Accurate Decoding of Natural Sentences From Non-Invasive Brain Recordings

- Meta (Facebook's parent company) has published research on decoding natural language sentences from brain activity using non-invasive recordings. - The study uses magnetoencephalography (MEG) and electroencephalography (EEG) — sensing technologies that measure magnetic fields or electrical signals from outside the skull, unlike older brain-computer interfaces that require surgical implants. - A deep-learning model (similar to the GPT-style architecture behind ChatGPT) is trained to map brain signals to representations of speech, then generate the corresponding sentences. - Prior work in "brain decoding" mostly relied on fMRI (slow and bulky) or invasive electrodes; this is a step toward practical, non-invasive language BCIs that could eventually help people who cannot speak or type. - Key challenge: MEG/EEG signals are noisy and vary between people; the model shows solid accuracy but is far from ready for real-world use.