AI “Mind-Reading” Tool Can Recreate What You’re Seeing From a Brain Scan

Meta description: A Weizmann Institute AI tool recreates what people see from fMRI brain scans using far less calibration data than rivals, raising both clinical hope and mental-privacy concerns.

Tags: brain-computer interface, fMRI decoding, neuroethics, AI image reconstruction, Weizmann Institute

A brain decoder developed at the Weizmann Institute of Science can reconstruct images a person has seen by reading their fMRI scans, and it can run in the opposite direction too—predicting brain activity from an image. The work, led by computer scientist Michal Irani, was presented last month at the Cognitive Computational Neuroscience conference in New York. It matters because the tool needs only about one hour of scan data from a new person, compared with roughly 40 hours for prior approaches, and because the same technique edges closer to decoding what someone is imagining rather than merely viewing.

Background: Why fMRI Decoding Has Been Stuck

Functional MRI tracks oxygenated blood flow through the brain using a large magnet. Active regions "light up," but the resolution is coarse: in a typical scanner each highlighted voxel covers about three cubic millimeters and roughly 16,000 neurons. Irani's team worked instead with newer, higher-resolution datasets in which each voxel covered about one cubic millimeter.

Earlier reconstruction attempts produced blurry, hard-to-interpret pictures. Even later models could generate a banana when a person saw a banana, but with the wrong structure and position, Irani says.

How the Two-Branch Decoder Works

The team trained a model on existing data from eight people, each shown around 9,000 images inside a high-resolution fMRI scanner. Their decoder has two branches: one predicts an image's structure, such as where colors sit, and the other predicts its content, such as bananas on a plate. Those predictions guide a diffusion model—AI that builds images by gradually cleaning up noisy pixels—to render a sharper reconstruction.

Data scarcity was the bottleneck. To solve it, the researchers trained a second model in reverse: an encoder that predicts brain activity from an image. Starting with a new picture, the encoder estimates what a person's fMRI scan would look like, and the decoder reconstructs the image from that estimate. Repeated cycles improved both tools, letting the team train on far more images than any volunteer ever viewed. Irani says about 70% of training data came from images never paired with fMRI scans.

Combining multiple studies also revealed brain regions that appear to share functions across individuals—one responding to food images, another to sports.

What the Tool Does Well, and Where It Fails

The universal brain encoder works on a new person with minimal calibration: one hour of fMRI data versus the typical 40. That cost matters, since imaging runs about $600 to $1000 per hour, according to neuroscientist Tommy Sprague of the University of California Santa Barbara.

The models still make mistakes. Irani showed a cake reconstructed as three stacked sandwiches and a dog in a bathtub rendered as a similarly colored goat. In a comparison test, though, the tool outperformed previously described methods by a significant margin. Irani calls "mindreading" a cute, jazzy label for the technique, and she is now extending it from still images to video and audio, with the goal of reconstructing thoughts, mental imagery, and dreams—something she says does not yet exist.

The Privacy Question

Neuroethicist Judy Illes of the University of British Columbia, who was not involved in the research, calls the work magnificent and says using the approach therapeutically for people with neurologic conditions is tremendously exciting. Irani likewise hopes it could help locked-in, completely paralyzed people communicate and shed light on PTSD flashbacks.

Others are more cautious. Sprague notes that 150 years of science fiction could come true if thoughts can be extracted surreptitiously, and he argues ethical considerations deserve more seriousness. He says he would have laughed at such worries a decade ago, when keeping a subject still and engaged in a scanner was hard enough. Irani and other scientists are now pursuing similar decoding from EEG, electrical brain activity captured through electrode caps or even headphones—a shift Sprague and Marcello Ienca of the Technical University of Munich both describe as a potential gamechanger, since a calibrated consumer EEG device could let companies extract extra brain information, possibly without consent. Ienca also imagines courts admitting mental-image reconstructions as evidence.

Irani acknowledges the misuse potential of EEG but says she is not concerned for now: "I'm trying to think only of good things."

What to Watch

The near-term test is whether the encoder-decoder pair holds up on video and audio, and whether one-hour calibration becomes standard in neuroscience labs that currently cannot afford 40-hour scans. The longer-term question is regulatory: EEG-based decoding has no established consent framework, and the tool's accuracy is still improving.