.png){kind=logo}
AI models analyse fMRI patterns and use visual features to reconstruct images that broadly reflect what participants see.
Generative models improve image quality, while newer systems explore reconstruction of imagined or remembered visuals under controlled experimental conditions and carefully designed studies.
Applications include neuroscience and assistive communication, but accuracy, training demands, individual variation, computational requirements and mental privacy remain significant challenges
For decades, scientists have tried to understand how the brain turns electrical activity into the images we see. Artificial intelligence is now giving researchers a new way to study that process: by using brain scans to reconstruct visual information.
Recent advances show that AI models can translate patterns of brain activity into images that capture important elements of what a person is looking at. The technology is not a literal window into the mind, but it is becoming a powerful tool for studying how the brain represents objects, scenes and meaning. In September 2026, researchers at the Weizmann Institute of Science unveiled Brain-IT, an AI system designed to reconstruct viewed images from brain activity while requiring substantially less training for a new individual than earlier approaches.
One of the main technologies behind this research is functional magnetic resonance imaging, or fMRI. Unlike a conventional MRI, which primarily shows anatomy, fMRI measures changes in blood oxygenation associated with neural activity. Researchers can therefore observe which parts of the brain become more active while a person looks at an image.
An fMRI scan does not produce a photograph of a thought. Instead, it generates complex patterns across thousands of small three-dimensional units called voxels. AI models can learn relationships between these patterns and the visual information presented to participants.
Researchers train a decoder by repeatedly pairing brain scans with known images. Machine-learning systems then identify statistical associations between neural activity and features such as colour, shapes, spatial arrangement, objects and broader semantic meaning. Once trained, the model receives a new brain scan and predicts the visual features represented by that activity.
Generative AI has significantly changed this process. Earlier systems often relied on approaches such as generative adversarial networks and variational autoencoders. Newer methods increasingly map brain signals into latent representations — compact mathematical descriptions of visual information — before feeding them into powerful image-generation models.
Diffusion models are particularly important. Rather than simply retrieving an image from a database, they can generate a new image guided by information predicted from brain activity. Research published in the Journal of Big Data in 2026 describes this shift toward multi-stage systems that combine neural decoding with representations such as CLIP and VDVAE before using diffusion-based generation.
The Weizmann team’s Brain-IT takes another approach to a major problem: limited training data. Its system uses an encoder to predict brain activity from images and a decoder to reconstruct images from brain activity. This helps create additional training examples without requiring every image to be physically shown to someone inside an MRI scanner.
A crucial distinction is whether the brain is responding to something a person is actually seeing or to an internally generated image.
Reconstructing a viewed image is currently the more established task. The system has a clear external stimulus and can learn the relationship between that stimulus and the resulting brain activity.
Decoding mental imagery is considerably harder. When people imagine an object, their brain activity can be weaker, noisier and less consistent than when they actually see it. A 2026 PLOS Computational Biology study found that models performing well on viewed-image reconstruction do not necessarily perform equally well on mental imagery. Its MIRAGE system used multimodal image and text features with a diffusion model to improve reconstruction of imagined images.
Also Read: AI Borrowing Boom: USD 80B Debt Wave Shakes Global Markets
The technology could eventually contribute to brain-computer interfaces, neuroscience research and assistive communication. For people with severe paralysis or communication disorders, systems capable of decoding meaningful visual or conceptual information could offer another channel for expressing themselves.
Researchers could also use reconstruction models to investigate how different brain regions process visual features and meaning. Recent work has shown that AI representations can help model high-level information encoded in visual brain regions, offering new ways to examine the relationship between biological and artificial intelligence systems.
Despite impressive demonstrations, reconstructed images are interpretations, not recordings of thoughts. AI systems can introduce details that were not present in the original stimulus because generative models are designed to produce plausible images.
Performance also varies between individuals. Brain anatomy and neural responses differ, while fMRI itself has limited temporal resolution. Large, carefully labelled datasets are expensive to collect, and training sophisticated models requires substantial computational resources. Researchers are also still testing how reliably models generalise to unfamiliar images and people. A 2026 Nature Communications dataset study highlighted the importance of testing models on visual stimuli outside their training distribution.
As brain-decoding systems improve, ethical questions will become increasingly important. Neural data can reveal information about a person’s responses and potentially aspects of their internal experiences. That raises questions about informed consent, ownership of brain data, and who should be allowed to analyse it.
For now, technology cannot freely extract a person’s thoughts. Most systems require controlled experiments, specialised scanners and substantial model training. But the possibility of increasingly sophisticated neural decoding makes mental privacy an issue that researchers, technology companies and policymakers will need to address.
AI-driven brain reconstruction therefore represents an important scientific step, not the arrival of literal mind reading. By linking patterns of neural activity with visual representations, researchers are gaining a clearer picture of how the brain processes what we see — and how artificial intelligence can help reveal that hidden process.
Also Read: Mecka AI Raises USD 60 Million as Humanoid Robot Race Fuels Data Demand
AI-based brain image reconstruction uses fMRI activity patterns to predict visual information, allowing machine-learning models to generate images that broadly represent what participants see or imagine.
fMRI measures blood-oxygen changes linked to neural activity. AI models analyse voxel patterns and learn associations between brain activity and visual features, including shapes, colours, objects, and meanings.
Yes, researchers are exploring imagined-image reconstruction, but it remains more challenging than decoding viewed images because mental imagery produces weaker, noisier, and less consistent brain signals.
Potential applications include neuroscience research, brain-computer interfaces, assistive communication, and studying visual processing. The technology could eventually provide alternative communication channels for people with severe disabilities.
No. Current systems require controlled experiments, specialised fMRI scanners, training data, and computational models. Reconstructed images are interpretations and may contain generated details absent from original thoughts.
