Israeli scientists develop AI that reconstructs mental images from brain scans with striking accuracy

A team of Israeli researchers has built an artificial intelligence system that can look at the electrical activity in your brain and reconstruct what you're picturing in your mind. The system, called Brain-IT, was developed at the Weizmann Institute of Science under the direction of Michal Irani and represents a significant jump in what neuroscience and machine learning can accomplish together. When a person lies in an fMRI scanner and imagines an image, Brain-IT analyzes the resulting brain scan data and generates a visual approximation of what that person was thinking about—not a vague blur, but something with recognizable detail and structure.

The work was submitted to the International Conference on Learning Representations and demonstrates a level of precision that earlier attempts at brain-image decoding could not achieve. The system can reconstruct everything from simple objects like street signs to complex scenes—a pizza with tomato and olive, rendered with the exact number of slices the viewer had in mind. The process takes roughly an hour from scan to image, a substantial reduction from the multi-hour processing times required by previous systems. What makes this notable is not just speed but fidelity. The visual outputs align closely with the actual images the subjects had viewed, capturing both the physical layout and the semantic content—the meaning—of what was in their minds.

This represents a dramatic improvement over earlier work in the field. In 2017, researchers at Purdue University developed models that could identify the general category of an imagined image but produced results that were noticeably blurry and lacked definition. By 2022, Japanese researchers had applied diffusion-based methods to sharpen the approximations, though their approach relied heavily on an intermediate step of converting brain data to text and then text to images. Brain-IT bypasses these intermediate translations and works directly from brain scan to image, achieving both speed and clarity that neither predecessor could match.

The central challenge in this kind of research is practical: gathering brain imaging data is expensive and time-consuming. The team trained their system on data from eight subjects who each viewed more than 70,000 images while inside fMRI machines. Eight people is a small cohort for training a frontier artificial intelligence system, so the researchers developed an ingenious workaround. They created what they call a dual-translation encoder—a mechanism that could generate synthetic training data without requiring additional human subjects in the scanner. The encoder works by taking a random image that was never actually viewed in an fMRI machine, translating it into a predicted brain scan, and then translating that predicted scan back into an image. By cycling through this process repeatedly during training, the models learned to generate brain scans that effectively encode visual information, even though those scans were never physically performed. Irani explained that this approach allowed the system to build itself a massive dataset, enabling the models to learn from far more examples than the eight subjects could provide.

Through this process, the researchers identified 128 distinct functional brain regions that participate in visual processing. Some of these regions were already known to neuroscience; others appear to be newly identified. The mapping revealed that different clusters of brain activity respond predictably to different categories of images. Certain areas light up when subjects view food, others activate for sports imagery, and separate neural pathways correspond to indoor versus outdoor scenes. This suggests that visual processing in the brain is organized by category and context in ways that the AI system has learned to read and decode.

The researchers see practical applications emerging from this work, particularly for people with severe paralysis who have lost the ability to speak or move. An assistive communication system based on Brain-IT could potentially allow such patients to express thoughts and images directly through brain activity, bypassing the need for physical speech or movement. The technology remains in the research phase, and significant questions remain about how it would scale beyond the eight subjects used in training, how it would perform with people whose brains might process visual information differently, and what ethical frameworks should govern its use. But the core achievement is clear: the system demonstrates that the relationship between what we imagine and the patterns of activity in our brains can be read and reconstructed with a precision that was not possible just a few years ago.