Brain2Qwerty v2

Nine volunteers sat in a magnetoencephalography scanner and typed 22,000 sentences, about ten hours each. From the magnetic traces alone — no implant, no electrode, nothing under the skin — Meta's Brain2Qwerty v2 reconstructs what they typed at a 39% average word error rate, meaning roughly three of every five words come back correct. For the best participant, over half the held-out sentences come back with one word wrong or fewer. No surgery-free system had reported a real sentence-level score before.

Adapting to patients who cannot generate labeled keystroke data the way healthy typists do is the critical open problem. — AI at Meta

Surgical implants still win outright: the best intracranial devices sit near 2% error, and this trails them badly. What matters is the shape of the curve. As Meta pooled more recording hours, accuracy kept improving in a straight log-linear line with no sign of flattening at ninety hours — which lets the authors argue the gap to implants is partly a data problem, not a physics ceiling. Buried in the ablations is the more surprising finding: one model trained on all nine people pooled beats a bespoke model fit to each person (a 47.8% vs 66.5% median error rate), the opposite of the field's usual per-person calibration instinct, hinting that the neural signature of typing intent has a common structure across brains.

The asterisk is the size of a room. MEG needs a magnetically shielded chamber and a subject who holds still, so this is a lab curve, not a device — and the volunteers are all healthy, proficient typists. The people a no-surgery reader would actually serve, ALS or locked-in patients, can't produce the labeled keystrokes the model trains on, and closing that gap is the unsolved problem the paper names for itself. The scaling line is the real news: it says the ceiling for reading language off the scalp is higher than anyone had shown, and reachable with scanner time rather than a scalpel.