A machine that reads private thoughts through the skull would change society overnight. Meta’s Brain2Qwerty research is not that machine—but it is close enough to the idea that the distinction deserves careful attention.
The experimental AI system converts non-invasive brain recordings into text while volunteers actively type memorised sentences. Its latest version produced an average word accuracy of 61%, rising to 78% for the best-performing participant, according to Meta’s 2026 report.
Those numbers are remarkable for a system that does not use a surgical implant. They do not show that AI can listen to an unsuspecting person’s inner monologue. The volunteers were inside a specialised scanner, performing a known task, after hours of individual data collection.
What Brain2Qwerty actually decoded
The first Brain2Qwerty study involved 35 healthy volunteers. Participants briefly memorised sentences and then typed them while researchers recorded their brain activity using either electroencephalography (EEG) or magnetoencephalography (MEG).
EEG uses electrodes on the scalp to measure electrical activity. It is relatively portable and familiar. MEG measures tiny magnetic fields produced by brain activity. It can capture cleaner spatial and temporal information, but normally requires a large machine, a magnetically shielded room and a participant who remains still.
In the peer-reviewed 2026 Nature Neuroscience paper, MEG achieved an average character error rate of 29%, compared with 65% for EEG. For the strongest MEG participants, the character error rate fell to 18%.
The AI was not discovering free-form thoughts. It combined signals related to language production, finger movements and the sensory feedback of typing. A language model then helped turn noisy predictions into plausible sentences.
What changed in Brain2Qwerty v2
Meta’s second version focused on more data, real-time processing and sentence-level accuracy. The company says it collected about 22,000 sentences from nine volunteers, with each person spending roughly 10 hours in an MEG scanner while actively typing.
According to the v2 research summary, the model reached an average word error rate of 39%—equivalent to 61% word accuracy. The best participant reached 78% word accuracy, and more than half of that person’s sentences were decoded with no more than one wrong word.
That is an impressive improvement over older non-invasive approaches. It is still too error-prone for dependable communication when every word matters, and the headline result came from the best of only nine participants.
Why Brain2Qwerty is not portable mind reading
Four constraints separate Brain2Qwerty from science-fiction telepathy.
First, the person is cooperating. Participants know the sentence and intentionally type it. The experiment does not recover an unrelated private idea.
Second, the model is personalised. Brain anatomy and signal patterns differ from person to person. The v2 volunteers each contributed around 10 hours of training data.
Third, the equipment is enormous. MEG is a room-scale laboratory system, not a hidden sensor in a phone, camera or pair of earbuds.
Fourth, the language model makes informed guesses. It uses statistical context to improve noisy neural predictions. That can make outputs more readable, but it also creates the possibility that a fluent-looking sentence is not what the participant intended.
Brain2Qwerty is therefore best understood as a controlled decoding system for a deliberate task—not a general-purpose reader of beliefs, memories or secrets.
Could it eventually help people who cannot speak?
That is the hoped-for application, but it has not yet been demonstrated by this research. Both major studies used healthy volunteers who could type normally. A person with paralysis, a brain injury or a condition affecting language may produce different signals and may not be able to perform the same training task.
Researchers would also need a method that does not depend on physical keystrokes, because typing itself provides useful motor and sensory information to the decoder. A clinically useful system might instead decode attempted typing, attempted speech or another intentional task.
The appeal of a non-invasive route is clear. It avoids surgery and could potentially reach more people than an implanted system. The trade-off is signal quality: the skull and scalp make brain activity harder to measure precisely, while the highest-performing non-invasive equipment is expensive and immobile.
Neural privacy cannot wait for a consumer product
Brain2Qwerty cannot secretly read the public’s thoughts. Even so, it demonstrates why rules for neural data should be created before the technology becomes convenient and commercial.
Brain signals may reveal information beyond the command a user meant to send. Future systems could infer attention, fatigue, emotion or health-related patterns. Combining neural recordings with personal profiles and powerful AI models would make that data especially sensitive.
UNESCO’s Recommendation on the Ethics of Neurotechnology calls mental privacy fundamental to personal identity and agency. It recommends informed consent, data minimisation, purpose limits, security and the ability for people to access, correct or erase neural data.
Those principles translate into practical questions for any future consumer BCI:
- Is raw neural data processed locally or uploaded?
- Is only the intended command retained?
- Can data be used to train another model?
- Can an employer, insurer or advertiser request access?
- Can the user permanently delete the recordings?
- Who is responsible when an AI decoder invents the wrong output?
The answers will matter as much as accuracy.
What Brain2Qwerty proves and what remains unproven
Brain2Qwerty shows that AI can extract surprisingly detailed information from non-invasive brain recordings during a tightly controlled typing task. It does not show that a company can point a device at someone and read unspoken thoughts.
The near-term story is assistive communication research, not mass-market telepathy. But the gap between those ideas will feel smaller as sensors and AI improve. That is why honest language and strong neural-data protections are needed now—before the technology becomes ordinary enough for people to stop asking what it can see.
Reporting note
This article discusses experimental research, not a commercially available medical or consumer system. Brain2Qwerty has been tested in small groups of healthy volunteers performing controlled typing tasks; clinical usefulness remains unproven.
Sources and further reading
- Nature Neuroscience: Noninvasive decoding of typed sentences from human brain activity
- Meta AI: Brain2Qwerty v2 research announcement
- Meta AI Research: Accurate Decoding of Natural Sentences from Non-Invasive Brain Recordings
- Meta AI Research: Brain-to-Text Decoding via Typing
- UNESCO: Recommendation on the Ethics of Neurotechnology
For a different route to intentional communication, read about brain-to-speech implants and early research converting attempted speech into a personalised voice.
Leave a Reply