Researchers at the Weizmann Institute of Science have developed an AI system that can reconstruct images a person is looking at based solely on their brain scans. The tool uses a two-branch decoder—one branch predicts spatial structure, the other predicts semantic content—and feeds those predictions into a diffusion model to generate a close visual match to the original image. The team also built an encoder that works in reverse, predicting brain activity from images, which let them train the system on far more data than was available from fMRI sessions alone.
The system was tested on public datasets from volunteers who viewed thousands of images in high-resolution fMRI scanners. Unlike earlier decoders that required roughly 40 hours of a new person's scan data, this one needs only about one hour, making it far more practical for research. The developers hope it will help neuroscientists understand how the brain represents visual information and eventually assist people with locked-in syndrome or even allow researchers to recreate dreams.
Outside experts are impressed but cautious. Neuroethicist Judy Illes called the work “magnificent” and said the therapeutic potential is exciting. But neuroscientist Tommy Sprague warned that if brain data can be extracted surreptitiously, it could enable unwanted mind reading. The team acknowledges the system is not perfect—some reconstructions fail—but the combination of improved fMRI resolution and AI training techniques marks a clear step forward in decoding human thought. The work was presented at the Cognitive Computational Neuroscience conference in New York. The source does not report any disagreement among the researchers themselves, only differing reactions among outside experts.}```json lines? The user asked to return strict JSON. We need to output only JSON object. The above is the JSON. Ensure no markdown code fences. The final answer should be raw JSON. Let's verify schema: headline, dek, keyPoints array of 3 strings, body string. Good. Headline under 80 chars: