{
  "id": 11177536,
  "title": "An AI “mind-reading” tool can reconstruct what you’re looking at based on a brain scan",
  "url": "https://urgent.news/2026/10/01/an-ai-mind-reading-tool-can-reconstruct-what-youre-looking-at-based-11177536",
  "topic": "ai",
  "section": "AI",
  "published": "2026-10-01T10:32:24.000Z",
  "source": {
    "name": "MIT Tech Review Biotech",
    "slug": "mit-tech-review-biotech",
    "url": "https://www.technologyreview.com/2026/10/01/1145588/ai-mind-reading-reconstructs-what-youre-looking-at/"
  },
  "original_language": "en",
  "account": "A new AI tool developed by Michal Irani and her colleagues at the Weizmann Institute of Science in Rehovot, Israel, can reconstruct images based on brain scans—a process referred to as \"mind-reading.\" This tool can also predict brain activity based on what a person is looking at. In the accompanying image, the left side of each pair represents the actual image the user saw, while the right side displays the AI's recreation derived from the brain scan.\n\nIrani and her team hope their mindreading tool will provide deeper insights into brain function and potentially assist individuals with locked-in syndrome in communicating. Additionally, the tool might enable scientists to recreate the content of dreams. However, neuroethicist Judy Illes, who was not part of the research, views the work as \"magnificent\" and finds the prospect of using such an approach to help people with neurological conditions \"tremendously exciting.\" However, she and others caution that similar techniques could be misused to surreptitiously reveal a person's thoughts and mental imagery without their consent. Neuroscientist Tommy Sprague from the University of California Santa Barbara agrees, expressing concern that \"150 years of sci-fi can come true anytime\" due to the potential for surreptitious information extraction.\n\nNeuroscientists have been working on reconstructing people's perceptions and inner thoughts from brain scans for years. Early attempts yielded blurry and incomprehensible results, but technological advancements in fMRI scans and image analysis tools have significantly improved the quality of reconstructions. Irani and her colleagues began their project by analyzing publicly available brain scan data collected from volunteers who were shown hundreds of images while in fMRI scanners. fMRI utilizes a powerful magnet to monitor blood flow to active brain regions, although each highlighted voxel represents an activity area covering about three cubic millimeters. To enhance image accuracy, Irani's team used newer datasets with higher resolution—a voxel covering approximately one cubic millimeter of neurons—allowing for better identification of brain regions involved in specific tasks.\n\nThe team trained an AI model on data from eight participants who each viewed around 9,000 images in a high-resolution fMRI scanner. This model incorporated two branches: one predicting image structure (such as colors and shapes) and the other predicting image content (like specific objects or scenes). By employing a diffusion model—a type of AI used for generating images from noisy data—they were able to create more accurate reconstructions. To further refine their models, Irani and her colleagues trained an encoder to predict brain activity from images. These tools were trained together, enabling the diffusion model to reconstruct images from the encoded brain activity patterns and the encoder to predict brain activity from newly presented images.\n\nInitially, reconstructions were not very accurate. For example, an image of a banana might appear distorted. To address this issue, the team employed a training loop where an image was first used to predict the corresponding fMRI scan, and then the decoder reconstructed the image from those predictions. Repeatedly cycling through this process led to significant improvements in image fidelity. Remarkably, 70% of the training data came from images not originally paired with fMRI scans from human subjects. By integrating data from multiple studies, Irani's team identified shared brain regions across individuals, such as one region responding to food images and another to sports images.\n\nIrani is currently collaborating with neuroscientists to explore potential applications of their tools to uncover new insights into brain function. The researchers presented their findings at the Cognitive Computational Neuroscience conference in New York, where they noted that their decoder requires only one hour of fMRI data per new subject, compared to the standard 40 hours needed by other tools. This efficiency makes the tool valuable for neuroscientists conducting research, as fMRI sessions can be costly, ranging from $600 to $1,000 per hour.",
  "summary": "A new AI tool can guess what you’re looking at just by analyzing your brain scans—and recreate that image with remarkable precision. It can go the other way too, and predict a person’s brain activity based on what they’re looking at. In the image above, for example, the left-hand image of each pair is what…",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 2,
    "also_reported_by": [
      {
        "outlet": "MIT Technology Review",
        "title": "An AI “mind-reading” tool can reconstruct what you’re looking at based on a brain scan",
        "url": "https://urgent.news/2026/10/01/an-ai-mind-reading-tool-can-reconstruct-what-youre-looking-at-based",
        "published": "2026-10-01T10:32:24.000Z"
      }
    ]
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}