Meta AI has unveiled Brain2Qwerty, a system that reconstructs words from non-invasive brain activity recorded while a person types. The technology uses a helmet of sensors, requiring no implants or surgery.
In trials with 35 volunteers, the system achieved a 32% character error rate using magnetoencephalography (MEG) sensors. This far outperformed EEG-based readings, which had a 67% error rate. The top MEG performers reached a 19% error rate.
The system works as a three-stage pipeline: a convolutional module analyzes short brain signal windows, a transformer processes sentence-level patterns, and a language model refines the character output. However, it is not real-time, as it requires an entire sentence sequence before decoding.
The research team, including collaborators from Paris Sciences et Lettres University and the Basque Center on Cognition, Brain and Language, positions this as a proof-of-concept. It shows non-invasive recordings can reconstruct language, contrasting with invasive approaches like Neuralink that prioritize performance.
For commercialization, significant challenges remain in sensor miniaturization and real-time computation. The current performance is scientifically notable but not yet viable for consumer devices.