Meta's Brain2Qwerty v2 Decodes Typed Sentences from Non-Invasive MEG Recordings

The gap between what your brain thinks and what your fingers can type has been shrinking for years. Invasive brain implants have let paralyzed patients type at reasonable speeds. But those require surgery — drilling into the skull, threading electrodes into brain tissue, accepting the risk of infection and rejection. For everyone else, decoding language from the brain without surgery has remained far less accurate.

Meta’s FAIR research team and academic collaborators narrowed that gap. Their system, Brain2Qwerty v2 (paper dated June 29, 2026, under review), uses magnetoencephalography — MEG, a helmet-like scanner that detects the magnetic fields generated by the brain’s electrical activity — to reconstruct sentences as participants type them. No surgery, no implants. It decodes typing activity, not arbitrary thoughts.

What Brain2Qwerty Actually Does

The setup is conceptually simple. A participant sits inside an MEG scanner and types sentences on a keyboard. The scanner records the magnetic fields produced by their brain activity. Brain2Qwerty v2 — an end-to-end deep learning pipeline — translates those MEG signals into the text the person was typing.

The team collected about 22,000 typed sentences from nine healthy volunteers, each recorded for roughly 10 hours. Across participants, the model achieves an average word error rate (WER) of 39% — about 61% word accuracy. The best participant reached a 22% WER (78% word accuracy), and the authors report that about half of that participant’s sentences were decoded with one word error or less. That is roughly a twofold improvement over Brain2Qwerty v1 (52% WER for the best subject). Invasive implants still perform better, but the authors note that accuracy improves log-linearly with more data, suggesting the gap could partly be narrowed by scaling.

The architecture is what makes it work. Earlier pipelines relied on hand-crafted steps to detect individual keystroke events. According to the official code repository, Brain2Qwerty v2 decodes whole sentences from a continuous MEG window without segmenting around keystrokes. It jointly trains three parts: a convolutional + Conformer encoder with a character-level CTC head, a word-level contrastive objective that aligns neural word embeddings to a language model’s embeddings, and a LoRA-adapted language model that generates the sentence. The team also used AI agents to explore optimizations of the decoding pipeline, with final configurations chosen by engineers.

Why MEG Matters

Most non-invasive brain reading to date has used fMRI or EEG. fMRI measures blood flow — a slow, indirect signal. EEG measures electrical activity at the scalp, but the skull smears the signal so badly that spatial resolution is poor. MEG sits in a sweet spot: it detects magnetic fields, which pass through the skull essentially undistorted, giving you both decent spatial resolution and millisecond-level temporal precision.

The catch is that MEG scanners are expensive, rare, and require magnetically shielded rooms. You’re not going to wear one at your desk. But the point of Brain2Qwerty isn’t to ship a consumer product tomorrow. It’s to demonstrate that the signal is there — that non-invasive brain reading can be good enough that we should take it seriously.

What This Means for LLM Research

This is where things get interesting for the AI community. Brain2Qwerty v2 uses a fine-tuned language model as one of its components, and the authors’ ablations show that removing the MEG-derived tokens from the language model’s input substantially worsens results — evidence that the system extracts semantic information from brain signals rather than relying on the language model alone. This fits a growing body of research linking brain activity to the internal representations of language models.

If that line of research holds up, improvements in language models could feed into brain-computer interfaces, and vice versa. A better understanding of how the brain encodes language could lead to more efficient AI architectures. Better AI architectures could lead to more accurate brain decoders. The loop closes.

The authors frame the work around restoring communication for people who have lost the ability to speak or move after brain injury, while also highlighting three ways AI drove the gains: deep learning replacing hand-crafted pipelines, fine-tuned language models, and agent-assisted code development.

The Privacy Elephant in the Room

Any technology that decodes language from brain activity — even crudely, even with consent — raises hard questions. Brain2Qwerty v2 only works when a participant is actively typing inside an MEG scanner. It can’t read passive thoughts, can’t be deployed remotely, and requires willing cooperation. The current system is about as invasive as getting an MRI.

But the trajectory is clear. If accuracy keeps improving and if MEG hardware becomes smaller and cheaper, the ethical questions shift from “can we?” to “should we?” A future version of this technology could, in theory, be deployed in settings where consent is murky — law enforcement interrogations, employment screening, security checkpoints.

The current system only works with active participation — the participant has to be typing. But the history of technology suggests that safeguards designed for one generation of hardware don’t always survive the next.

The Bottom Line

Brain2Qwerty v2 is a research milestone, not a product announcement. But it’s the kind of milestone that changes the conversation. A few years ago, decoding sentences from non-invasive brain recordings at 78% word accuracy for the best participant would have sounded like science fiction. Now it’s a paper under review, with code available.

The gap between non-invasive and surgical brain-to-text decoding is narrowing. How society handles the ethical questions will matter as much as the accuracy numbers.

Sources: Meta AI publication · Meta AI blog · Project page · Code