roughly a few percent. Memorability barely budged, and the AI's own scoring agreed only weakly with human raters. So the method works better as a hypothesis-testing tool than as a finished treatment.

There's a clear catch. Push the AI too hard and the altered images drift away from the original, meaning you trade realism for effect. The model is also a frozen snapshot, not a living brain, and it can't yet capture how real stimulation unfolds over time. Still, the approach offers something psychiatry rarely gets: a way to design and pre-test candidate brain treatments cheaply and safely before anyone signs up for a clinical trial.