AI Tool Reconstructs Images From Brain Scans

A brain scanner watches you look at a photograph of a face. Minutes later, a computer program that has never seen the photograph draws its own version—close enough that you recognize the original at a glance. That scenario is no longer hypothetical. Researchers at the Weizmann Institute of Science in Rehovot, Israel, have built an AI system that reconstructs what a person is viewing from their brain activity alone, and it works in both directions: it can also predict a person's neural responses from an image they are looking at.

The work, reported by MIT Technology Review, sits at the intersection of neuroscience, machine learning, and ethics. It raises the prospect of new tools for people with severe neurological conditions—and fresh concerns about how far neural data can be decoded without consent.

What Is the AI Mind-Reading Tool and How Does It Work

Start with the mechanics of a single trial. A volunteer views an image—in the published example, a face—while their brain activity is recorded. The model then analyzes that scan and generates a reconstruction of the image. Side by side, the pairs tell the story: the left-hand image is what the person actually saw, and the right-hand image is what the model produced from the brain scan. The reconstructions are described as remarkably precise.

The system is bidirectional. Beyond decoding a scan into an image, it can work in reverse, predicting a person's brain activity based on what they are looking at. That two-way capability matters scientifically. A model that can both read neural activity and forecast it demonstrates a deeper internal representation of how visual information maps onto brain signals—not just a pattern-matching trick in one direction.

This kind of neural decoding has a track record. Peer-reviewed fMRI-based image reconstruction has been an active research area since at least 2022, when brain-scan decoding methods began producing recognizable—if blurry—reconstructions of viewed images. The field has advanced steadily since, with generative AI models improving the fidelity of reconstructions. What distinguishes the Weizmann tool, per the report, is the precision of its output and its ability to operate in both directions.

Terminology matters here. Researchers generally prefer "neural decoding" or "brain decoding" to "mind reading," because the technology reconstructs visual content from measured brain signals rather than accessing thoughts directly. The tool does not retrieve a verbal narrative of what someone is thinking. It reconstructs visual imagery from patterns of brain activity.

The Research Team Behind the Breakthrough

The tool was developed by Michal Irani and her colleagues at the Weizmann Institute of Science. Irani's involvement signals the project's dual nature: it draws on computer vision and machine learning expertise as much as on neuroscience. The Weizmann Institute is one of the world's leading basic-research institutions, with a long record in neuroscience, computer science, and the physical sciences—a pedigree that lends the work substantial credibility in the scientific community.

Irani's stated motivation is not surveillance or consumer technology. She hopes the tool will ultimately reveal more about how the brain works. That framing places the research in the tradition of basic science: understand the system first, then consider applications.

The team's ambitions, as described, extend to some of the hardest problems in neuroscience. One goal is helping people who are locked in—conscious but unable to move or communicate—by offering a potential channel for expressing what they are seeing or imagining. Another is reconstructing the content of dreams, which would give researchers a window into visual experience that occurs entirely offline from the external world.

Both goals are explicitly framed as hopes for what the technology could eventually do, not as capabilities it has demonstrated. The published work concerns reconstructing viewed images from brain scans. Dream decoding and communication for locked-in patients remain aspirational.

Potential Medical and Scientific Applications

For patients with severe neurological conditions, the implications are concrete enough to take seriously. Locked-in syndrome—most famously associated with patients after certain brainstem strokes—leaves people aware and cognitively intact but largely or entirely unable to move. Existing communication aids depend on residual movement, such as eye blinks or small eye movements. A system that could decode visual imagery directly from brain activity might offer a channel that requires no movement at all.

The scientific value may be just as significant. If a model can predict brain activity from an image and reconstruct an image from brain activity, it becomes a tool for probing how the visual system encodes information. Researchers could use it to test hypotheses about which brain regions carry which kinds of information, how abstract or detailed visual representations are, and how imagery differs from perception.

The dream angle is more speculative but scientifically tantalizing. Dreams are visual experiences generated without external input, and studying them has always been methodologically difficult—subjects must wake and report, and reports are incomplete and distorted by memory. A decoding approach that could reconstruct dream imagery from brain activity during sleep would give scientists a direct handle on one of the least accessible corners of human experience. Irani's team frames this as a hope, not a result.

Expert Reactions: Excitement and Caution

Judy Illes, a neuroethicist and professor of neurology at the University of British Columbia in Canada, who was not involved in the research, called the work "magnificent." Her enthusiasm centers on therapeutic potential. The idea of using this approach to help people with neurologic conditions, she said, is "tremendously exciting."

That endorsement carries weight. Illes is one of the field's most established voices on the ethical dimensions of neurotechnology—someone who typically pairs excitement with scrutiny. Her reaction suggests the scientific achievement itself is not in dispute.

Not everyone is focused on the upside. Other scientists warn that a similar approach could be used to reveal people's inner thoughts and mental imagery, potentially without their consent. That caution is not abstract. Neural decoding research has advanced to the point where the technical barrier to reconstructing visual experience is falling, while the legal and ethical frameworks governing brain data remain thin.

The disagreement is less about whether the technology works than about what should be done with it—and who gets to decide.

The Ethical Debate Around Neural Decoding Technology

Consent is the crux. Traditional privacy protections govern what people say and do. Brain data is different: it can, in principle, reveal mental content that a person has not chosen to externalize and may not even be aware of. If a scan can reconstruct what someone is looking at, the question of whether that person agreed to the decoding—and understood what they were agreeing to—becomes urgent.

The risk profile scales with the technology. Today's systems typically require a cooperative subject, often in a controlled experimental setting, and specialized equipment such as fMRI. As the science matures and hardware improves, the gap between "research subject" and "person going about their day" could narrow.

Therapies raise the same questions in a gentler form. A communication tool for a locked-in patient is unambiguously beneficial. But continuous neural monitoring in a clinical setting generates vast amounts of data, and the rules for storing, sharing, and analyzing it are still being written. Illes's field exists precisely because these questions arrive before the answers do.

There is also a broader cultural dimension. The phrase "mind reading" invites both hype and fear, and neither serves public understanding well. What the technology does—decode visual content from brain signals—is remarkable without being magical. Precise language helps regulators and the public assess actual risks rather than imagined ones.

What This Means for the Future of Brain-Computer Interfaces

Brain-computer interfaces have moved from laboratory curiosity to clinical reality over the past decade, with implanted devices restoring some communication and control to people with paralysis. The Weizmann tool points toward a different branch of the field: non-invasive decoding of rich visual content from brain scans, without implants.

The trajectory suggests three near-term developments. First, reconstructions will likely continue to improve as generative models get better at producing naturalistic images from limited neural data. Second, the bidirectionality demonstrated here—predicting brain activity from images and images from brain activity—will become a standard test of how well a model captures the brain's internal code. Third, the ethical conversation will accelerate, because the capability is arriving faster than the governance.

What the tool cannot currently do is just as important as what it can. It reconstructs viewed images; it does not read propositional thoughts, intentions, or memories in verbal form. Dream reconstruction and communication for locked-in patients remain goals, not deliverables. The research community's task now is to pursue those goals while building the consent frameworks and data protections that a technology this intimate will require.

Irani's stated hope—that the tool reveals more about how the brain works—is the most defensible frame for the work. Understanding precedes application, and in neural decoding, understanding also precedes regulation. The science is magnificent, as Illes says. The governance is not yet there.

Source: MIT Technology Review