Meta’s New AI Can Read 61% of What You Type — Straight From Your Brain

Nine volunteers spent ten hours apiece inside a helmet that looks like something built for a different decade, typing ordinary sentences while machines tracked the faint magnetic flickers of their own neurons. Out of that came a number that matters more than it first appears: 61 percent. That is how often Meta’s newest brain-reading system now gets the word right, with no surgery, no implant, and no opened skull.

On June 25, Meta’s Fundamental AI Research division published Brain2Qwerty v2, a system that decodes full typed sentences from magnetoencephalography, or MEG, recordings in something close to real time. It is the clearest sign yet that the gap between invasive and non-invasive brain-computer interfaces, long assumed to be unbridgeable without wires inside the skull, is narrowing faster than most neuroscientists expected.

What Happened

Meta detailed Brain2Qwerty v2 in a paper published in Nature Neuroscience and in a companion post on its AI research blog, alongside the full training code for both the new model and its 2025 predecessor. The team trained the system on roughly 22,000 typed sentences gathered from nine volunteers, each of whom wore an MEG helmet for about ten hours while typing on a keyboard. MEG captures the faint magnetic fields produced by electrical activity in the brain, picked up by sensors arrayed outside the head rather than implanted within it.

The headline figure is a 61 percent average word accuracy rate across participants, against roughly 8 percent for earlier non-invasive approaches Meta cites as the prior baseline. For the best-performing volunteer, accuracy reached 78 percent, with more than half of that person’s sentences decoded with one word error or fewer. Meta also released the dataset behind the original Brain2Qwerty model through its research partner, the Basque Center on Cognition, Brain and Language in Spain, and pointed to a $5 million fund the company has committed toward open neuroscience datasets more broadly.

The Mechanism: How a Helmet Reads Typing

From Hand-Built Rules to End-to-End Learning

Earlier non-invasive decoders relied on hand-built pipelines: a researcher would design rules to spot specific neural events, then string those events into letters and words. Brain2Qwerty v2 throws that scaffolding out in favor of end-to-end deep learning, building on the architecture first detailed in the peer-reviewed Brain2Qwerty v1 paper in Nature Neuroscience. A convolutional encoder reads the raw MEG signal directly, learning which patterns of magnetic activity correspond to which intended keystrokes without being told in advance what to look for. A transformer layer then models the longer structure of a sentence, the way a string of likely letters tends to assemble into real words. A character-level language model sits on top, nudging ambiguous output toward plausible English.

That layering is what lets the system recover from noise. A raw MEG signal is messy: it captures muscle movement, ambient interference, and the ordinary chaos of a working brain alongside the typing-related activity researchers actually want. By fine-tuning large language models on neural data, Meta’s team gave the decoder a way to lean on semantic context, essentially letting it guess that a garbled signal probably meant “the” rather than some implausible four-letter fragment, the same way autocorrect leans on probability rather than pure signal fidelity. Meta also used AI agents to explore configurations for the decoding pipeline during development, though engineers selected the final training setup by hand rather than letting the agents choose autonomously.

Why Data Volume, Not Architecture, Drove the Jump

One detail explains much of the year-over-year jump: accuracy rose in a roughly log-linear relationship with the volume of training data. Each additional hour of recording bought a predictable, if diminishing, accuracy gain. That is a different kind of promise than a one-off architectural trick: it suggests Meta, or any lab with enough MEG hours, could keep closing the remaining gap with invasive systems simply by collecting more data, rather than waiting on some unproven algorithmic breakthrough.

The Backstory: Why Non-Invasive Mattered All Along

Brain-computer interfaces have split for years into two camps with very different risk profiles. Invasive systems, the kind built on electrodes implanted directly onto or into brain tissue through neurosurgery, have already demonstrated real-time cursor control and, in limited clinical settings, restored some communication for paralyzed patients. Elon Musk’s Neuralink is the most visible example of that approach, and it has shown what is achievable once electrodes sit in direct contact with neurons. The open-source posture behind Brain2Qwerty v2 also tracks with Meta’s broader research strategy, the same instinct toward releasing weights and training code that shaped Meta’s Llama 4 model family, even as the company has simultaneously moved toward closed, commercial models elsewhere in its product lineup. The cost of the invasive approach’s performance is steep: a craniotomy, infection risk, and a device that must be maintained inside a person’s skull indefinitely.

