An AI “mind-reading” tool can reconstruct what you’re looking at based on a brain scan

A new AI tool developed by Michal Irani and her team at the Weizmann Institute of Science in Rehovot, Israel, can reconstruct what a person is looking at by analyzing their brain scans. The tool uses functional magnetic resonance imaging (fMRI) scans, which track blood flow in the brain to identify active areas. While traditional fMRI scans highlight brain activity in voxels (3D pixels) covering about 3 cubic millimeters, Irani’s team used higher-resolution scans with voxels of 1 cubic millimeter. The AI model consists of two parts: a decoder that predicts the structure and content of an image from brain activity, and an encoder that does the reverse—predicting brain activity based on an image. The tool was trained on data from eight people who viewed around 9,000 images each in an fMRI scanner. The reconstructed images are compared to the originals, showing remarkable accuracy in recreating details like colors, shapes, and positions.

The AI model uses a two-branch approach to improve image reconstruction. The first branch predicts the structural details of an image, such as colors and shapes, while the second branch identifies the content, like objects or scenes. This dual approach allows the model to generate more accurate representations. To overcome limited training data, the team developed an encoder that predicts brain activity from an image, which is then used to train the decoder. This iterative process enhances both models. The team combined data from multiple studies to identify brain regions with shared functions across individuals, such as areas responding to food or sports images. The model requires only one hour of fMRI data from a new person for calibration, unlike previous tools that needed up to 40 hours. This efficiency was presented at the Cognitive Computational Neuroscience conference in New York in 2026.

Despite its advancements, the AI tool is not perfect. Examples of failures include reconstructing an image of a cake as a pile of sandwiches or a dog in a bathtub as a goat in a bathtub. The tool represents the current state of the art but still has room for improvement. In comparison tests, it outperformed previous models significantly, though accuracy varies depending on the complexity of the image. The team plans to expand the tool’s capabilities to include video and audio reconstruction, as well as decoding mental imagery and dreams. These goals remain unachieved but are part of ongoing research. The tool’s potential applications include helping locked-in patients communicate and aiding in the study of brain functions, such as understanding PTSD flashbacks.

The development of brain-decoding technology raises significant ethical questions about mental privacy. Scientists like Tommy Sprague warn that such tools could be misused to extract thoughts or images without consent, potentially violating personal privacy. Marcello Ienca, a neuroscientist and philosopher, highlights the risk of commercial exploitation or legal misuse of reconstructed mental images. While Irani’s tool currently requires fMRI scans, which are difficult to obtain without a person’s cooperation, future technologies like EEG (electroencephalography) could make brain decoding more accessible. EEG measures electrical brain activity and could be used with minimal calibration, making it easier to extract information from brain activity. Irani acknowledges these concerns but focuses on the positive potential of the technology.