Meta has revealed a groundbreaking artificial intelligence model, Brain2Qwerty, capable of decoding brain activity into text with up to 80% accuracy.
This achievement, developed in collaboration with the Basque Center on Cognition, Brain, and Language, represents a significant step in non-invasive brain-computer interface (BCI) research.
The system uses magnetoencephalography (MEG) technology, a non-invasive method for recording brain activity, to analyze neural signals as participants type sentences. However, while promising, the system remains confined to laboratory settings due to technical and logistical constraints.
Meta’s work reflects the growing interest in BCIs, both invasive and non-invasive. The company originally entered the field in 2017 with the ambitious goal of creating a brain-reading wearable device for consumer use.
While the project was discontinued after four years, Meta pivoted to fundamental neuroscience research. According to Jean-Rémi King, the head of Meta’s Brain & AI team, the current goal is to uncover “the principles of intelligence” by studying the brain’s architecture, rather than focusing on commercial products.
This research aligns with broader efforts in the field, such as Neuralink’s work with invasive implants and recent advances in non-invasive techniques like functional magnetic resonance imaging (fMRI).
How Brain2Qwerty Works
At the core of Meta’s achievement is its multi-layered AI model, which deciphers neural activity into readable text. Participants in the study typed sentences while seated in an MEG scanner, which captured their brain activity.
The AI processed this data in three stages: a convolutional module first analyzed raw MEG signals, a transformer model interpreted sentence-level context, and a language model refined the results into text. This approach enabled the system to decode up to 80% of typed characters accurately, setting a new benchmark for non-invasive neural decoding.
MEG significantly outperformed electroencephalography (EEG), another non-invasive technique, in decoding accuracy. While MEG achieved a character error rate (CER) of 32%, EEG’s CER was much higher at 67%. According to King, “MEG significantly narrows the gap between invasive and non-invasive systems.”
The system performed even better for certain participants, achieving error rates as low as 19%, highlighting its potential as a tool for future applications, Meta’s researchers write in their paper:
“Our analyses indicate that this decoding benefits from two main factors. First, the use of MEG signals instead of EEG signals resulted in a two-fold improvement. Second, our deep learning architecture, combined with a pretrained character-level language model, substantially outperforms standard models.
With MEG, Brain2Qwerty reaches a character-error-rate (CER) of 32%, and for the best participants, the model achieves a CER as low as 19%. These findings highlight the narrowing gap between non-invasive and invasive methods of decoding brain activity.”
The Role and Limitations of MEG
Magnetoencephalography is a non-invasive method that records the magnetic fields generated by neuronal activity. Compared to EEG, MEG offers a higher signal-to-noise ratio, enabling more precise neural decoding.
However, the technology is not without challenges. MEG scanners are large, expensive—costing millions of dollars—and require magnetically shielded rooms to operate. Even minor movements by participants can disrupt the accuracy of the recordings.
These limitations make MEG impractical for real-world applications at present. However, emerging technologies like optically pumped magnetometers (OPMs) could offer a path forward, Meta’s researchers explain in their paper.
“Finally, while MEG outperforms EEG, current MEG systems, including the one used in the present study, are not wearable. This, however, may be resolved by the development of new MEG sensors based on optically pumped magnetometers (OPMs).”
These portable MEG devices, still in development, promise to reduce costs and improve accessibility, potentially enabling applications beyond controlled laboratory settings. As Meta’s research shows, such innovations could bring non-invasive BCIs closer to everyday use.
The broader implications of non-invasive technologies like MEG are also evident in global efforts to advance the field. For example, China recently launched initiatives to standardize BCI technologies, reflecting the growing international focus on brain-machine communication.
Exploring Language Production in the Brain
In addition to decoding brain activity into text, Meta’s research delves into how the brain processes language. The team discovered that the brain organizes language production hierarchically.
At the highest level, abstract ideas are transformed into sentences, which are then broken down into words, syllables, and finally individual letters. These transformations are closely synchronized with motor actions, such as the act of typing on a keyboard.
Using MEG, the researchers captured neural activity at millisecond intervals, revealing how thoughts progress into actions.
According to co-author Mingfang Zhang, “Our study shows that the brain generates a sequence of representations that start from the most abstract level of representations – of the meaning of a sentence – and gradually transform them into a variety of actions, such as the actual finger movement on the keyboard.”
This discovery not only advances our understanding of human cognition but also provides valuable insights for designing AI systems that emulate human intelligence.
However, studying these neural dynamics presents challenges, particularly for non-invasive methods. Movements like speaking or typing can heavily distort neuroimaging signals.
Despite these difficulties, Meta’s findings underscore the potential of MEG as a tool for capturing the brain’s linguistic processing in real-time. Their work complements ongoing research in the field, including recent advancements in fMRI-based thought decoding, offering alternative approaches to understanding human communication.
Meta’s Approach vs. Invasive Alternatives
Meta’s focus on non-invasive brain-computer interfaces sets it apart from competitors like Neuralink, which specializes in invasive solutions. Neuralink’s N1 chip, for example, last year enabled a paralyzed patient to control a computer cursor and play chess using only their thoughts.
While such implants provide higher accuracy and faster communication rates, they require brain surgery, limiting their accessibility and raising ethical concerns.
In contrast, Meta aims to develop safer, non-invasive systems, though current technologies like MEG still face significant hurdles. King noted that “Our effort is not at all toward products. In fact, my message is always to say I don’t think there is a path for products because it’s too difficult.”
This research-first approach highlights Meta’s long-term commitment to advancing scientific understanding rather than pursuing immediate commercial applications.
Nevertheless, non-invasive systems have their own limitations. Current MEG setups are cumbersome and expensive, requiring stationary environments that restrict their use to research facilities.
The emergence of portable alternatives like optically pumped magnetometers could bridge this gap, bringing non-invasive BCIs closer to real-world deployment. For now, however, invasive systems like Neuralink’s remain more practical for immediate applications, particularly in medical settings.
