Type with your thoughts? Meta's Brain2Qwerty v2 explained

Meta has introduced Brain2Qwerty v2, an upgraded artificial intelligence system that translates brain activity into written sentences without requiring brain implants, marking a significant step forward in non-invasive brain-to-text technology.The company said the research is aimed at helping people who lose the ability to communicate following a stroke, an accident or neurological disorders. Unlike approaches that depend on surgically implanted electrodes, Brain2Qwerty v2 relies on magnetoencephalography (MEG), which captures brain activity from outside the head.Meta described Brain2Qwerty v2 as its most advanced end-to-end system for decoding complete sentences in real time from non-invasive brain recordings.Also Read: Meta bets on AI to replace human moderators as content policing shiftsTrained on thousands of sentencesThe model was trained using data from nine volunteers, each of whom spent about 10 hours wearing an MEG scanner while typing nearly 22,000 sentences.Rather than using manually designed methods to identify specific neural signals, the system processes raw brain activity through end-to-end deep learning. Meta also said it fine-tuned large language models on neural data so the system could use semantic context to better interpret noisy brain recordings.The company added that AI agents were used to test different optimisations for the decoding pipeline, while engineers selected the final training configuration.Accuracy improves sharplyAccording to Meta, Brain2Qwerty v2 achieved an average word accuracy of 61%, substantially higher than the roughly 8% reported by previous non-invasive approaches.For the best-performing participant, the system reached 78% word accuracy, with more than half of the decoded sentences containing one word error or fewer.Researchers also observed that performance improved as the volume of training data increased, suggesting the remaining gap with invasive brain-computer interfaces could narrow further through larger datasets.Focus on accessible communicationMeta said invasive techniques such as stereotactic electroencephalography and electrocorticography have already demonstrated that AI-powered neuroprostheses can restore communication, but those procedures require brain surgery and are difficult to deploy widely."Our noninvasive approach can help bridge that gap," the company said.Alongside the announcement, Meta said it is releasing the full training code for Brain2Qwerty v1 and Brain2Qwerty v2. Its research partner, the Basque Center on Cognition, Brain, and Language (BCBL), will also make the Brain2Qwerty v1 dataset publicly available to support further neuroscience research.The company said the work forms part of its broader effort to build open foundational brain models and accelerate research into diagnosing and treating neurological disorders.