Meta’s FAIR lab just decoded speech from non-invasive brain recordings — jumping from 8% to 61% word accuracy. Here’s why that’s real science, and why the headline still oversells it.

What happened

Meta’s FAIR lab trained a deep learning pipeline on roughly 22,000 sentences from nine volunteers, each recorded for 10 hours while typing inside a magnetoencephalography (MEG) scanner. Instead of hand-built signal processing, the system decodes raw brain signals end to end, then uses a fine-tuned language model to clean up the noise into coherent sentences. Meta released the training code for both v1 and v2; its partner BCBL is releasing the v1 dataset.

Why it matters

Restoring speech to people who’ve lost it currently means brain surgery: electrodes placed directly on the cortex. That works, but it doesn’t scale, because most people won’t accept an implant. A non-invasive method that approaches surgical accuracy would change who this technology can reach. Going from 8% to 61% is the difference between a lab curiosity and something worth taking seriously.

Here’s the part almost nobody is mentioning: MEG is not a headset. It’s a room-sized machine that costs millions, needs magnetic shielding, and requires the person to sit still inside it. So “no surgery” is true, but it quietly swaps a one-time surgical risk for a permanent hardware prison. The bottleneck for real-world use was never really the AI. It’s the sensor. And Meta’s own framing leans on a scaling law — that accuracy keeps improving log-linearly with more data. Log-linear is the honest word doing a lot of work there: it means each equal gain costs you an order of magnitude more data. That’s not a wall, but it’s not a runway either.

My take

I think this is excellent science and oversold as a communication tool, and the gap between those two things is the sensor, not the model. The decoder is clearly good enough to ride any hardware improvement that comes. So the real question isn’t “can AI read the brain” (answered, yes) but “can anyone build a wearable non-invasive sensor with MEG-level signal.”

I’m wrong if, within three years, a portable or wearable non-invasive setup posts word accuracy in this same range. If that doesn’t happen, the ceiling here was always physics, not intelligence.