Meta AI unveils Brain2Qwerty system that decodes typed text from brain activity without implants

Meta AI researchers have developed Brain2Qwerty, a three-stage deep neural network that decodes typed text from non-invasive brain activity recordings, achieving a 29% average character error rate with MEG and 65% with EEG. The system, trained on data from 35 participants typing memorized sentences, represents a significant advance over previous methods, with the best MEG participant reaching an 18% character error rate. Published in Nature Neuroscience on June 29, 2026, the breakthrough offers a non-invasive alternative to surgical brain-computer interfaces that currently enable paralyzed patients to communicate but carry risks of hemorrhage and infection. The technology demonstrates that magnetoencephalography significantly outperforms EEG for brain-to-text decoding, achieving 2.5-fold improvement over existing models.

Brain2Qwerty noninvasive BCI system

• ▪Brain2Qwerty includes three core stages: a convolutional module input with 500 millisecond windows of MEG or EEG signals, a transformer module trained at the sentence level, and a pretrained language model to correct the outputs of the transformer
• ▪Brain2Qwerty is a three-stage deep neural network trained to decode text from brain signals
• ▪The article about Brain2Qwerty was published on June 29, 2026
• ▪Meta AI's Brain2Qwerty is an AI model trained to decode text production from noninvasive recordings of brain activity
• ▪Brain2Qwerty model inputs are defined as 0.5-second time windows extracted from −0.2 seconds to +0.3 seconds relative to each keystroke

MEG versus EEG performance comparison

• ▪Magnetoencephalography has a higher signal-to-noise ratio than EEG
• ▪The best MEG participant using Brain2Qwerty reached a character error rate as low as 18%
• ▪Character accuracy peaks at 22% for MEG and 16% for EEG in linear classification tasks, both significantly above the 14% chance level
• ▪MEG significantly outperforms EEG in Brain2Qwerty decoding performance with a P-value of 1.0 × 10−7
• ▪A public brain-computer interface benchmark using EEG achieves an accuracy of only 43.3% on a four-class classification task with a motor imagery dataset
• ▪Brain2Qwerty achieves an average character error rate of 29% with MEG compared to 65% with EEG
• ▪The 40-millisecond latency for peak classification accuracy in left versus right hand decoding is consistent with the physiological convergence of the efferent motor command and the afferent somatosensory feedback in EEG and MEG
• ▪MEG achieves a peak accuracy of 74% for left versus right hand classification, significantly outperforming EEG which achieves 64% with a P-value of 8.9 × 10−7

Deep learning decoder architecture

• ▪EEGNet outperforms linear models on both hand error rate and character error rate for both MEG and EEG with P-values less than 10−5
• ▪Linear ridge classifiers trained to categorize left versus right-handed responses achieve peak classification accuracy at 40 milliseconds after the keystroke
• ▪Brain2Qwerty achieves a 1.2-fold improvement in character error rate compared to EEGNet with EEG recordings with a P-value of 1.9 × 10−6

Model training on typing tasks

• ▪Brain2Qwerty was evaluated on EEG recordings from 20 participants comprising 146,000 characters, 23,000 words, and 4,000 sentences
• ▪Brain2Qwerty was evaluated on MEG recordings from 20 participants comprising 193,000 characters, 30,000 words, and 5,000 sentences
• ▪In the Brain2Qwerty typing protocol, sentences were displayed word-by-word on a screen, and following the final word, a visual cue prompted participants to begin typing the sentence without visual feedback
• ▪Brain2Qwerty was trained on data from 35 participants who typed briefly memorized sentences on a keyboard while their brain activity was recorded with either EEG or MEG