Meta Brain2Qwerty v2 decodes typed sentences from brain scans at 61% accuracy without implants

Meta Brain2Qwerty v2 Achieves Major Leap in Non-Invasive Brain Decoding

Meta has introduced Brain2Qwerty v2, an AI system that translates brain activity into text without requiring surgery or brain implants 1

1

. The non-invasive brain-computer interface achieved an average word accuracy of 61%, with the best participant reaching 78% accuracy—more than half of decoded sentences contained one word error or less 2

2

. This marks a dramatic improvement from previous non-invasive methods that managed only around 8% word accuracy, and surpasses last year's Brain2Qwerty v1, which topped out at 48% .

The system was trained at the Basque Center on Cognition, Brain, and Language in San Sebastián, Spain, using approximately 22,000 sentences from nine healthy volunteers aged 25 to 56 5

5

. Each participant spent roughly 10 hours wearing a magnetoencephalography (MEG) scanner—a helmet-like device that measures the tiny magnetic fields produced by neuronal activity—while actively typing .

How AI Brain Activity Decoding Works Through Deep Learning

The AI system decodes typed sentences using the same pattern-recognition technology behind ChatGPT and Meta's Llama 2

2

. Brain2Qwerty v2 employs a multi-tiered approach that first converts neural signals from non-invasive magnetoencephalography scans into individual character tokens using deep learning models including Transformers and Convolutional Neural Networks 3

3

. A second AI system called an aligner then organizes these characters into complete words. Finally, large language models fine-tuned on the brain data use semantic context to reconstruct coherent sentences, much like a smartphone predicting your next word .

This represents the first time an LLM has successfully decoded noisy brain activity into structured, intelligible sentences 2

2

. Meta also deployed "auto-research" AI agents trained to iteratively modify the code base and invent novel architectures, producing substantial improvements in word error rate, though human researchers remained critical to the scientific process 2

2

.

Communication Restoration Potential for Paralysis Patients

Meta frames the technology as potentially helping millions who have lost the ability to speak due to brain lesions, locked-in syndrome, amyotrophic lateral sclerosis (ALS), and other paralyzing neurodegenerative disorders 2

2

. The system converts brain activity into text without surgery, eliminating the risks associated with invasive neuroprosthetics that typically require complex and expensive brain operations .

However, significant limitations remain. The MEG scanner is a room-sized, expensive machine that belongs in hospitals rather than homes 1

1

. The system cannot work in real time—models need a complete typing session to finish before producing output, so there's no live feedback 1

1

. More fundamentally, the technology learns from brain signals of people actually typing, yet its intended users—those fully locked in by paralysis—cannot type at all 1

1

. This would require rebuilding the task around imagined movement rather than real keystrokes.

Neuroscience Research Advances Through Open-Source Release

Meta is releasing the full training code for both Brain2Qwerty v1 and v2, while its research partner at the Basque Center on Cognition, Brain, and Language will release the v1 dataset 5

5

. The company stated, "Our hope is that this work, done in the open, advances neuroscience to identify, diagnose, and treat neurological disorders faster than in siloes" . The open-source release is part of Meta's broader Digital Brain Project, which includes a $5 million fund to support open neuroscience datasets .

Researchers found that decoding accuracy improved as training data increased, suggesting that simple scaling laws could build more capable systems in the future 2

2

. If extended training on non-invasive MEG data can eventually eliminate the need for neurosurgery, it would represent a transformative shift in patient care 2

2

. Meta also tested cheaper EEG technology but found MEG far superior, with a character error rate of 29% against 65% for EEG 1

1

.

For now, invasive approaches from companies like Neuralink still win on results, with recent surgical work hitting 92% sentence-level accuracy 1

1

. But Meta's system turns thoughts into text without cracking open skulls, and the company's pitch is that it can close the performance gap as it feeds models more data 1

1

. For people who have lost the ability to communicate, this non-invasive approach could prove more meaningful than any chatbot or image generator 3

3

.