Meta unveils non‑invasive BCI that types from neural signals

Meta announced Brain2QWERTY, a non‑invasive brain‑computer interface that translates raw EEG signals into typed text in near‑real time [Meta Blog].

The prototype uses a 64‑channel EEG cap sampling at 1 kHz and runs a 12‑million‑parameter transformer on Meta’s Edge TPU. Median latency is 180 ms per character. In a closed‑beta with ten participants the system recorded a 15% word‑error rate on a 100‑word dictation benchmark, beating the previous best non‑invasive result of 22% [Meta Blog].

The cap‑based design removes the need for surgical implants, allowing deployment in ordinary office environments without medical clearance. For users with motor impairments, the system offers a text‑entry method that bypasses voice or eye‑tracking, replacing assistive devices that cost thousands of dollars. Running a transformer on a consumer‑grade Edge TPU shows that high‑bandwidth neural decoding can be handled on mass‑produced hardware, forcing competitors to reassess their hardware roadmaps.

Editor’s take Meta’s move is less about pioneering neuroscience and more about securing a foothold in the emerging BCI market before Apple or Google ship comparable products. By publishing the full architecture, Meta forces the ecosystem to standardize on cap‑based data pipelines, raising privacy concerns around raw brain data. The real test will be whether developers can build applications that make the 15% error rate acceptable for everyday tasks.

Reader poll Which hands‑free input technology would you rely on for daily work?

• Meta’s Brain2QWERTY BCI
• Voice assistants (e.g., Alexa, Siri)
• Eye‑tracking devices (e.g., Tobii)
• Traditional keyboard