Meta Achieves Reading Thoughts and Converting Them to Text Without Surgery

What happened?

Meta has announced the second version of Brain2Qwerty, a non-invasive system that decodes a person's brain activity while typing and converts it into text. Unlike alternatives like Neuralink, which require surgery to implant electrodes, Brain2Qwerty uses magnetoencephalography (MEG) and electroencephalography (EEG) to capture neural signals from outside the skull. In tests with 35 volunteer participants, the system reconstructed sentences with an average accuracy of around 80% in the character error rate (CER) metric, according to data published by Meta. However, this accuracy was achieved only after each user trained the model for about 20 minutes, and under highly controlled laboratory conditions, with MEG equipment costing over $2 million and requiring magnetic shielding.

The study, published on arXiv, details that the system combines MEG (which measures magnetic fields) and EEG (which measures electrical activity) to capture both motor preparation and finger movement execution. Meta trained a deep learning model with data from over 1000 hours of typing, enabling the system to predict key sequences from brain signals. Although the technical advance is significant, the reliance on physical typing limits its direct application to people with severe paralysis who cannot move their hands. Meta acknowledges this issue and states that its long-term goal is to adapt the technology to decode inner speech or the intention to communicate, but for now the system only works with actual finger movements.

Why is it important?

This advance represents an important step toward brain-computer interfaces (BCIs) that are accessible without surgical risks. Most current BCIs that achieve high accuracy, such as Blackrock Neurotech's Utah Array or Neuralink's implantable device, require invasive surgery. Brain2Qwerty v2 demonstrates that comparable performance (in terms of typing speed, around 12 words per minute vs. 8 for implanted systems) can be achieved without opening the skull. However, the paradox is evident: the system learns from signals generated by typing, making it useless for those who cannot move their hands. Meta acknowledges that the ultimate goal is to help people with paralysis, but the current method depends on the motor skills that these patients have precisely lost.

In market terms, non-invasive BCIs have attracted significant investments. Companies like Synchron (which uses a stent in blood vessels) and NextMind (acquired by Apple) have raised hundreds of millions of dollars. However, most commercial non-invasive solutions, such as Emotiv's EEG headsets, have much lower accuracy and only detect simple commands. Brain2Qwerty v2, by reconstructing full sentences, marks a milestone in decoding natural language from non-invasive brain signals.

Consequences and challenges

The development raises ethical questions about mental privacy and the use of neural data. Unlike traditional biometric data, brain signals can reveal information about emotions, intentions, or even involuntary memories. Meta, which has faced criticism for its handling of user data in the past, assures that Brain2Qwerty data is stored anonymously and not shared with third parties. However, there is still no specific regulatory framework for neurodata protection, although countries like Chile have advanced in neuro-rights laws.

Additionally, reliance on bulky equipment like MEG limits its application outside the lab. MEG requires a magnetically shielded room and cryogenic temperatures for sensors, making the system unsuitable for home use. Meta is exploring versions based solely on EEG, which are portable and low-cost (around $500), but accuracy with EEG alone is significantly lower (around 60% accuracy in initial tests). The scientific community notes that while impressive, the system is still far from practical for everyday use. Dr. José del R. Millán, a BCI expert at the University of Texas, commented: "Meta has achieved a technical milestone, but the gap between the lab and real life remains huge. To be useful, we need portable systems that work without extensive training and in noisy environments."

What readers should know

• Brain2Qwerty v2 does not read arbitrary thoughts, only those associated with the act of typing. It cannot access memories, emotions, or thoughts unrelated to the task.
• The reported accuracy is high (around 80% CER), but only under controlled conditions and with prior user-specific model training. Without that training, accuracy drops to 30%.
• There is no commercial release date; Meta presents it as basic research. The company has published the data and model on GitHub to encourage academic collaboration.
• Alternatives like Neuralink (surgery) or low-cost EEG-based systems (like OpenBCI) are advancing in parallel. Neuralink has enabled paralyzed patients to control computer cursors, but requires surgical implants.
• The cost of MEG equipment (over $2 million) makes the technology inaccessible to most hospitals and research centers. Meta is working on solid-state MEG sensors that could reduce the cost to $100,000 in the coming years.

In summary, Brain2Qwerty v2 is a remarkable scientific advance demonstrating the potential of non-invasive BCIs, but its practical application for people with paralysis remains a challenge. The gap between the lab demonstration and a viable consumer product is still wide, and will require advances in portability, accuracy, and ethics before it can transform patients' lives.