Not in the unrestricted sense suggested by “psychic AI.” Brain-IT is a research system that reconstructs images people have viewed from their fMRI recordings. It does not show that AI can secretly read arbitrary private thoughts, decode spontaneous mental imagery, or work without a brain scanner and a specific experimental setup.
What Brain-IT actually does
In a 2026 ICLR paper, Roman Beliy, Amit Zalcher, Jonathan Kogman, Navve Wasserman and Michal Irani describe Brain-IT, a method for reconstructing viewed images from functional magnetic resonance imaging (fMRI) recordings. Its task is visual decoding: a person views images while their brain activity is recorded, and the model uses those recordings to produce an image reconstruction. Read the ICLR 2026 paper record.
| # | Preview | Product | Price | |
|---|---|---|---|---|
| 1 |
|
Inferno Application | $15.99 | Buy on Amazon |
As an Amazon Associate I earn from qualifying purchases.
The system predicts two kinds of image information. A low-level branch estimates structural features and helps establish a coarse layout; a semantic branch estimates content-related features that guide a diffusion model. The authors also use clusters of functionally similar brain voxels shared across subjects, rather than treating every voxel as an isolated signal. These components help generate a plausible reconstruction, but the output is not necessarily an exact copy of the viewed photograph.
Recommended Free Tools
How the experiment was trained
fMRI datasets pair brain recordings with images that participants actually saw. The Weizmann Institute’s account says the Natural Scenes Dataset used in this work contains scans from eight participants and about 73,000 image-fMRI pairs. Those figures describe this particular dataset, not a representative survey of people or a universal test of mind-reading. Weizmann UK’s explanation of the work.
#1 Best Overall
To address the limited amount of scan data available for training, the team used a paired encoder and decoder approach. The encoder predicts fMRI activity from images, while the decoder reconstructs images from fMRI. The encoder can therefore generate estimated scans for images that were not actually shown to a participant in the scanner. Those generated scans are a training aid, not real measurements from a person.
What the one-hour and 15-minute results mean
The ICLR abstract reports that one hour of fMRI data from a new subject produced results comparable to methods trained on full 40-hour recordings. Separately, the authors’ project page displays transfer-learning reconstructions after 15 minutes of data from a new subject. These are distinct claims from different demonstrations; they should not be combined into a promise that any brief scan will reliably reveal a person’s thoughts. ICLR 2026 paper record · Brain-IT project page.
The project page also reports results averaged across Natural Scenes Dataset subjects 1, 2, 5 and 7, with Brain-IT outperforming baselines on seven of eight metrics. That is a comparison on the authors’ image-reconstruction task and evaluation setup. It is not a general accuracy score for decoding arbitrary thoughts, and no single broadly applicable percentage for such a task is established by the cited sources.
What it cannot establish
- It is not unrestricted thought-reading. The demonstrated input is fMRI data recorded while someone views images; the demonstrated output is a reconstruction of viewed visual content.
- It is not covert decoding without equipment. The method depends on brain scans and a trained model, not an ordinary phone, webcam or consumer AI device.
- It has not demonstrated dream reading or spontaneous imagery decoding. The work does not show that the model can reconstruct mental content a participant was not prompted to view.
- Reconstruction is not photographic certainty. A model-generated image guided by predicted features can capture content or broad structure without matching every detail of the original.
The institute reports that the team identified 128 functional regions shared across people, including a division within the place-processing area (PPA): one part responded to indoor scenes and another to outdoor scenes. This is the team’s reported finding within its study, not proof that every person’s brain is mapped identically or that all thought can be assigned to shared regions.
What researchers say may come next
Michal Irani, a professor at the Weizmann Institute of Science, says auditory decoding is an area the lab is working to extend toward. She describes video decoding as especially challenging because video changes across many frames per second while an fMRI scan takes about two seconds. Dream reading is presented as a possible future prospect only if substantial obstacles are overcome; it has not been demonstrated. Weizmann UK’s account.
The headline phrase “psychic AI” therefore overstates the result. Brain-IT is a significant research method for reconstructing images from recorded brain activity under controlled conditions. The evidence does not support the claim that it can secretly extract anyone’s private thoughts.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Fix the driver behind crashes, sound loss and screen glitches3Clear out junk files and repair common Windows errors
