Meta's new AI can decode images directly from brain activity in real time

Sci-fi movies love the trope of a helmet that projects whatever a character is thinking directly onto a screen. Meta just took a real-world step in that direction.

Researchers from Meta AI unveiled an artificial intelligence system capable of decoding human brain activity in real time. Using magnetoencephalography (MEG), a non-invasive neuroimaging technique, the system can reconstruct the images a person is seeing with millisecond-level precision.

A three-step pipeline from brain waves to pictures

Most prior breakthroughs in brain decoding leaned heavily on functional Magnetic Resonance Imaging (fMRI). While fMRI captures crisp spatial details, it is painfully slow, measuring blood flow changes over several seconds. MEG, on the other hand, records thousands of magnetic brain measurements every single second, effectively acting as an ultra-high-speed camera for neural pulses.

Meta split the decoding pipeline into three separate components: an image encoder, a brain encoder, and an image decoder.

The image encoder first creates detailed visual representations of an image independently of the human brain. Next, the brain encoder learns to map raw MEG signals directly to those visual representations. Finally, the image decoder takes those reconstructed representations and generates a plausible picture.

The team trained the system using a public MEG dataset released by the Things academic research consortium. Along the way, Meta made a fascinating discovery: the brain signals aligned best with modern self-supervised vision models, specifically Meta's own DINOv2.

Because DINOv2 learns to understand images without human annotations, its artificial neurons ended up activating in patterns remarkably similar to biological neurons in the human visual cortex.

High on concepts, shaky on the small stuff

While the system is impressive, it is not quite a full-blown mind reader yet.

Meta noted that the system is exceptionally reliable at capturing high-level features. If you are looking at a dog, a car, or a person, the AI almost always gets the general category right.

Where it stumbles is low-level details. The decoder often misplaces objects, flips orientations, or generates the wrong specific animal breed. When compared against fMRI studies using the Natural Scene Dataset, the spatial fidelity of MEG-generated images is clearly rougher around the edges.

However, what MEG sacrifices in pixel-perfect sharpness, it makes up for in pure speed. Because it captures brain activity continuously, it can produce an ongoing flux of decoded images that mirrors human perception as it happens.

Why it matters

We think Meta's smartest move here was bypassing the race for static pixel perfection and betting entirely on temporal speed.

For patients and clinicians, this points toward truly practical, non-invasive brain-computer interfaces. Invasive implants like Neuralink require risky brain surgery. If non-invasive hardware can evolve alongside lightweight MEG sensors, people who have lost the ability to speak or move due to brain injuries could one day communicate their thoughts and visual perceptions in real time, without undergoing surgery.

For the broader AI industry, this work provides concrete evidence that self-supervised architectures are on the right track. By proving that DINOv2 spontaneously develops brain-like representations, Meta gives serious scientific backing to Yann LeCun's vision of autonomous machine intelligence. Models that learn the world abstractly rather than memorizing pixels might actually be our best shot at building human-like reasoning.

The clear limitation, of course, is generalization. The current system was trained on controlled datasets where participants viewed predetermined images. Decoding a spontaneous, unprompted human thought or daydream is an entirely different beast, and turning this lab demo into a versatile clinical device remains a massive hurdle.