Meta’s Fundamental AI Research (FAIR) group released version two of its Brain2Qwerty system on June 29, demonstrating the highest accuracy ever achieved for a non-invasive brain-to-text decoder. The system translates magnetoencephalography (MEG) recordings into full sentences, no surgery, no implants.
How it works
Volunteers wear a helmet of magnetic sensors that measure the faint magnetic fields produced by neuronal electrical activity. As they type, the system reconstructs the intended words by mapping distributed neural activity patterns to text. The upgrade from v1, which decoded individual characters, represents a leap in ambition: v2 processes continuous MEG recordings and reconstructs entire sentences.
The system was trained on approximately 22,000 typed sentences collected from nine volunteers, each spending roughly 10 hours inside an MEG scanner.
Accuracy numbers
Brain2Qwerty v2 achieves 61 percent average word accuracy across all participants, with the best performer reaching 78 percent. At the character level, the word error rate averages 39 percent, dropping to 22 percent for the top performer.
These figures are modest compared to near-perfect speech-to-text systems or invasive brain-computer interfaces, Neuralink’s N1 implant has demonstrated character error rates below 5 percent, but they represent a breakthrough for a fully non-invasive approach that requires nothing more than a head-mounted sensor array.
Open science, not proprietary
Meta is releasing the system’s code and training dataset openly on GitHub at facebookresearch/brain2qwerty, alongside a $5 million (approximately 4 million pounds) fund for open neuroscience research. The work was simultaneously published in Nature Neuroscience.
The open-source approach stands in contrast to Meta’s broader pivot toward proprietary AI products exemplified by the Muse Spark product line. FAIR’s continued commitment to open neuroscience research suggests the company sees strategic value in maintaining an independent research identity.
The non-invasive advantage
The target population for Brain2Qwerty is people who have lost the ability to speak due to neurological conditions such as ALS, stroke, or locked-in syndrome. For these individuals, even 61 percent word accuracy represents a communication channel where none existed before.
MEG technology still has major practical limitations: the scanners are room-sized machines costing millions of dollars, requiring magnetically shielded environments. Portable deployment remains years away.
Nevertheless, the v2 results shift the conversation around brain-computer interfaces. The field has long assumed that high-fidelity neural decoding requires breaching the skull. Brain2Qwerty v2 suggests that algorithmic improvements can narrow the gap without crossing the surgical threshold, potentially making the technology accessible to millions rather than the few willing to undergo invasive procedures.