Brain2Qwerty: Meta's Non-Invasive Path From Brain Signals to Text

Meta's AI division has unveiled Brain2Qwerty, a deep-learning system that reconstructs typed sentences from recorded brain activity without any implanted hardware, positioning magnetoencephalography-based decoding as a potential communication channel for patients who cannot speak.

The name describes the mechanism. A participant simply attempts to type sentences on a standard QWERTY keyboard while brain activity is recorded, and the model learns to infer which key was pressed — or intended — from the neural signals alone. Once trained, the system can generate text from brain activity in a person who makes no physical movement.

This matters for clinical neurophysiology because it changes the recording conversation. Most high-performance brain-to-text systems demonstrated to date rely on intracranial arrays — electrocorticography (ECoG) grids or depth electrodes placed during epilepsy monitoring or in dedicated BCI trials. Those devices deliver spatial resolution but require surgery. Brain2Qwerty works with non-invasive recording, chiefly magnetoencephalography (MEG), which captures the millisecond-scale dynamics of cortical activity that scalp EEG partly attenuates and blurs spatially.

What does the system actually do?

According to the announcement framing the release, Brain2Qwerty was evaluated on participants typing sentences while their brain activity was recorded with MEG. The pipeline has three stages:

• A neural network trained on brain recordings and the corresponding keystrokes
• A language model that scores and corrects the character-by-character output
• An evaluation measuring character and word error rates between decoded and typed sentences

The language-model component is what pushes the error rate down. Even when the raw neural decoder is uncertain about individual characters, the language model exploits the statistics of English to recover the intended word — the same principle that powers autocorrect on a phone, applied to a neural signal instead of a thumb.

Why MEG and not scalp EEG?

The choice of modality is the technical crux. Scalp EEG remains the workhorse of clinical neurophysiology, but volume conduction and skull attenuation smear the signal, making single-character decoding at a keyboard's spatial granularity a hard problem. MEG records magnetic fields generated by cortical currents, preserving temporal resolution in the millisecond range with better spatial discrimination than scalp EEG, though still far below ECoG.

For reading-room readers, the practical takeaway is this: non-invasive decoding of fine motor intentions — individual finger movements mapped to specific keys — is now demonstrably feasible in a laboratory MEG setting. It has not been demonstrated in an ambulatory, bedside, or home environment, and MEG hardware is bulky, expensive, and magnetically shielded. No Brain2Qwerty device is FDA-cleared; this is investigational research in healthy participants, not a clinical communication aid and not validated in patients with aphasia, locked-in syndrome, or amyotrophic lateral sclerosis.

How does it compare with implanted decoders?

Intracranial systems from academic groups and companies have produced faster and more accurate typing in participants with electrodes placed for epilepsy monitoring or implanted for chronic BCI use. Brain2Qwerty accepts lower performance in exchange for zero surgical risk — the trade that would matter most for a patient population deciding between a craniotomy and a helmet.

The claim from Meta is therefore narrow and should stay narrow: a trained neural network, given MEG recordings of people typing, can reconstruct the sentences they typed without any implant. Claims about restoring communication to non-verbal patients remain prospective.

What comes next?

Meta indicates that improving non-invasive decoding accuracy and moving toward more practical recording setups are the open problems. For clinical neurophysiologists, the work signals a future in which MEG — and eventually perhaps high-density EEG — protocols designed for epilepsy and presurgical evaluation could double as acquisition platforms for communication decoding research, provided laboratories begin banking the temporally rich, movement-annotated datasets such models require.