This AI system can turn brain activity into images: Israeli scientists develop new visual decoding system

What if an AI could look at your brain activity and work out what image you are seeing? Researchers at Israel’s Weizmann Institute of Science have developed a system that brings that idea closer to reality. Called Brain-IT, the AI uses functional magnetic resonance imaging (fMRI) data to study patterns of brain activity and reconstruct the visual information being processed by a person.

The system was developed in Professor Michal Irani's lab and was selected for presentation at the International Conference on Learning Representations (ICLR) 2026. According to the researchers, Brain-IT can reconstruct images while preserving both their overall content and finer details. More importantly, the system needs about one hour of fMRI data from a new person, compared with the dozens of hours required by earlier approaches.

Must Read: Anthropic brought a Hindu monk to San Francisco to help Claude tackle AI ethics and consciousness

How Brain-IT turns brain scans into images

The researchers worked with a dataset containing fMRI scans from eight people, each of whom viewed several thousand images. Building a much larger dataset would require scanning many more people for long periods, making the process difficult and expensive.

To get around this problem, the team developed an encoder alongside the image decoder. The encoder can take an image that has never been shown to a person inside an MRI scanner, predict the brain activity it could produce and then send that information back through the system to reconstruct the original image. This approach effectively helped the researchers generate additional training examples without collecting a fresh brain scan for every image.

Must Read: OpenAI agent sparks chaos again; Wikimedia flags millions of requests, unauthorised edits

The model works by breaking brain activity into functional patterns and linking them with visual information. The researchers say Brain-IT can reproduce the content of images while also improving details such as composition and structure. The official ICLR paper says its results are comparable to methods that previously required around 40 hours of recordings, despite using only one hour of data from a new subject.

What could this mean for AI and the brain?

During development, the encoder identified 128 functional brain regions shared across people that perform different roles in processing images. Researchers found, for example, differences in how parts of the brain's place-processing region respond to indoor and outdoor scenes.

The technology could eventually help researchers understand how the brain processes visual information and could have applications in assistive communication. The team is also exploring whether similar methods could decode other forms of information, including sound.