Scientists using artificial intelligence have managed to observe a person's brain activity and reconstruct images of what they were seeing; this work was conducted by researchers at the Weizmann Institute of Science using functional magnetic resonance imaging (fMRI). During the experiment, volunteers viewed various images while the team recorded changes in activity across different regions of the brain.

Based on these signals, the artificial intelligence had to solve an extremely difficult problem: determining which image might have triggered that specific pattern of brain activity. To achieve this, the researchers split the task into two parts: one system attempted to extract information regarding the image's structure—such as shapes, positions, and composition—while another sought to interpret higher-level information about the scene's content.

This information was then combined and used by an image-generation model to reconstruct what the person was viewing. The results are impressive, particularly because the AI ​​does not receive the original photograph itself, but rather the signals produced by the brain while someone looks at that photograph. While the reconstructions successfully preserve key features of the original scene in some instances, we remain far from having a machine capable of perfectly reading the contents of someone's mind—a fact made quite clear by the errors that occur.

In one of the examples described by the researchers, a person was looking at a dog in a tub; the artificial intelligence reconstructed a goat in a tub—capturing much of the structure and even the meaning of the scene—but it failed specifically on the animal's identity.

There is another issue as well: training systems of this type requires vast amounts of brain data, and obtaining thousands of high-resolution fMRI scans is far more complex than simply gathering millions of photos from the internet. Consequently, researchers also developed a reverse process: they show an image to a model and ask it to estimate the brain activity pattern that image would likely elicit.

This synthetic brain data can then be used to help train the system attempting to map the reverse path—from brain activity back to the image. For now, however, the process relies on large MRI scanners and extremely controlled experimental conditions; nor does it mean that someone can simply place a person in a machine and uncover any thought passing through their mind.

What researchers are demonstrating is something far more specific: the possibility of reconstructing visual content represented in brain activity. Yet, when comparing this experiment to what Neuralink is doing, a rather intriguing situation emerges: on one hand, machines are learning to translate our intentions into digital commands; on the other, they are beginning to transform our brain activity back into images. Gradually, the barrier that—throughout history—kept the inner workings of the brain accessible only to the individual is beginning to open up small windows to the outside world.