Meta's new non-invasive brain-computer interface hits 78% accuracy with AI

Meta has revealed its latest brain-computer interface project, Brain2Qwerty v2. It uses a device called MEG to pick up tiny magnetic fields from brain activity, then uses AI to turn those signals into text—no surgery or implants needed. The goal is to help people who can't speak or move due to brain injuries or diseases to communicate just by thinking.

The AI model was trained on data from 9 volunteers, including 22,000 sentences and about 10 hours of brain recordings. It can even correct noisy brain signals using context. In tests, the system hit an average word accuracy of 61%, and the best user reached 78%. More than half of the test sentences had no more than one wrong word.

Still, the tech has a big catch: it needs a huge, expensive lab-grade MEG machine, so it's far from ready for everyday use. Meta has open-sourced the training code on GitHub, and the datasets will be shared soon.