BCIwiki (bciwiki.com) — A deep learning model called Brain2Qwerty decoded sentences that people were typing from noninvasive recordings of brain activity, reaching an average character error rate of 29% with magnetoencephalography (MEG) in 35 healthy volunteers, researchers reported in Nature Neuroscience on June 30, 2026. Authors are affiliated with Meta AI in Paris, École Normale Supérieure, Université PSL, CNRS, the Hospital Foundation Adolphe de Rothschild and Univ. Lille's SCALab.
Participants briefly memorized sentences and then typed them on a QWERTY keyboard while their brain activity was recorded with either electroencephalography (EEG) or MEG. MEG substantially outperformed EEG, which yielded an average character error rate of 65%. For the best participants, the error rate fell to 18%, and the model perfectly decoded a variety of sentences outside its training set.
The authors note that implanted neuroprostheses can restore communication for patients who have lost the ability to speak or move, but carry the risks of neurosurgery. They say their results narrow the gap between invasive and noninvasive methods and open a path toward safe brain-computer interfaces (BCIs) for noncommunicating patients. All participants were healthy volunteers; the approach has not been tested in patients. The paper's PMID is 42374156 and its DOI is 10.1038/s41593-026-02303-2.