Meta built AI that turns brain activity into text, no chip required

Signals Inbox·July 8, 2026·NeuroTech

Meta says Brain2Qwerty v2 can decode typed sentences from brain activity in real time, without an implant. So it looks like non-invasive brain-to-text is moving from messy lab demos toward readable sentences, but still inside a very expensive MEG setup.

Q1What actually happened?

Meta announced Brain2Qwerty v2, a system that turns brain activity into typed sentences using MEG, a scanner that reads tiny magnetic signals from the brain. No electrodes go into the brain. No skull is opened. The person types sentences, the scanner records the brain activity, and the AI tries to reconstruct the text.

Q2What changed from v1?

V1 was already strong for non-invasive decoding, but it was more like a careful research pipeline. V2 is the bigger step because it decodes continuous brain recordings and tries to produce sentences in real time. Meta says it uses character, word, and sentence-level modules together, instead of only guessing one key at a time.

Q3How good is it?

Good for a lab system, not good enough for daily use. The v2 paper reports about 39% word error rate on average, which means roughly 61% word accuracy. The best participant did better, but this is still far from something you would trust to write messages for a locked-in patient without heavy correction.

Q4Why does the comparison matter?

Because invasive brain implants already work better for some patients. Neuralink, Synchron, Blackrock, and academic teams can record closer to the neurons, so the signal is cleaner. Meta is taking the opposite bet: accept a weaker signal, but avoid surgery.

Q5Is this really mind reading?

Not really. The model is not reading random private thoughts. It is decoding brain activity while people are typing or preparing typed sentences in a controlled experiment. That is a big difference. It is closer to reading the brain’s typing process than opening someone’s inner monologue.

Q6What is the biggest catch?

The hardware. MEG is not a cheap headset. It usually needs a large scanner, controlled conditions, and serious lab infrastructure.

Q7Why does this matter now?

Because the field is splitting into two paths. One path is invasive implants with better signal and higher medical risk. The other is non-invasive decoding with safer access but much noisier data. Meta just showed that the second path is not stuck at toy-level demos anymore, even if it is still far from a product.