Researchers at the Weizmann Institute of Science in Rehovot, Israel have built an artificial intelligence system that can reconstruct what a person is looking at based only on their fMRI brain scan — and predict a person's brain activity from an image in the other direction — MIT Technology Review reported on October 1, 2026.
The tool, developed by computer scientist Michal Irani and her colleagues, recreates viewed images with a precision that earlier brain-decoding efforts have not matched. In side-by-side comparisons, the model's reconstructions preserve not just the general subject of a scene but its structure, position and content — a persistent weakness of previous systems, which could generate a plausible banana but rarely the banana the person actually saw. For more context on this story, see our ongoing AI industry coverage.
How the brain decoder works
The system is a two-branch "brain decoder." One branch predicts the structure of the image — where colors and shapes sit in the frame — while the second predicts its content, for example a bunch of bananas on a plate. Those combined predictions then steer a diffusion model, the same family of generative AI techniques behind modern image and video generators, to produce the final reconstruction.
The team started from publicly available brain scan data: volunteers had been shown hundreds of images while lying in fMRI scanners, which track oxygenated blood flow through the brain as a proxy for neural activity. Irani's group used newer datasets captured on higher-resolution scanners, where each voxel of activity covers roughly one cubic millimeter of tissue rather than the roughly three cubic millimeters of a standard scanner.
Training on borrowed data
The central obstacle was data hunger. High-resolution fMRI scans from a person viewing thousands of images are scarce, and the models needed far more than existed. The researchers' workaround was to train a second model in reverse — an encoder that predicts what a brain scan would look like if a person were shown a given image.
By running the encoder and decoder against each other, the team could generate effectively unlimited training pairs. Start with a photo of a leopard, predict the fMRI activity it would produce, reconstruct the image from that activity, and refine both models on the mismatches. Around 70 percent of the training data ultimately came from images that were never actually shown to a scanned volunteer, according to Irani.
The approach produced what the team calls a universal brain encoder, and its efficiency is the headline result. Previous decoding tools typically required about 40 hours of fMRI data to calibrate to a new person. Irani's decoder needs roughly one hour — a difference that matters because scanner time is expensive. "None of us can afford 40 hours of imaging for a new subject," said Tommy Sprague, a neuroscientist at the University of California, Santa Barbara, who was not involved in the research, noting that imaging can run roughly $600 to $1,000 per hour. The findings were presented at the Cognitive Computational Neuroscience conference in New York in September.
What it gets right — and wrong
In comparison tests, the tool outperformed previously described brain-decoding systems "by a significant margin," Irani told MIT Technology Review. It is not infallible. Over a video call, she pointed to an image of a cake that the system reconstructed as a pile of three sandwiches, and a dog in a bathtub that came back as a similarly colored goat in the same tub.
By pooling data across multiple studies, the team also identified brain regions that appear to share functions across individuals — one region responded to images of food, another to sports. Irani says she is now working with neuroscientists to use the tools to map how the brain organizes meaning more systematically.
Promises for medicine
The most immediate hopes are clinical. Irani believes the approach could eventually help people with locked-in syndrome — fully paralyzed patients who cannot speak or move — communicate using brain activity alone. She also wants to extend the technique beyond static images to video and audio, with the long-term goal of reconstructing imagined content, including dreams and the sensory content of PTSD flashbacks.
Judy Illes, a neuroethicist and professor of neurology at the University of British Columbia who was not involved in the work, described the research as "magnificent." "The idea [of using this approach] to help people with neurologic conditions ... therapeutically is tremendously exciting," she told MIT Technology Review.
The privacy problem
The same precision raises the prospect the field has long dreaded: extracting a person's mental imagery without meaningful consent. "The results seem very impressive," said Sprague. "But if there's a way to surreptitiously extract information about what you're thinking about, then ... 150 years of sci-fi can come true anytime, and that's worrisome in a lot of ways."
Those concerns remain, for now, bounded by physics. The system requires a person to lie inside a multi-ton fMRI magnet for hours, and every previous decoder has needed extensive per-person calibration. But the calibration requirement is precisely what Irani's team just cut by a factor of dozens, which is why ethicists are paying attention to a tool that also happens to be a research breakthrough.
The Weizmann group's next targets — moving images and sound, then imagined and dreamed content — will test whether the technique's accuracy survives outside the carefully controlled conditions of a scanner lab. If it does, both the medical applications and the privacy debates will escalate together.
A fast-moving field
Brain decoding has progressed from blurry, barely recognizable approximations to structurally faithful reconstructions in under a decade, driven by better scanners, larger shared datasets and generative models that can fill in realistic detail from coarse signals. Weizmann's contribution — using a bidirectional encoder-decoder loop to break the data bottleneck — is the kind of engineering insight that tends to spread quickly, because it does not depend on new hardware. Expect other labs to replicate the training scheme, and expect the question of who controls and consents to neural data to move from the margins of AI policy toward its center.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
