Meta's Brain2Qwerty v2 Reads Text from Brain Activity Without Implants
AI at Meta has announced Brain2Qwerty v2, the second iteration of its non-invasive brain-to-text pipeline that reconstructs typed text directly from brain activity. While v1 (just published in Nature) decoded individual keystrokes from magnetoencephalography (MEG) signals, v2 makes a qualitative leap—decoding the semantics of entire phrases in real-time, end-to-end from raw signal, without character-by-character guessing.
The main limitation persists: high accuracy requires a bulky, expensive MEG scanner (shielded room, still head), not a wearable device. With low-cost EEG the character error rate in v1 was 67%, making it nearly unusable. Meta has not yet disclosed v2’s accuracy, typing speed, vocabulary size, or whether it still depends on MEG—but even an imperfect non-invasive communication channel could be transformative for people with neurological disorders who have lost the ability to speak.