New Delhi: Tech giant Meta has introduced Brain2Qwerty v2, a revolutionary artificial intelligence system capable of decoding non-invasive brain recordings into real-time text without requiring surgical implants. This development marks a major milestone in non-surgical brain-computer interface research, closing the performance gap with traditional invasive neurological procedures. The breakthrough holds immense potential to restore seamless communication capabilities for millions of people worldwide currently suffering from severe brain lesions, paralysis, and complex neurological disorders.
Unlike gold-standard neuroprosthetic methods such as stereotactic electroencephalography and electrocorticography, which rely heavily on high-risk skull surgery to implant internal electrodes, Brain2Qwerty v2 captures neural activity externally. The technology utilizes a specialized helmet equipped with magnetoencephalography sensors to measure the microscopic magnetic fields produced by natural neural activity. Meta trained the advanced model on roughly 22,000 sentences gathered from nine healthy volunteers who spent 10 hours actively typing while wearing the non-invasive scanning equipment.
By transitioning from older hand-engineered analytical pipelines to a unified end-to-end deep learning framework, the AI directly processes raw, unfiltered neural signals to identify linguistic patterns. To counter the inherently chaotic nature of external brain scanning, Meta engineers fine-tuned large language models on the structural neural data, allowing the system to use grammatical and semantic context to resolve ambiguous readings. The system achieved a remarkable average word accuracy rate of 61 per cent, presenting a massive technological leap over the meager 8 per cent accuracy baseline recorded by previous non-invasive frameworks.
For the top-performing participant in the clinical study, the decoding system hit an exceptional 78 per cent word accuracy rate, with over half of the translated sentences containing one word error or less. Crucially, the researchers observed that decoding accuracy scales log-linearly with data volume, implying that expanding the dataset could soon allow non-invasive systems to fully match the precision of brain surgery. To foster global collaboration and accelerate open-source clinical research, Meta has publicly released the full training code for both iterations of the system, while its research partner, the Basque Center on Cognition, Brain and Language, will open-source the accompanying foundational dataset.
The project operates under Meta’s broader Digital Brain Project, an initiative dedicated to constructing comprehensive, open-source foundational models of the human brain to improve the diagnostic and therapeutic landscapes of global neurology. As part of this expansion, the tech firm highlighted several adjacent systems, including its TribeV2 model for complex multi-sensory perception encoding, NeuralSet for large-scale neurological data processing, and NeuralBench for standardizing model evaluations. To cement its commitment to open science, Meta also announced a brand-new USD 5 million global research fund explicitly intended to support academic institutions in building open-access neuroscience datasets.
