Meta AI has released Brain2Qwerty v2, an end-to-end AI pipeline that converts magnetoencephalography (MEG) brain recordings into written text without requiring surgical implants.
Published June 29, 2026, the system achieves 61% average word accuracy across nine participants — a dramatic improvement over the roughly 8% accuracy typical of other non-invasive brain-computer interface methods, and approaching performance levels previously attainable only with electrode arrays implanted directly in the brain.
How it works
Participants wore MEG headsets — which measure the tiny magnetic fields produced by neural activity — and typed sentences while their brain signals were recorded. Meta trained an end-to-end deep-learning model on roughly 22,000 sentences per participant, with each of the nine volunteers logging about 10 hours of recording time.
Rather than relying on hand-crafted signal features, the model learns directly from raw brain data. Large language models fine-tuned on the neural recordings help bridge the gap between noisy MEG signals and coherent text.
What the numbers mean
The best-performing participant reached 78% word accuracy. Meta’s researchers say accuracy improves log-linearly with data volume, suggesting further gains are plausible as more training data is collected.
The work was conducted in collaboration with the Basque Center on Cognition, Brain, and Language (BCBL), which also released the v1 dataset publicly. Meta has made training code for both v1 and v2 open to the research community.
Why it matters
Brain-computer interfaces typically divide into two camps: high-accuracy systems that require invasive surgery, and non-invasive approaches that sacrifice most of that accuracy. Brain2Qwerty v2 narrows that gap meaningfully. According to Meta, the technology could eventually help people with communication disabilities caused by brain lesions or neurodegenerative conditions, giving them a pathway to text-based communication without an operation.
No consumer product timeline was announced. The nine-participant study is small, and real-world accuracy may vary — but the paper argues credibly that scaling training data further could push non-invasive decoding performance higher still.