Meta Unveils Brain2Qwerty v2: AI Decodes Sentences Straight from Brain Waves, No Surgery Needed

Meta's AI research division has announced the second generation of Brain2Qwerty, a system that translates brain activity into written text without requiring any surgical implant. The update, revealed June 29, 2026, marks a major leap from the original version released last year, with Meta describing it as the highest-performing end-to-end pipeline capable of real-time sentence decoding from non-invasive brain recordings, closing in on accuracy levels once achievable only through brain surgery.

How it works

The system was trained on roughly 22,000 sentences gathered from nine volunteers, each recorded for 10 hours while wearing a magnetoencephalography (MEG) device and actively typing. Rather than relying on manually engineered processing steps to spot neural events, Meta's researchers built an end-to-end deep learning model that decodes language directly from the raw brain signal. The team also fine-tuned large language models on the neural data, letting the system use semantic context to fill in gaps between noisy brain readings and fluent sentences, and used AI agents to help explore pipeline optimizations before engineers locked in the final configuration.

The results

According to Meta, the new version reached a 61% word accuracy rate overall, a sharp jump from the roughly 8% accuracy of prior non-invasive decoding methods. For its top-performing participant, accuracy climbed to 78%, with more than half of all sentences decoded with no more than one word wrong. Meta also reports that accuracy scales log-linearly with the amount of training data, hinting that simply collecting more data could continue narrowing the gap with invasive, surgery-based brain-computer interfaces.

Why it matters

Meta frames the work as a potential lifeline for people with brain lesions or conditions that block their ability to communicate. Invasive techniques such as stereotactic EEG and electrocorticography have already shown that pairing implanted sensors with an AI decoder can restore communication, but those approaches are hard to scale to large patient populations. A non-invasive alternative like Brain2Qwerty could make such technology far more accessible.

Open research push

To support outside researchers, Meta is releasing the full training code for both Brain2Qwerty v1 and v2. Its research partner, the Basque Center on Cognition, Brain, and Language (BCBL), is separately releasing the dataset behind version 1. The project sits alongside other open brain-modeling efforts at Meta, including its Tribev2 perception-encoding model, the NeuralSet framework for processing brain data at scale, and NeuralBench for evaluating such models — all part of a broader $5 million Digital Brain Project fund aimed at encouraging open neuroscience datasets.

Source: https://ai.meta.com/blog/brain2qwerty-brain-ai-human-communication/