Meta has released Brain2Qwerty v2, an artificial intelligence system that non-invasively converts brain recordings into text in real time, without the need for surgical implants.
The system achieved an average word accuracy of 61%, compared to about 8% for previous non-invasive methods. The strongest participant’s accuracy reached 78%, with more than half of his sentences deciphered but only one wrong word.
Mehta trained his model on about 22,000 sentences collected from nine healthy volunteers. Each participant spent about 10 hours typing sentences inside the magnetoencephalogram scanner, giving the model about 10 times more training data per person than previous versions.
Brain2Qwerty v2 processes continuous brain activity through separate components designed to detect letters, align words, and reconstruct complete sentences. Meta also fine-tuned a large-scale language model on neural data so that the system can use semantic context when interpreting noisy signals.
The new version removes a major limitation of the original system, which required researchers to know in advance when to press each key. Brain2Qwerty v2 instead generates sentences directly from continuous recordings, allowing the pipeline to work in real time.
This study does not represent open-ended mind reading. Participants were healthy volunteers who actively entered memorized sentences, and the system relied on large MEG scanners, which remained impractical in most clinical settings. Mehta also acknowledged that current error rates are still too high for everyday communications.
Mehta said performance improved as the amount of training data increased, but there was no clear plateau. The company believes that additional data and advances in wearable MEG sensors can close the remaining gap to brain-computer interfaces that require surgical implants.
The company has released the complete training code for both versions of Brain2Qwerty, and the Basque Center for Cognition, Brain and Language has published the datasets used to develop the first version. Data for the second version will not be available while the journal is pending publication.
