A volunteer typed “the music is not too loud.” Meta’s brain-decoding model returned “my homework is due tomorrow.” For another volunteer typing the same words, it reproduced the sentence perfectly.
Both results appear in a research paper released June 29. Their difference reveals a problem more immediate than science-fiction mind reading: artificial intelligence can produce fluent language that its user never intended.
Nine Typists and a 306-Sensor Brain Scanner
On June 29, 2026, Meta introduced Brain2Qwerty v2, a noninvasive system designed to turn brain activity into text. Developed with researchers at the Basque Center on Cognition, Brain and Language in Spain, it aims eventually to help people who have lost the ability to communicate without requiring surgically implanted electrodes.
The experiment, however, depended on an activity many potential patients cannot perform.
Nine healthy, right-handed, English-speaking volunteers listened to sentences through headphones, paused, then physically typed them while magnetoencephalography (MEG) measured faint magnetic fields produced by their brains. Each contributed roughly ten hours of recordings. Across participants, researchers gathered approximately 22,000 sentences.
Unlike its predecessor, which needed the timing of individual keystrokes, v2 takes continuous MEG recordings as its input. A neural decoder estimates character sequences, a second component associates signals with words, and a fine-tuned large language model assembles the sentence. Meta also used AI agents to test improvements to the pipeline, with researchers selecting the final configurations.
Meta reported 61% average word accuracy, equivalent to a 39% word error rate. Its best-performing participant reached 78% accuracy.
Those results represent a substantial advance in noninvasive decoding. But they do not demonstrate the ability to read private thoughts. Participants were reproducing sentences they had heard and were actively typing, not silently imagining unrelated ideas.
Nor does coherent output guarantee faithful interpretation. In the researchers’ comparison, adding the language model reduced errors measured by words but increased errors measured character by character. When the neural evidence was weak, the system sometimes generated plausible sentences detached from the intended text.
Meta described the system as capable of real-time decoding, yet its paper acknowledges that the current architecture must process an entire sentence before returning words. The experiment also required a bulky, 306-sensor cryogenic MEG apparatus.
Most importantly, researchers have not demonstrated that the technique works for people unable to move or type. The study remains a research preprint, rather than evidence of a clinically available communication device.
The distinction matters. Brain2Qwerty offers a possible path around the surgical risks of implanted interfaces. Whether that path leads to a practical means of communication remains an open scientific question.
The Right to Correct the Machine
For a generation raised on predictive text, a machine completing a sentence is familiar. Usually, we can backspace.
But a neural communication device could eventually become the very means through which someone unable to speak asserts what they meant. In that setting, a fluent error may be more consequential than gibberish: another person could mistake the software’s interpretation for the user’s own words.
Students have already articulated a related problem with the interpretation of bodily data.
In a 2020 Georgia Tech study, researchers interviewed students about classroom wrist sensors that collected physiological information—not brain signals. One undergraduate business student, identified as P11, questioned attempts to infer emotions from those measurements: “Maybe I wasn’t upset about the class, I was just upset that day.”
An undergraduate computer science student, P3, expressed greater comfort with data sharing: “I don’t personally mind sharing my data with anyone really.” Yet she also acknowledged others’ concerns and preferred aggregated information that would better protect their privacy.
The two perspectives expose separate vulnerabilities. A system might misinterpret what someone's data means. Even when its interpretation is accurate, the person may have limited control over where that information goes.
For Brain2Qwerty, those questions become inseparable. Meta’s findings suggest that collecting more neural data can improve decoding accuracy. The same information that could help restore someone's voice may therefore become increasingly valuable to developers training future models.
The volunteers consented to a defined research experiment. That consent should not be mistaken for permission to reuse recordings indefinitely or to make neural interpretations available to employers or insurers. Brain2Qwerty has not been deployed for such purposes, but those potential uses expose gaps that technical accuracy alone cannot resolve.
Governments have begun responding. Colorado and California adopted protections covering neural data in 2024. A 2025 UNESCO recommendation also warned against workplace neuro-monitoring and emphasized explicit consent, particularly in technologies affecting young people.
The easiest mistake is to dismiss neural privacy concerns because today's decoders cannot extract secret thoughts. The harder possibility is a future in which institutions treat an imperfect neural prediction as more authoritative than the person whose brain produced the signal.
The right to communicate may need to include something deceptively simple: the right to say, even when a machine has made the words sound convincing, that is not what I meant.
Explore more exclusive insights at nextfin.ai.
