Researchers at the Weizmann Institute of Science in Rehovot, Israel, have built an artificial intelligence (AI) tool that analyzes a person's brain scan and reconstructs what they are looking at with remarkable accuracy. Called Brain-IT, the system recovers the seen image from high-resolution functional magnetic resonance imaging (fMRI) scans — and in reverse can predict a person's brain activity from an image. The findings were presented at the Cognitive Computational Neuroscience conference in New York.
Two-Branch "Brain Decoder"
At the heart of the system is a two-branch "brain decoder". The first branch predicts the image's structure — where colors are located and how shapes are distributed; the second branch identifies the image's content — what it depicts. Both predictions are fed into a diffusion model. Diffusion models are a type of AI known for progressively denoising a collection of noisy pixels to produce a sharp image. In Brain-IT, this model combines the predictions of the two branches to generate a far more accurate copy of what the person saw.
In practical terms, this means the system doesn't just find the object in the image — it tries to recover its position, size and colors. While earlier tools arrived at the general conclusion that "the person saw something", Brain-IT aims to reconstruct the structure of the scene that was viewed.
To train the system, the scientists used brain scans of eight people, each of whom viewed nearly 9,000 images in high-resolution fMRI scanners. But the volume of existing scans wasn't enough — so the team built a paired model that performs the reverse task: an "encoder" that predicts brain activity from an image. The encoder and decoder were trained together: the encoder creates an estimated brain scan for a new image, the decoder turns it back into an image — and the errors improve both models. As a result, about 70 percent of the training data ultimately consisted of images that the person in the scanner had never actually seen.
The essence of this approach is that the encoder predicts the "expected" brain scan from an image, while the decoder reconstructs the image from that prediction. Both models learn from each other's errors and improve step by step — artificially expanding the limited set of real scans.
One-Hour Calibration
The institute's official announcement says Brain-IT can identify brain activity patterns shared across people — combining data from different studies showed that certain brain regions perform similar tasks in everyone. This shared nature gives the system its biggest advantage: while earlier brain-decoding tools required about 40 hours of fMRI data to work with a new person, Brain-IT needs just one hour of calibration data.
This difference matters a great deal in practice: instead of tens of hours of scanner sessions for a new participant, one hour of calibration is enough — so researchers spend less time and can test more people. The system thus doesn't need lengthy individual preparation to "read" a new person's brain. The institute itself described this result as "unprecedented speed and accuracy" in reading a new person.
According to the institute, the technology could in the future enable fully paralyzed ("locked-in") patients to communicate through brain activity alone, and also reveal how the brain works.
Warnings on Mental Privacy
But independent scientists are pointing to the flip side of this breakthrough: the same approach could expose people's inner thoughts and mental images — possibly without their consent. Tommy Sprague, a neurobiologist at the University of California, Santa Barbara, and an independent expert not involved in the study, told MIT Technology Review:
"The results look really impressive. But if there were a way to surreptitiously extract information about what you're thinking about, 150 years of science fiction could come true at any moment — which is concerning in many ways."
Adding to the concern: the team is now working on extending a similar approach to EEG-based decoding — the brain's electrical activity is collected via electrode caps or even earbuds. EEG is a method of recording the brain's electrical activity through electrodes placed on the scalp.
At the same time, the Weizmann Institute points to helping fully paralyzed patients and uncovering how the brain works as the technology's primary goals — the researchers designed this tool first and foremost for medicine and fundamental science. So on one side stand great prospects like a communication channel for paralyzed patients, and on the other serious questions about mental privacy — and both stem from this very research.
