Meta uses AI to decode thoughts into text without brain implants

For years, getting a computer to read human thoughts meant taking the extreme route: opening up the skull and planting microelectrodes directly into brain tissue.

Meta just showed there might be a less invasive path. In two newly published research papers, the company demonstrated that artificial intelligence can decode brain signals into real-time text without a single surgical cut.

Listening to keystrokes inside the skull

To pull this off, researchers recorded neural activity using non-invasive magnetoencephalography (MEG) and electroencephalography (EEG) while 35 typists typed out sentences on a standard keyboard.

They used that neural data to train an AI model named brain2qwerty. Instead of trying to guess abstract thoughts out of thin air, the deep learning model focused on predicting the precise characters people were intending to type based purely on their brainwaves.

With MEG sensors, the model achieved an average character error rate of 32%. That means it accurately predicted the intended letter 7 to 8 times out of 10, hitting up to 80% character accuracy.

In the best-performing subject, the system managed to flawlessly decode complete sentences that were entirely outside the original training dataset.

A companion paper mapped out the hierarchy of neural dynamics that supports language production, explaining how the brain transitions from an abstract thought to physical motor execution.

Why it matters

Brain-computer interfaces have long been split into two camps: invasive hardware like Neuralink that offers high fidelity at surgical risk, and clunky consumer headsets that can barely tell if you are blinking.

Meta’s clever pivot here is anchoring language decoding to the physical act of typing. Typing generates structured, highly repeatable motor signals in the cortex, giving the AI far cleaner patterns to latch onto than ambiguous internal monologues.

For non-verbal patients and people suffering from paralysis or ALS, this brings non-invasive speech and text restoration much closer to reality without requiring high-risk brain surgery. Further down the road, it points to silent communication for operators, soldiers, or anyone needing hands-free stealth control.

That said, the technology is nowhere near everyday consumer use. MEG machines are still room-sized, expensive contraptions. There is also the obvious dystopian baggage: as online observers quickly pointed out, once non-invasive tech gets good enough to decode internal signals from the scalp, personal mental privacy becomes a very urgent legal and ethical battleground.