Non-invasive alternatives, EEG and MEG chief among them, sidestep the surgery entirely but have historically paid for that safety with a severe accuracy penalty. Meta’s own first-generation Brain2Qwerty, published in February 2025, reported a character error rate around 32 percent for MEG and as high as 67 percent for EEG-based recordings, results researchers at the time described as scientifically meaningful but commercially unworkable. The model did not operate in real time and required a full sentence before producing any output. Version 2’s shift to genuine sentence-level, near-real-time decoding, combined with the accuracy jump, is the first signal that non-invasive systems might eventually approach what surgery alone could previously deliver.

Reactions

Independent AI researchers covering the release have largely focused on what Meta chose to do alongside the science. By open-sourcing the full training pipeline for both Brain2Qwerty generations and releasing the underlying BCBL dataset rather than keeping the work proprietary, Meta handed outside labs the means to stress-test its claims and extend the architecture, rather than asking the field to take the company’s word for it. That posture stands out at a moment when competition for AI research talent between Meta, DeepMind, and a wave of well-funded startups has fueled an exodus of researchers out of the major labs, and publishing high-profile, reproducible work in the open is one lever Meta’s FAIR division can pull to keep its research division an attractive place for scientists to stay, even as commercial pressure pulls the rest of the company toward closed models.

Coverage from outlets following brain-computer interface research has also been careful to separate the achievement from any near-term consumer application. The MEG scanners used in the study are large, expensive machines that belong in a research lab, not a living room, and the participants were healthy volunteers rather than people with the speech-affecting conditions Meta says ultimately motivate the work. Whether the underlying decoding approach generalizes to patients with ALS, locked-in syndrome, or anarthria, the populations Meta names as its long-term target, remains an open question the current study does not answer.

The Dispute: Real Time, Or Real Time Enough?

There is some daylight between how Meta frames the system and how the underlying mechanics actually work. Meta describes Brain2Qwerty v2 as capable of real-time sentence decoding, and several outlets have repeated that framing without qualification. But the model still processes information at the sentence level rather than continuously streaming individual keystrokes as they happen, meaning a user would need to finish a thought before the system renders it as text, a meaningfully different experience from typing on a keyboard and seeing letters appear instantly. Whether that counts as “real time” in any practical sense is more a matter of definition than a dispute over the underlying numbers, but it is worth flagging for readers who might otherwise picture something closer to live transcription than the system currently delivers.

What Happens Next

Meta’s own framing points toward data scaling as the clearest near-term lever: if accuracy keeps climbing in step with recording hours, the company’s evident next step is simply collecting more MEG data across a wider range of participants, rather than redesigning the architecture from scratch. The release of the v1 dataset and the broader $5 million open-data fund both suggest Meta is betting that other labs, given the tools, will help generate exactly that additional training volume faster than Meta could alone.

The harder open problem is portability. Today’s MEG scanners are room-sized, shielded, and far too expensive and immobile for any clinical or consumer deployment outside a research setting. Meta has not announced a timeline for miniaturized or wearable MEG hardware, and without one, Brain2Qwerty v2 remains a proof of concept rather than a path to a device a patient could use at home. The next meaningful checkpoint for this research is less likely to be a higher accuracy number than evidence the same decoding approach works on patients who have actually lost the ability to speak, rather than on healthy volunteers typing in a lab.

Why It Matters

Brain2Qwerty v2 lands at a moment when Meta is simultaneously courting controversy over a pivot toward closed, commercial AI models elsewhere in the company, which makes the open-source treatment of this particular research notable by contrast. It also arrives as a quieter counterpoint to the industry’s more headline-grabbing AI storylines this year, chatbots, agents, and frontier model races, by pointing toward a use of AI that is harder to monetize but potentially more directly consequential for the people it is built to help: those who have lost the ability to communicate and have, until now, had only the surgical option available to them.

Sources

Meta AI Research blog (ai.meta.com), “From Brain Waves to Words: Brain2Qwerty Offers a New Path to Communication Without Surgery,” June 2026. Additional reporting from MarkTechPost, Digital Trends, and Storyboard18.