AI That Reads Minds: What Japanese Scientists Built

Updated 6 November 2026 · Standard formula, shown below

No, Japanese scientists have not created an AI that can literally read your private thoughts, but they have built a system that can reconstruct images and words from brain activity with surprising accuracy. Researchers at the National Institutes for Quantum Science and Technology (QST) in Japan developed a technique that uses functional magnetic resonance imaging (fMRI) to decode visual perceptions and mental imagery, then feeds that data into a generative AI model to recreate what a person is seeing or imagining. The result is a rough, often blurry approximation of the original image, not a telepathic mind reader.

How the Technology Works

The system, described in a 2023 paper in the journal Neural Networks, combines two established tools: fMRI scanning and a latent diffusion model (the same type of AI behind image generators like Stable Diffusion). Here is the step-by-step process:

Brain scanning: A participant lies inside an fMRI machine and views hundreds of images (e.g., a bear, a clock, an airplane) while the scanner records blood flow changes in the visual cortex. This produces a map of which brain regions activate for each image.
Training a decoder: The researchers train a separate AI model to associate specific fMRI patterns with the corresponding images. This decoder learns the mapping between brain activity and visual features.
Generation: When the participant views a new image (or imagines one), the fMRI data is fed into the decoder, which produces a text description. That text is then passed to a latent diffusion model, which generates a new image based on the description.
Refinement: The generated image is compared against the brain activity again, and the system iteratively adjusts the image to better match the neural signal, improving accuracy.

The key innovation is that the system does not read thoughts directly; it reads the neural correlates of perception and then uses a separate AI to “translate” those correlates into a plausible image. This is a reconstruction, not a direct readout.

What It Can and Cannot Do

The QST system achieved around 75% accuracy in matching reconstructed images to the originals from a set of 1,200 images. However, that accuracy is relative to a fixed set of options; it is not a general-purpose mind reader. The technology has several hard limitations:

Requires fMRI: The system needs a massive, expensive, immobile scanner (costing millions of dollars) and a controlled lab environment. It cannot work with a portable EEG headset or a smartphone.
Training per individual: The decoder must be trained on each person’s brain for hours before it works. There is no universal “mind reading” model.
Limited to visual content: It reconstructs images and simple text descriptions, not abstract thoughts, emotions, or memories. You cannot ask it “what is the person thinking about their mother?” and get an answer.
Slow and noisy: fMRI captures brain activity over seconds, so it cannot track real-time thoughts. The reconstructed images are often distorted, missing colors, and sometimes completely wrong.

In short, it is a proof-of-concept for decoding visual perception, not a tool for reading minds in everyday life.

Why This Is Not Telepathy

The word “mind reading” implies accessing another person’s private mental state directly. This technology does not do that. It measures physical brain activity (blood flow) and uses statistical patterns to guess what visual stimulus produced that activity. It is analogous to looking at a fingerprint and reconstructing a face—possible only if you already have a database of fingerprints and faces from that same person. The AI is not “reading” thoughts; it is matching patterns. Furthermore, the system cannot decode imagined scenarios that have no visual component, like a memory of a conversation or a feeling of anxiety. It is strictly limited to visual imagery, and even then, only after extensive personal calibration.

Ethical and Privacy Concerns

Even though the current technology is crude, the ethical implications are significant. If brain decoding improves, it could threaten mental privacy—the right to keep your thoughts to yourself. Researchers in this field have proposed guidelines, including:

Informed consent: Participants must understand exactly what data is collected and how it will be used.
Data security: Brain data is highly sensitive and should be encrypted and stored securely, like medical records.
No coercion: Brain scanning should never be used without voluntary consent, especially in legal or employment settings.
Regulation: Governments may need to update privacy laws to cover neural data, which is currently not protected in most jurisdictions.

These concerns are speculative for now, but they are the reason many scientists urge caution. The technology is not here yet, but the groundwork is being laid.

Comparisons to Other Brain-Computer Interfaces

It is easy to confuse this fMRI-based decoding with other brain-computer interface (BCI) projects, but they are fundamentally different:

Elon Musk’s Neuralink: Implants electrodes directly into the brain to record electrical signals. It aims to let paralyzed people control a cursor or type with their thoughts. It reads motor intent, not visual imagery, and requires surgery.
EEG headsets: Consumer devices like Emotiv or Muse measure electrical activity from the scalp. They can detect simple states (attention, relaxation) but cannot reconstruct images or complex thoughts.
Speech decoding (e.g., from UC San Francisco): Uses electrode arrays on the brain surface to translate attempted speech into text. Again, this is motor-related, not visual.
The QST system: Non-invasive, uses fMRI, and focuses on visual reconstruction. It is the closest to “seeing” what someone sees, but it is slow and lab-bound.

Each approach has trade-offs: invasive methods get higher resolution but require surgery; non-invasive methods are safer but noisier. The QST system is notable because it achieves high-quality reconstruction without surgery, but it is not practical for daily use.

Future Possibilities and Limitations

What could this lead to? Potential applications include helping people who are locked-in (fully paralyzed but conscious) communicate by imagining images, or creating a new form of art from dreams. However, significant hurdles remain:

Speed: fMRI takes seconds to capture a single brain state. Real-time decoding is impossible with current hardware.
Cost and portability: fMRI machines are huge and expensive. Functional near-infrared spectroscopy (fNIRS) is portable but has lower resolution.
Individual variability: Brain patterns differ widely between people, so each user needs a custom decoder.
Semantic gap: The system reconstructs what a person sees, not what they think it means. A picture of a dog might be reconstructed as a furry blob if the decoder misinterprets the neural pattern.

Scientists are optimistic but cautious. The QST team has stated that their goal is to aid communication for disabled people, not to invade privacy. The technology will likely improve, but it will be decades before it is anything resembling a portable mind reader—if ever.

Frequently Asked Questions

Can this AI read my private thoughts?

No. It can only reconstruct visual images that a person is currently looking at or imagining, and only after the system has been trained on that person’s brain for hours. It cannot access memories, emotions, or verbal thoughts.

Is this the same as the “mind reading” AI in sci-fi movies?

No. Sci-fi mind readers can extract any information from a brain instantly. This system is a slow, noisy, statistical reconstruction of visual stimuli. It is more like a very advanced “guess the picture” game than telepathy.

Could this be used to interrogate criminals or witnesses?

Not currently. The technology requires a cooperative, trained participant in an fMRI machine, and it only works for visual content. It cannot detect lies or retrieve episodic memories. Even if it improves, legal use would raise major privacy and consent issues.

How accurate is the reconstruction?

In the QST study, the system correctly matched the reconstructed image to the original from a set of 1,200 images about 75% of the time. That sounds high, but it is a forced-choice test—the AI is picking from a known set. In open-ended scenarios, accuracy would be much lower.

Do I need to worry about companies reading my brain through a headset?

No. Consumer EEG headsets cannot decode thoughts. They measure general brainwave patterns, which are too noisy to reconstruct specific images or words. The QST system requires a multi-million-dollar MRI machine, not a consumer gadget.

What are the ethical rules for this research?

Researchers follow standard human-subjects ethics: informed consent, data anonymization, and institutional review. But there is no specific law for neural data in most countries. Scientists are calling for new regulations to protect mental privacy as the technology advances.

Method: the standard formula for this calculation, rounded as each result states. See how the calculators work or report an error.