Can AI Read Your Mind? How Close Are We Really?

An AI That Turns Brain Activity Into Sentences

Meta’s Brain2Qwerty v2 works with magnetoencephalography, or MEG, a technique that detects the extremely weak magnetic fields generated by electrical activity in the brain. Unlike an implanted brain-computer interface, MEG requires no surgery. Researchers trained the system using recordings from nine healthy volunteers. Each participant spent around 10 hours in the MEG scanner while typing, producing roughly 22,000 sentences for the study. The AI learned to associate patterns of brain activity with letters, words and entire sentences.

The results were striking. Brain2Qwerty v2 achieved an average word accuracy of 61%, compared with around 8% for previous non-invasive approaches. The best-performing participant reached 78% word accuracy, with more than half of the decoded sentences containing one word error or fewer. The previous Brain2Qwerty version reached 48% in its best participant.

The researchers found another important pattern: performance continued to improve as more training data was added. So far, they have not found a clear point where the improvement stops. That raises the possibility that larger datasets and more powerful AI models could push brain-to-text systems considerably further.

The crucial detail is what the system is actually decoding. Brain2Qwerty is not scanning someone’s mind and producing a transcript of every thought. It is learning the neural patterns associated with a highly specific task: producing typed sentences. The breakthrough lies in how effectively AI can extract linguistic information from an extremely noisy biological signal.

Click "Allow" to continue