Meta Unveils Brain-to-Text AI That Decodes Thoughts Without Surgery
Meta has introduced Brain2Qwerty v2, a non-invasive brain-computer interface that converts brain activity into typed text without requiring surgery. The announcement coincided with the publication of the underlying research in Nature Neuroscience.
The system achieves an average word accuracy of 61% using magnetoencephalography (MEG), which measures magnetic fields produced by brain activity. Moreover, the highest-performing participant reached 78% accuracy, while more than half of the decoded sentences contained one or fewer word errors.
Unlike earlier systems, Brain2Qwerty v2 decodes complete words and semantic meaning rather than individual characters. Consequently, the technology delivers faster and more accurate text generation from neural signals.
AI model improves brain signal decoding
Researchers trained the system on approximately 22,000 sentences collected from nine volunteers. Each participant spent about 10 hours wearing an MEG device while typing sentences.
Additionally, the model combines end-to-end deep learning with fine-tuned large language models to translate noisy brain signals into meaningful text. According to the research, performance improves as more training data becomes available, suggesting further gains in accuracy over time.
The new version also represents a significant improvement over Brain2Qwerty v1, which focused primarily on character-level decoding. Furthermore, previous high-accuracy brain-to-text systems relied on surgically implanted devices that carried medical risks such as infection and signal degradation.
Technology could help people with communication disorders
The research highlights the potential to assist people with brain injuries or neurological disorders that limit communication. Therefore, the non-invasive approach could eventually offer a safer alternative to implanted brain-computer interfaces.
To encourage further research, Meta has released the complete training code for both Brain2Qwerty v1 and v2. In addition, its research partner, the Basque Center on Cognition, Brain and Language, has made the v1 dataset publicly available.
Public reaction has remained mixed. While many welcomed the breakthrough for its medical potential and accessibility, others expressed concerns about privacy and Meta’s role in developing brain-decoding technology.
