Related: AI Screens 6 Million Molecules to Find New Antibiotics for Drug Resistant Gonorrhea
Meta has shown that an AI model can read the sentences a person types directly from their brain activity, without any surgery or implant, and the work has now cleared the bar of peer review. The system, called Brain2Qwerty, was published in the journal Nature Neuroscience on June 29, and it represents one of the more striking demonstrations yet that thought can be turned into text using only sensors placed outside the skull. The result is a non-invasive counterpart to the implanted brain chips pursued by companies like Neuralink, and it sharpens a long running question about how far the brain can be decoded without ever opening it.
The setup behind the research is deliberately constrained. Working with the Basque Center on Cognition, Brain and Language, researchers at Meta‘s Fundamental AI Research team recorded the brain activity of 35 healthy volunteers as they typed briefly memorized sentences on a standard QWERTY keyboard. The recordings used two non-invasive techniques, magnetoencephalography, which measures the tiny magnetic fields produced by neural activity, and electroencephalography, which reads electrical signals from the scalp. The AI then learned to map those signals back to the specific keys the person was pressing, reconstructing the sentence character by character.
The model itself stacks three stages. A convolutional module first digests short windows of the brain signal, a transformer then works at the level of whole sentences to capture context, and finally a pretrained language model cleans up the output, much as autocorrect fixes a sloppy text message. That last step matters, because it lets the system lean on what words and sentences are statistically likely rather than relying on the brain signal alone, which is faint and noisy when read from outside the head.
The performance gap between the two recording methods is the headline finding. Using magnetoencephalography, Brain2Qwerty averaged a character error rate of about 29 percent, and for the best performing participant it fell to roughly 18 percent, low enough to reconstruct many sentences accurately, including ones the model had never seen in training. Electroencephalography, the cheaper and more portable option, lagged badly at around 65 percent, a reminder that the richer magnetic signal carries far more usable information than scalp electrodes can. The takeaway is that non-invasive decoding works, but only when paired with the most sensitive equipment available.
That equipment is also the catch. Magnetoencephalography scanners are enormous, expensive machines that must sit inside a specially shielded room to block out the magnetic noise of the everyday world, which means Brain2Qwerty in its current form is a laboratory instrument and nothing like a wearable. The volunteers were healthy people typing sentences they had just memorized, not patients trying to communicate and not people thinking freely, so the system decodes a deliberate, structured act rather than open ended thought. The distance between this and a practical assistive device for someone who cannot move or speak remains large.
Still, the direction is what makes the work significant. The dominant approach to high accuracy brain decoding has relied on electrodes placed directly on or in the brain through surgery, an option that carries real medical risk and will only ever reach a small number of people. Showing that external sensors can decode language at all, and that an AI language model can carry much of the load, points toward a future in which brain interfaces might not require an operation. Even modest gains in non-invasive accuracy compound, because the same machine learning tricks that lifted Brain2Qwerty can be applied as the sensors themselves improve.
Meta did not stop at the version described in the paper. Alongside the publication the company released the full training code for Brain2Qwerty and for a second iteration it calls Brain2Qwerty v2, which it describes as a faster end to end pipeline capable of decoding sentences in real time from non invasive recordings, while the Basque Center released the dataset behind the original work. Putting the code and data in the open lets other labs reproduce, test, and build on the results rather than take Meta’s word for them, which is how a research claim hardens into a shared benchmark, and it signals that the company treats non invasive decoding as a field to grow rather than a one off demonstration.
For now the result sits squarely in the realm of research, a careful demonstration rather than a product anyone will use soon. But it lands in a field moving quickly, where implanted interfaces are advancing on one side and AI models are getting better at extracting signal from noise on the other. Brain2Qwerty stakes out the non-invasive middle ground, and by putting peer reviewed numbers behind the claim, it gives the rest of the field a concrete benchmark to chase as the long effort to read language from the brain continues.
Related on Entrelligence: Neuralink’s through-the-dura brain implant, and AI screening 6 million molecules for new antibiotics.
Your First 10 AI Skills
10 practical AI skills, copy-paste prompts and a 7-day plan to start using AI with confidence.
Download the guide →