Meta Unveils a Non-Invasive Brain-to-Text AI System

Meta unveiled Brain2Qwerty v2, a non-invasive brain-computer interface designed to decode typed sentences from raw neural signals in real time.

The system marks what Meta describes as the highest-performing technology of its kind. Rather than relying on surgical implants, Brain2Qwerty v2 uses magnetoencephalography, or MEG, a non-invasive method that measures magnetic fields generated by brain activity.

The announcement coincided with the publication of the original Brain2Qwerty research in Nature Neuroscience, placing the work within a formal research setting while drawing wider attention to the progress of non-invasive brain decoding.

Brain2Qwerty v2 Accuracy and Performance

Brain2Qwerty v2 achieved an average word accuracy of 61% across participants using MEG.

For the top-performing participant, accuracy reached 78%. In that case, more than half of the decoded sentences contained one or fewer word errors.

That level of word-level performance represents a notable step forward for non-invasive brain-computer interfaces, especially because high word-level accuracy in brain decoding had previously been available only through surgical implants.

Average and Top Participant Results

 

Brain2Qwerty v2 Measure

 

 

Reported Result

 

 

Average word accuracy across participants

 

 

61%

 

 

Top participant word accuracy

 

 

78%

 

 

Top participant sentence decoding

 

 

More than half of sentences had one or fewer word errors

 

How Brain2Qwerty v2 Works

Brain2Qwerty v2 was trained on approximately 22,000 sentences from nine volunteers.

Each volunteer was recorded for 10 hours while wearing an MEG device and typing. The system was built to learn from raw brain activity associated with typing and convert those signals into decoded language.

The pipeline combines end-to-end deep learning on raw brain signals with fine-tuned large language models. Meta described this approach as a way to bridge “the gap between noisy neural data and coherent language.”

Training Data From Typed Sentences and MEG Signals

The training setup centered on volunteers actively typing while their brain activity was recorded.

The system did not depend on invasive implants. Instead, it used MEG recordings collected while participants wore the device, giving the model access to neural signals linked to typed sentence production.

The dataset used for Brain2Qwerty v2 included:

  • Approximately 22,000 sentences
  • Nine volunteers
  • 10 hours of recording per volunteer
  • MEG recordings captured during typing

Moving Beyond Character-Level Brain Decoding

Brain2Qwerty v2 advances beyond the character-level decoding used by its predecessor.

Brain2Qwerty v1, published the same day in Nature Neuroscience, achieved a character error rate of 32% using MEG. Brain2Qwerty v2 moves from that character-level approach toward decoding words and semantics directly.

Meta says the system’s performance scales log-linearly with data volume. That suggests further improvements may be possible with more training data.

Clinical Promise for Communication Disorders

Meta framed Brain2Qwerty v2 as research with potential relevance for people who cannot communicate because of brain lesions or neurological disorders.

The company said, “We believe this research has the potential to make a real difference for the millions of people who suffer from brain lesions or disorders that prevent them from communicating.”

That promise is central to the clinical interest around non-invasive brain-computer interfaces: the possibility of helping people communicate without requiring surgical implantation.

Non-Invasive Brain Decoding Compared With Surgical Implants

Until now, high word-level accuracy in brain decoding had been available only through surgical implants.

Those implants can carry risks, including infection and signal degradation over time. Brain2Qwerty v2 is notable because it reaches a reported 61% average word accuracy using MEG, a non-invasive method.

That distinction matters because the system points toward brain-to-text decoding without surgery, while still producing word-level sentence output in real time.

Open Science and Research Access

Meta released the full training code for both Brain2Qwerty v1 and Brain2Qwerty v2.

Its research partner, the Basque Center on Cognition, Brain and Language, released the v1 dataset. The release was intended to accelerate further work around the technology.

Jean-Rémi King, a researcher involved in the project, clarified on social media that the peer-reviewed paper was published in Nature Neuroscience.

Public Reaction to Meta’s Brain-Reading AI

Public reaction to Brain2Qwerty v2 was divided.

Some praised the advance for its accessibility potential, especially because it is non-invasive and aimed at communication challenges. Others expressed distrust of Meta’s involvement in brain-reading technology, citing the company’s advertising-driven business model.

That mixed response reflects the tension around the work: Brain2Qwerty v2 points to meaningful progress in non-invasive communication technology, while also raising concerns about who develops brain-decoding systems and how the public understands their use.