Researchers in Israel have developed an artificial intelligence system that can “read” patterns of activity inside a person's brain to work out what they are looking at.
It then reconstructs images of what they are seeing with remarkable accuracy.
The team based at the Weizmann Institute of Science showed participants a series of images - a surfer, a clock tower, a beach, a giraffe, a banana next to a bowl, a sleeping woman, a man holding a glass, two birds on a branch, and people playing a game of frisbee - while they were inside an MRI scanner.
They then asked AI to work out what images the participants had seen, based just on the record of brain activity in the MRI scan, and without AI being shown the actual pictures.
The AI system was able to reconstruct images, preserving key elements of the originals.
Called Brain-IT, it could eventually help doctors assess awareness in patients who have suffered a catastrophic stroke, severe brain injury or other conditions that leave them unable to speak or move.
The system could allow neuroscientists to investigate how the brain represents objects, scenes, colours and spatial relationships.
One day, it may even help researchers study what happens in the brain when people imagine, dream or hallucinate, although Brain-IT has not been shown to reconstruct dreams or arbitrary thoughts.
Conventional MRI can show doctors what the brain looks like and reveal physical damage.
Functional MRI, or fMRI, uses the same machinery but goes a step further, showing which parts of the brain are becoming active by detecting changes in blood flow.
The combination of fMRI and AI can then attempt to interpret those patterns of activity - in this case, reconstructing a picture of what the person is seeing.
The potential advantage is that instead of simply showing doctors where brain activity is occurring, the AI attempts to interpret what that activity represents.
The idea of putting fMRI and AI together is not new. What sets Brain-IT apart is the reported improvement in the quality of the reconstructions and the much smaller amount of brain-scan data needed to adapt the system to a new person.
"There exist nowadays models that translate brain activity into images, and they can even produce impressive reconstructions that preserve the semantic meaning of the image reasonably well," says Michal Irani, professor at the Department of Computer Science and Applied Mathematics, who leads the team.
Professor Michal Irani (middle) and her team (left-right) Navve Wasserman, Jonathan Kogman, Roman Beliy and Amit Zalcher (credit-Weizmann Institute of Science)[Missing Credit]
"However, they tend to make mistakes in basic features such as composition and colour.
"The new model we developed outperforms them in reconstructing both the content of the image and its details. What's more, while every other model requires dozens of hours of brain scans to learn to 'read' a new person, our model needs only one hour."
The Weizmann Institute of Science described Brain-IT as “a ‘mind-reading’ AI model developed in the lab of Prof Michal Irani” which can reconstruct a scene “based solely on the brain activity of a person viewing it”.
Prof Irani’s team are Roman Beliy, Amit Zalcher, Jonathan Kogman and Navve Wasserman.
The researchers first show a participant a large number of known images while recording their brain activity with fMRI.
The scanner detects tiny changes in blood oxygen across thousands of small three-dimensional areas of the brain, known as voxels.
The researchers know exactly which picture the person was looking at and can match it with the pattern of brain activity recorded at the same time.
Over time, the AI learns that particular patterns of activity are associated with different visual information, such as objects, faces, colours, shapes and their approximate positions in an image.
Once it has learned these relationships, the researchers can show the participant a new picture.
The AI system is then given the brain scan, but not the picture itself.
It extracts clues from the brain activity about what the image contains and how its elements are arranged. It then uses an image-generation system to produce a new picture that fits those clues.
The result is not a photograph being pulled out of the brain.
It is an AI-generated reconstruction based on the pattern of brain activity produced when the person looked at the original image.
One of the main challenges is that people's brains are different. The same image does not produce exactly the same pattern of activity in every person.
Brain-IT addresses this by sharing most of the AI model across participants, rather than training a completely new system for every individual.
A relatively small amount of additional brain-scan data is used to adapt the model to each person's brain.
The researchers report that about one hour of data from a new person can produce results comparable to methods using around 40 hours of data. They also report meaningful reconstructions using as little as 15 minutes of data.
That could make a substantial difference to future research.
MRI scanning is expensive and time-consuming, so requiring dozens of hours from every participant would make large studies difficult.
A system that can use an existing model and adapt it to a new person with much less scanning could allow researchers to work with more participants and conduct studies that would otherwise be difficult or expensive.
It could also make the technology more practical to investigate for medical applications.
One possible use is in patients who cannot communicate.
A person with severe brain injury, stroke or another neurological condition may be unable to speak or move even when some brain functions remain intact. Doctors can use conventional examinations and brain imaging to look for signs of brain function, but determining what an unresponsive patient is actually processing can be difficult.
A system such as Brain-IT could eventually provide another source of information by attempting to show what a patient is seeing or responding to.
For now, Brain-IT is a research system rather than a clinical tool.
But Prof Irani and her team are working on decoding auditory information.
She added: “What remains especially challenging is decoding video - for example, during dreaming.
“Dozens of images change every second, while an fMRI scan takes about two seconds.
“If we overcome all these obstacles, it’s possible that in the future we may even be able to read dreams.”
The research was presented at the International Conference on Learning Representations (ICLR) in Rio de Janeiro in April 2026.
To get more news, click here to sign up for our free daily newsletter.