Meta has unveiled a new artificial intelligence system called Brain2Qwerty v2, designed to decode brain activity and convert it into text—without requiring surgery or implanted devices.

The company says the technology could eventually help people living with paralysis or other conditions that impair communication by enabling them to express themselves using only their thoughts.

Translating Brain Signals into Text

According to India Today, Brain2Qwerty v2 is designed to decode what a person is typing directly from brain activity without the need for invasive brain implants.

The system relies on magnetoencephalography (MEG), a non-invasive technique that measures the tiny magnetic fields produced by neural activity using a helmet equipped with highly sensitive sensors rather than surgically implanted electrodes.

The helmet, which resembles a large hair dryer, captures faint magnetic signals generated by the brain while a person types.

Meta says this approach offers a safer alternative to existing communication-restoration technologies such as stereoelectroencephalography (sEEG) and electrocorticography (ECoG), both of which require brain surgery, making them expensive, risky, and difficult to deploy at scale.

Unlike Neuralink, which relies on surgically implanted brain-computer interface chips, Meta's system can read brain activity externally through the specialized MEG helmet.

How Brain2Qwerty v2 Works

Meta says Brain2Qwerty v2 was trained using approximately 22,000 sentences collected from nine volunteers, each of whom spent around 10 hours wearing the MEG device while typing.

Unlike the original Brain2Qwerty model, which relied on manually engineered pipelines to detect specific neural events associated with keystrokes, the new version uses end-to-end deep learning to decode language directly from raw brain signals.

Instead of depending on predefined neural patterns, Brain2Qwerty v2 enables AI to identify meaningful patterns directly from raw brain activity.

Significant Accuracy Improvements

The new system achieved an average word recognition accuracy of 61%, compared with approximately 8% for other non-invasive brain decoding methods.

Among the best-performing participants, word recognition accuracy reached 78%, with more than half of the decoded sentences containing one word error or fewer.

Meta also fine-tuned large language models using neural data, allowing the system to leverage semantic and grammatical context to fill in missing information when brain signals are noisy or ambiguous—similar to how smartphone autocorrect predicts intended words.

The company added that decoding performance continued to improve as more training data became available, suggesting that larger datasets could further narrow the performance gap between non-invasive systems and surgically implanted brain-computer interfaces.