Overview
- Meta unveiled an updated Brain2Qwerty system on Monday, June 29, 2026, and published research showing noninvasive recordings can be decoded into text and that the team has released code and datasets to encourage open research.
- The system uses a three‑stage AI pipeline — a convolutional front end on 500 ms windows, a transformer for sentence‑level patterns, and a pretrained language model for character correction — to turn brain signals into typed text.
- Results differ sharply by sensor: magnetoencephalography (MEG) produced the strongest outcomes with roughly 70–80% character accuracy in lab tests, while EEG performance was much lower and noisier.
- Brain2Qwerty is not real time and was tested on 35 healthy volunteers typing memorized sentences, so it decodes motor planning for finger movements rather than free‑form thoughts and depends on full sentence sequences for prediction.
- Major barriers remain before any clinical or consumer use: MEG machines cost millions and are room‑sized, sensor miniaturization and real‑time decoding are unsolved engineering problems, and regulatory and clinical validation will take years.