Meta unveils new AI system that can turn your brain activity into text

Meta has announced Brain2Qwerty v2, a new AI system that is designed to read your thoughts and turn it into text. With this, Meta says, you may be able to use your thoughts to type without needing any implants or surgery.

In Short

• Meta’s new AI system can turn your brain activity into text
• Brain2Qwerty v2 said to be up to 78 per cent accurate
• All of this can happen without any surgery or implants, says Meta

Meta wants to read your thoughts. Yes, the tech giant has designed a new AI system, Brain2Qwerty v2, that can read your brain activity and then turn it into text. And, all of this can happen without the need for surgery or implants.

Meta says that the new system can help people with brain lesions, paralysis and other conditions that prevent them from communicating. As a person would be able to communicate just by their thoughts.

This is version 2 of Brain2Qwerty, Meta's research project that came out last year.

Reading brain signals for text?

Brain2Qwerty v2 is designed to decode what a person is typing directly from brain activity without any implant.

Meta says that the system uses magnetoencephalography, or MEG, a non-invasive method that measures the tiny magnetic fields produced by neural activity through a sensor-laden helmet rather than implanted electrodes.

In simple terms, it works through a special helmet, which resembles a big hairdryer, that picks up faint magnetic signals from brain activity. Meta said this matters because the current gold-standard methods for restoring communication, including stereotactic electroencephalography and electrocorticography, rely on brain surgery, making them risky, expensive and difficult to scale.

Unlike Elon Musk’s Neuralink, which requires brain-interface chips to be implanted inside your brain, Meta’s solution can read your brain signals via this special helmet.

How accurate is this?

Meta says that Brain2Qwerty v2 was trained on around 22,000 sentences from nine volunteers, each of whom spent about 10 hours wearing an MEG device while actively typing.

Unlike the first version, which relied on a hand-engineered pipeline to detect specific neural events linked to keystrokes, v2 uses end-to-end deep learning to decode language directly from raw brain signals. That is, while in the past, the AI system could only translate text based on essentially a handbook of signals, Brain2Qwerty v2 allows AI to analyse raw brain signals to figure out patterns itself.

According to the company, this new system reached an average word accuracy of 61 per cent, compared with 8 per cent for other non-invasive methods. For the best-performing participant, it reached 78 per cent word accuracy, with more than half of all decoded sentences containing one word error or less.

The company also fine-tuned large language models on neural data so the system could use semantic and grammatical context to fill gaps when signals were noisy or ambiguous, much like autocorrect predicting what a person meant to type.

Meta states that decoding accuracy also improved with more training data, suggesting the remaining gap with surgical approaches could narrow further through data scaling.

This research forms part of Meta’s Digital Brain Project, which also includes Tribev2 for perception encoding, NeuralSet for processing brain data at scale and NeuralBench for evaluating brain models. The company also has a $5 million fund for open neuroscience datasets.

Meta is not the only company who is working on such a device. Sabi, a Silicon Valley startup, is said to be developing a cap that may also be able to read your brain signals.

- Ends