Researchers can now rebuild a rough picture of what you were looking at from a brain scan. Not your private thoughts. Not a transcript of your inner voice. A sketch of the scene your eyes just saw — close enough that a banana stays a banana in roughly the right place, and a leopard still looks like a leopard.
A team at the Weizmann Institute of Science, led by computer scientist Michal Irani, built an AI system that reads functional MRI (fMRI) — a scan that tracks blood-oxygen changes as a stand-in for brain activity — and turns those signals into images. In tests, their reconstructions beat prior state-of-the-art methods on standard scores of structure and meaning. With about one hour of calibration data from a new volunteer, the tool roughly matched older systems that needed around forty hours of scanning.
That is a real jump in how faithfully machines can decode vision from the scanner. It is also a reminder of the privacy stakes if similar methods someday move from hospital magnets to cheaper EEG caps — electrode headsets that listen to electrical brain signals without a giant machine.
As an Amazon Associate I earn from qualifying purchases. If you buy through links on this page, I may earn a commission at no extra cost to you.
What the tool actually does
Volunteers in large public datasets lay in high-resolution fMRI scanners and looked at thousands of everyday photos. The scanner recorded which tiny cubes of brain tissue — called voxels — lit up for each picture. Earlier AI “decoders” could often guess the category (a banana, a dog) but scrambled layout, color, and position. Irani’s group wanted reconstructions that kept both the meaning of the scene and its rough structure.
Their system, called Brain-IT (Brain Interaction Transformer), was presented for ICLR 2026 and detailed on arXiv. It splits the job into two cooperating parts. One branch predicts high-level semantic features — what the picture is “about” — to steer a diffusion model, the same family of AI that builds images by gradually cleaning noise into a clear picture. The other branch predicts low-level structural features (layout, color blobs, coarse edges) and turns those into a rough draft using a method called Deep Image Prior. At generation time, the rough draft initializes the diffusion process, then the semantic branch refines it.
Think of it like handing an artist a blurry thumbnail plus a short note that says “two giraffes, outdoor light,” then letting them finish the painting. Separately, either cue is incomplete. Together, the results look much closer to what the person actually saw.
Why one hour of scanning matters
Collecting fMRI is expensive — researchers quoted in MIT Technology Review put typical rates near $600 to $1,000 an hour. Older decoding tools often needed on the order of forty hours of a new person’s data before they worked well. Brain-IT leans on functional clusters of voxels that are shared across people, plus network weights that are also shared. Only a thinner set of person-specific settings needs to adapt.
On the Natural Scenes Dataset (NSD) — a 7-Tesla fMRI collection where subjects viewed thousands of COCO photos — Brain-IT with one hour of new-subject data scored in the same ballpark as leading prior methods trained on the full ~40-hour recordings. The paper even shows meaningful reconstructions with as little as fifteen minutes of subject-specific data, though quality continues to improve with more time.
That efficiency is why neuroscientist Tommy Sprague, of UC Santa Barbara, told MIT Technology Review the work looks valuable for labs that cannot buy dozens of hours per volunteer. Faster calibration means more studies, more subjects, and less wasted magnet time.
The encoder that feeds the decoder
Brain-IT sits on top of earlier Weizmann work: a Universal Brain Encoder sometimes nicknamed “Wisdom of a Crowd of Brains.” That encoder runs the other direction. Given an image, it predicts what each brain voxel would do. Why bother? Because real paired fMRI-and-photo data is scarce. By predicting fake-but-useful fMRI for many extra photos, the team can train the decoder on far more examples than any single scanner session provides. Irani has said roughly seventy percent of their training material can come from images that never had a real fMRI pair.
The encoder is voxel-centric: each tiny brain cube gets its own learned “embedding” vector that captures what that spot seems to care about (edges vs. faces vs. scenes, for example), while most of the network is shared across people and even across different scanner strengths. That design is what lets a crowd of brains improve encoding for each individual, and what lets Brain-IT transfer to a new person with little data.
The tools still fail. Irani has pointed to a cake reconstructed as a stack of sandwiches, and a dog in a bathtub that came out looking more like a goat. Faithful is not perfect. The paper’s own failure gallery shows cases where structure looks right but the object is wrong, or the other way around.
Not sci-fi mind reading — yet
Irani calls “mind reading” a jazzy nickname for what they are doing now: reconstructing seen images from willing volunteers who stay still in a scanner. The group hopes the same line of work could someday help locked-in patients communicate, or let scientists study mental imagery, PTSD flashbacks, and dream content. Those goals are still ahead. Reconstructing imagination and dreams is not solved.
Neuroethicist Judy Illes of the University of British Columbia called the research “magnificent” and highlighted the therapeutic upside for people with neurologic conditions. Sprague praised the results but flagged the darker reading: if methods improve enough to extract what someone is thinking about without clear consent, decades of science-fiction worry stop being fiction.
Philosopher and neuroscientist Marcello Ienca of the Technical University of Munich told MIT Technology Review that moving similar decoding to EEG would be a “game changer.” Once a consumer headset is calibrated to your brain, companies might pull more information than you meant to share. He also raises the prospect of courts treating reconstructed mental imagery as evidence someday. Irani acknowledges EEG-era misuse risks and says she is trying, for now, to focus on constructive uses.
What to take away
Here is the sober version. AI can already sketch what you saw from a research fMRI scan with surprising fidelity, and it can adapt to a new brain after about an hour instead of a workweek in the magnet. That helps neuroscience and, potentially, clinical communication. It does not mean an app can silently dump your thoughts onto a screen today.
The near-term story is scientific: better maps of how vision is represented across brains, cheaper experiments, and clearer failure modes. The medium-term story is ethical: rules for consent, data access, and consumer brainwear before the hardware catches up to the algorithms. The pictures are getting sharper. The privacy conversation should keep pace.
Further reading
The Idea of the Brain — Matthew Cobb — A clear history of how we have imagined the brain — from hydraulics and telephone exchanges to today’s computer metaphors — and why the organ still resists a neat explanation.
Being You — Anil Seth — A leading consciousness researcher’s tour of how the brain builds a sense of self and a stable world — useful context when AI starts decoding visual experience from scans.