Morning Edition · Saturday, July 11, 2026Published at 1:46 AM EDT · New York

Meta's Brain2Qwerty Decodes Typed Sentences From Brain Scans at 61% Word Accuracy, Without Surgery

The non-invasive pipeline reads magnetoencephalography signals as volunteers type, up from about 8% word accuracy for prior non-surgical methods, though it still requires a room-sized scanner.

Save

Meta's Brain2Qwerty Decodes Typed Sentences From Brain Scans at 61% Word Accuracy, Without Surgery

Meta's research team detailed Brain2Qwerty, a deep learning system that decodes sentences from non-invasive brain recordings while participants type memorized text on a QWERTY keyboard. The architecture, described in the accompanying paper, stacks a convolutional module over 500-millisecond windows of magnetoencephalography (MEG) or electroencephalography (EEG) signal, a sentence-level transformer, and a pretrained language model that corrects the transformer's output.

The reported numbers mark a real step for surgery-free decoding. The first version reached a 32% character error rate with MEG and 19% for the best participant, against a 67% error rate for cheaper EEG. The second version reaches 61% average word accuracy, up from roughly 8% for prior non-invasive methods, with the best participant at 78% and more than half of sentences decoded with one word error or fewer, across 35 healthy volunteers.

The trade is fidelity for accessibility, and the constraint is hardware. MEG requires a shielded room and a magnetometer array, not a wearable, and the cheaper EEG modality still performs far worse. This is a laboratory result about what non-invasive signals can carry, not a product for patients who have lost the ability to speak.

Weighed against invasive brain-computer interfaces, which achieve higher accuracy through implanted electrodes and the risks of neurosurgery, Brain2Qwerty takes the opposite position: no operation, lower accuracy, and a dependence on immobile scanning equipment.

What this means

The advance is in the machine learning stack, not the sensor, which is what makes it relevant to engineers rather than only to neurosurgeons: a convolution-transformer-language-model pipeline extracts far more from the same noisy MEG signal than earlier decoders. That favors labs with large model-training capacity over pure hardware groups, and it keeps the accessibility question tied to scanner size. Whoever shrinks MEG-grade sensing toward a wearable, not whoever tunes the model further, decides whether this reaches assistive use.

What to watch

Any move from shielded-room MEG toward portable optically-pumped magnetometers, since sensor portability, not model accuracy, is the binding constraint on real-world use.
Whether accuracy holds for imagined or attempted typing rather than actual keystrokes, the harder task that matters for people who cannot move.

Observations to monitor, not financial advice.

2 sources

Synthesized from: Meta AI (Brain2Qwerty) · arXiv 2502.17480

Part of a tracked trend

Non-Invasive Neural Decoding

AI labs increasingly apply machine learning to decode language from non-invasive brain signals, trading fidelity for accessibility and pushing neurotechnology toward broader assistive use.