Meta Achieves Brain-to-Text Decoding Without Surgery Using AI With 61% Accuracy

For years, the development of brain-computer interfaces has been trapped in a rigid tradeoff: achieve high accuracy by drilling into the skull to implant electrodes, or prioritize safety by using external sensors that struggle to decode anything beyond basic commands. Meta's Fundamental AI Research (FAIR) division has now fractured that compromise. With the release of Brain2Qwerty v2, the company has demonstrated a non-invasive system capable of decoding typed sentences directly from brain activity with an average word accuracy of 61%.[1][3]

The breakthrough relies on magnetoencephalography (MEG), a neuroimaging technique that measures the microscopic magnetic fields generated by electrical activity in the brain. Unlike fMRI, which tracks slow blood flow, MEG operates fast enough to capture the millisecond-level timing of individual keystrokes. However, MEG signals are notoriously noisy, which is why previous non-invasive text decoding methods languished at roughly 8% word accuracy.[1][4][5]

To solve the noise problem, Meta discarded hand-crafted signal processing rules in favor of an end-to-end deep learning pipeline. When a participant sits in the MEG scanner and types a sentence, the system feeds the raw magnetic data into a three-stage neural network designed to interpret the continuous stream of neural activity.[3][6]

First, a convolutional encoder translates the raw brainwaves into tokens representing individual characters. Next, an AI aligner groups those characters into words. Finally, a large language model (LLM)—operating much like the predictive text on a smartphone—uses semantic context to clean up the output, guessing the most likely intended word when the neural read is ambiguous or corrupted by noise.[4][6]

The results represent a paradigm shift for non-invasive neurotechnology. Across nine volunteers who each provided 10 hours of typing data, the system reconstructed 22,000 sentences with 61% average word accuracy. The best-performing participant reached 78% accuracy, with more than half of their sentences decoded with one word error or less.[1][2][3][5]

Perhaps more importantly for the future of the field, Meta's researchers observed that the system's accuracy scales log-linearly with the volume of training data. The model showed no signs of plateauing, suggesting that the remaining performance gap between external MEG scanners and surgical implants like Neuralink could be closed simply by feeding the AI more hours of recorded brain activity.[1][3][5]

To accelerate that progress, Meta has open-sourced the training code for Brain2Qwerty under a non-commercial license, while its research partner, the Basque Center on Cognition, Brain and Language, is releasing the underlying datasets. The goal is to provide the broader neuroscience community with the tools to build upon the foundation model and adapt it for clinical use.[1][5]

Despite the software triumph, the system remains tethered to severe hardware limitations. MEG scanners are multi-million-dollar, room-sized machines that require magnetically shielded environments to function. They are highly sensitive to movement, meaning the technology is currently confined to clinical research labs rather than wearable consumer devices.[4][6]

Yet, for the medical community, the proof of concept is profound. For patients suffering from amyotrophic lateral sclerosis (ALS), brain stem strokes, or locked-in syndrome, the risks of surgical implants—including infection, tissue scarring, and hardware degradation—are steep barriers to entry. By proving that AI can extract coherent language through the skull, Meta has illuminated a viable, non-invasive horizon for restoring human communication.[2][6]