Brain2Qwerty v2 decodes natural sentences from MEG with 39% word error

Researchers at the intersection of neuroscience and AI have introduced Brain2Qwerty v2, a brain-to-text model that reads natural sentences directly from magnetoencephalography (MEG) recordings — a non-invasive alternative to intracranial implants. Published on arXiv, the study collected 22,000 typed sentences from nine participants, each recorded for 10 hours. Using character, word, and sentence-level representations, the model reaches a word error rate (WER) of 39% on average. For the best-performing participant, over half of the sentences were decoded with zero or one error, bringing near-practical communication within reach.

The paper attributes the leap to three AI-driven innovations: replacing hand-crafted event-detection pipelines with deep learning, fine-tuning large language models to extract semantic representations, and using AI agents to iteratively refine the decoding pipeline through automated code development. Critically, decoding error rate improves log-linearly with training data volume, suggesting that performance can scale toward intracranial levels simply by collecting more recordings. Brain2Qwerty v2 doesn't yet match the accuracy of surgical electrode arrays, but it offers a safe, scalable path forward — one that could restore communication for locked-in patients without the risks of brain surgery.