News · 2026-06-30

Meta reads full sentences from brain waves - without surgery

Meta's AI research group has built a non-surgical brain-reading system that recovers typed sentences with about 61% of words correct, up from roughly 8% for prior non-surgical methods — closing most of the accuracy gap with approaches that require implanted electrodes. The system, called Brain2Qwerty v2, uses magnetoencephalography to decode brain activity into text in real time, and Meta has released the training code and data for other researchers to build on.

Key facts

• What: A new version of Meta's brain-to-text system decodes typed sentences from magnetic brain signals far more accurately than before, closing much of the gap with implanted electrodes.
• When: 2026-06-30
• Primary source: read the source

For people who have lost the ability to speak or type — through injury, stroke, or diseases like ALS — a system that turns thought into text could restore something close to normal communication. The most accurate approaches so far require brain surgery to place electrodes directly on or in the brain tissue. That gives a clean signal, but it is a major operation with real risks, and it will only ever reach a small number of people. The goal has been to get comparable accuracy from outside the skull, with no surgery at all.

Brain2Qwerty v2 narrows that gap. It uses magnetoencephalography — a scanner that picks up the faint magnetic fields generated by the brain's electrical activity. As a person types, the system captures those magnetic signals and an AI model translates the patterns into the actual text being typed. The result: it recovers sentences coherently with about sixty-one percent of words correct. Prior non-surgical methods managed around eight percent — barely better than guessing and nowhere near usable. Going from eight percent to sixty-one percent is the difference between noise and something you could almost hold a conversation through. Meta says the pipeline works end to end and can decode sentences in real time, and the company released the training code and data so other researchers can build on it. The work is described in a post from Meta AI.

The AI model's task is to learn the statistical link between those chaotic magnetic patterns and the letters a person intends — the kind of find-the-signal-in-the-noise pattern-matching that modern neural networks excel at. Raw brain signals are extraordinarily messy: faint, noisy, and different from person to person and moment to moment. The system is not reading thoughts in any general sense; it is decoding the specific, physical brain activity that accompanies the motor act of typing, and mapping it back to characters.

The essential caveat — which Meta is upfront about — is that the magnetic scanner that makes this work is room-sized, specialized laboratory equipment. It is not a headband, not a wearable, and not anything you could use at home or carry around. This is a research milestone about what is possible with non-surgical brain reading, not a product on the way to market. The value is in the proof: it shows you can get near-implant accuracy without cutting into the brain, which reframes what the goal even is. If the accuracy can be preserved as the hardware shrinks — a very big if, and likely years of work — it points toward a future where restoring communication does not require surgery. For now, the honest framing is a lab result that dramatically raised the ceiling on what reading the brain from the outside can achieve, while leaving the hard problem of doing it with practical, affordable equipment wide open. Even bounded that way, closing most of the gap to invasive methods is the kind of step that changes what researchers dare to aim for.

Primary source, verified: read the paper →

Topics: Meta · neuroscience · brain-computer-interface · accessibility · research