AI Can Now Recreate Thoughts From Brain Scans With Startling Accuracy

Israeli scientists have developed a new artificial intelligence program called Brain-IT that can analyze brain scans and recreate visual depictions of a person's thoughts with remarkable accuracy. The system produces reconstructions that researchers say are structurally and semantically more accurate than previous iterations.

Researchers from the Weizmann Institute of Science demonstrated the technology by showing participants various images and then using fMRI scans to reconstruct those images. The AI was able to recreate visuals ranging from specific objects like stop signs and pizzas to more complex scenes, matching details such as the number of slices on a pizza that a person envisioned.

This latest development builds on previous research in mind-reading AI. In 2017, Purdue University researchers created a model that could predict thoughts but only produced grainy recreations. More recently, in 2022, Japanese scientists developed an AI model that could approximate thoughts more precisely but relied on text-to-image technology.

Brain-IT, however, can reportedly recreate images within an hour, a significant improvement over the multiple hours required by earlier systems, according to Michal Irani, who led the development at the Weizmann Institute. "We realized that by translating back and forth – from a random image that had never been viewed in an fMRI machine, to a predicted brain scan, and then back to the image we started with – the models would effectively build themselves a massive dataset," Irani explained in a statement. "In this way, during training, the models would learn to generate scans that encode images effectively, even though those fMRI scans had never actually been performed."

The system was trained using over 70,000 images and corresponding brain scans from eight participants. Because acquiring large datasets of direct brain scans can be time-consuming and expensive, the researchers employed an "encoder" shortcut to create a substantial training set. This method allowed the AI to decode images from brain scans and generate scans that encoded images without requiring every scan to be performed.

During the experiments, Brain-IT identified 128 functional regions in the brain that process visual information, some of which were previously known and others newly discovered. These regions responded to different categories of images, such as food or sports, and to spatial concepts like indoor versus outdoor scenes.

Researchers suggest that Brain-IT could have significant applications, such as helping paralyzed individuals communicate or enabling more efficient and cost-effective brain studies. Tommy Sprague, a neuroscientist at the University of California, Santa Barbara, noted the potential for cost savings, as extensive imaging sessions can cost hundreds or even thousands of dollars per hour.

Looking ahead, the research team hopes to expand the technology's capabilities, with Irani suggesting that future advancements could potentially allow for the reading of dreams. However, the development also raises privacy concerns, with some experts worried about the potential for surreptitiously extracting thoughts, echoing long-standing themes in science fiction.