Meta unveils AI system that translates brain signals into text with 61% accuracy

Meta just demonstrated that reading someone’s mind, or at least their typing intentions, no longer requires drilling into their skull. The company’s new AI system, Brain2Qwerty v2, translates brain activity into text using entirely non-invasive techniques, achieving an average word accuracy of 61%.

The best-performing participants in the study hit 78% decoding accuracy. To appreciate how significant that is, consider that prior non-invasive methods managed roughly 8% accuracy. That’s the difference between a system that produces gibberish and one that produces mostly coherent sentences.

How it works, minus the neuroscience degree

Think of it like this: when you type on a keyboard, your brain fires specific electrical patterns for each key you intend to press. Brain2Qwerty v2 captures those patterns using two established brain-scanning technologies, magnetoencephalography (MEG) and electroencephalography (EEG), and then feeds that data through a sophisticated AI pipeline to reconstruct what you were trying to type.

Neither MEG nor EEG requires surgery. MEG measures magnetic fields generated by neural activity, while EEG uses electrodes placed on the scalp to detect electrical signals.

The system employs a hybrid approach that combines three distinct layers of machine learning. Convolutional neural networks handle the initial feature extraction from raw brain signals. Transformer models manage the sequence modeling. And pretrained language models refine the output to produce coherent text rather than random character soup.

From v1 to v2, and why it matters

Brain2Qwerty v2 builds on research Meta published in February 2025 as Brain2Qwerty v1. That earlier version achieved up to 80% character accuracy using MEG and EEG data collected from 35 volunteers. The distinction between character accuracy and word accuracy is important here. Getting individual letters right 80% of the time is impressive, but reconstructing full words correctly at a 61% average rate (and up to 78% for top performers) represents a qualitatively different achievement.

The project is a collaboration between Meta and the Basque Center on Cognition, Brain and Language (BCBL), a research institution specializing in the neuroscience of language processing.

The primary target application is assistive communication. People who have lost the ability to speak due to brain injuries, strokes, or neurodegenerative conditions could potentially use this technology to communicate through text. The system remains firmly in the lab. Meta has not announced any timeline for real-world deployment, and the technology currently requires participants to be in a controlled research environment with access to MEG equipment, which is expensive and not exactly portable.

What this means for investors and the broader tech landscape

The privacy implications are the elephant in the room. A system that can decode what someone is thinking, even if only what they’re trying to type, raises questions that existing regulatory frameworks aren’t equipped to answer. Who owns brain data? How is it stored? Can it be subpoenaed?

The risk for investors is the classic research-to-product gap. Lab results with 35 volunteers in controlled settings don’t automatically translate into viable commercial products. MEG machines cost millions of dollars and require magnetically shielded rooms. EEG is cheaper and more portable, but generally less precise.

The 78% top-performer accuracy suggests that individual brain anatomy and signal quality play a significant role, meaning the technology may work dramatically better for some people than others.

Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.