Brain2Qwerty: Meta builds the AI that reads your thoughts and turns brain signals into text

There is a question humans have been asking for decades, somewhere between science fiction and philosophy: what if a machine could read our thoughts? For a long time the answer remained in the realm of imagination, between Minority Report, Black Mirror and Philip K. Dick's novels. But in 2026 Meta AI turned that question into a concrete research project, with measurable and published results. It is called Brain2Qwerty and it is an artificial intelligence system that can decode the electrical signals of the human brain and turn them directly into written text: words, sentences, thoughts formed mentally, without the person moving a muscle, without a keyboard, without a voice. Just thought, and the text appears. This is not science fiction, it is the result of years of research into Brain-Computer Interfaces combined with the most recent deep learning architectures developed by Meta's research team, and the medical, technological, ethical and social implications are deep enough to demand a serious, thorough reflection.

Brain2Qwerty, whose name is a direct reference to the QWERTY keyboard that has been the primary interface between the human mind and digital text for over a century, is an end-to-end system that combines brain signal acquisition hardware with AI models for decoding and language generation. Its operation unfolds in four main phases. The first is signal acquisition: the system uses two non-invasive technologies in parallel, EEG with sensors placed on the surface of the skull, cheap and safe but with a noisy signal and low spatial resolution, and MEG, magnetoencephalography, which measures the magnetic fields generated by neural activity with higher signal quality but requires far more expensive and cumbersome equipment. The great challenge at this stage is to isolate the specific signal associated with linguistic thought from the continuous electrical and magnetic noise of the brain, an extremely complex signal processing problem.

The second phase is preprocessing: the raw signal is filtered from noise, normalized, cleaned of muscular and ocular artifacts and time-aligned with the cognitive tasks performed by the subject. Preprocessing quality largely determines the quality of the subsequent decoding. The third phase is the core of the system, neural decoding with AI: the preprocessed signal is passed to a transformer architecture specifically designed to process time series of neural signals, which has learned through supervised training to associate specific patterns of brain activity with specific sequences of characters or words. Training requires calibration sessions in which subjects imagine or think specific words while the system records the corresponding brain activity, thus learning the neural dictionary specific to each individual, because brain patterns vary significantly from person to person. The fourth phase is text generation and correction, in which a language model corrects errors, disambiguates uncertain interpretations and produces coherent text, similarly to a smartphone's autocorrect but operating on brain signals rather than imprecisely pressed keys.

The results published by the Meta team are impressive, not in the sense that the system is already perfect, but in the sense that it demonstrates capabilities that until a few years ago were considered unreachable with non-invasive technologies. Under optimal experimental conditions with calibrated subjects, high-quality MEG signal and a structured mental typing task, Brain2Qwerty reaches decoding speeds of 40-50 characters per minute, not comparable to a physical keyboard but functionally useful for many applications and orders of magnitude better than previous BCI systems. The error rate, measured as Character Error Rate, sits around 8-12% under optimal MEG conditions and 15-25% with EEG, remarkable numbers given the non-invasive nature of the decoding. The supported vocabulary includes thousands of common English words, with performance degrading on rare terms and proper nouns; research in other languages is ongoing. After a calibration session of 30-60 minutes performance improves significantly and keeps refining with use, as the model accumulates more data on the user's specific neural pattern.

The question many are asking is why Meta, a company whose primary business is social networks and digital advertising, is investing significant resources in BCI. The answer is strategic and unfolds on multiple levels. The first is the vision of spatial computing: Mark Zuckerberg has built a narrative in recent years in which the future of computing is not the smartphone but AR/VR glasses and immersive environments where keyboard and mouse are past interfaces, and the ideal input is the most natural one possible, thought itself. The second is the legacy of CTRL-Labs, the non-invasive neural interface startup Meta acquired in 2019 for an estimated 500 million to one billion dollars, which brought in the research team and the technology portfolio on which Brain2Qwerty was built. The third is positioning in AI hardware, with the Ray-Ban Meta Smart Glasses already on the market and the next generation of AR/VR devices in development, to build an ecosystem where BCIs can become the next-generation input interface. The fourth is research as a reputational positioning tool: publishing frontier research on technologies with medical applications positions Meta as a first-tier scientific lab, a significant image asset at a time when the group's public reputation has occasionally been problematic.

Beyond consumer and metaverse applications, Brain2Qwerty has medical potential that is probably its most urgent and impactful application. The most immediate use case concerns people affected by conditions that prevent verbal and motor communication, such as ALS, locked-in syndrome and severe paralysis from spinal cord injuries: for them the ability to communicate through thought is not a tech gadget but a radical transformation of quality of life, and the non-invasive nature of Brain2Qwerty offers a path without requiring neurosurgery. The brain signals acquired can also be used as feedback for neurological rehabilitation protocols, helping patients retrain neural circuits damaged by stroke or brain trauma. On the diagnostic front, the recorded patterns could contain early markers of neurodegenerative conditions such as Alzheimer's, Parkinson's or multiple sclerosis, opening possibilities for early diagnosis that could radically change care pathways. In high-criticality environments like cockpits, operating rooms and power plant control rooms, real-time monitoring of cognitive state through brain signals could detect fatigue, distraction or altered states before they produce errors with serious consequences.

Brain2Qwerty is one of the AI technologies with the deepest and most immediate ethical implications developed in recent years, and the questions it raises are not theoretical. If a system can decode brain signals and turn them into text, who has access to that data? What happens if it is acquired without consent through a wearable device that continuously monitors neural activity? Thought is the last frontier of individual privacy, and Brain2Qwerty in its future applications could put that frontier at risk irreversibly. When the system is calibrated, a personalized neural dictionary is created that is as unique as a fingerprint: who owns it, can it be sold, shared or transferred to third parties? Existing privacy laws, including GDPR, were not written with this kind of biological data in mind. If a system can read thoughts, it can also in principle influence them, and the history of big tech does not inspire unlimited confidence in its ability to resist the temptation to exploit this capability for commercial purposes. There is also the accessibility issue: BCI technologies are expensive, and if Brain2Qwerty becomes a standard interface for future computing, those who cannot afford the hardware will find themselves in an even more marked position of digital disadvantage than today. Finally, cybersecurity: the data acquired literally reflects the content of a person's mind, a breach of this kind of information is not comparable to a stolen password and is a violation of unprecedented depth that requires absolute security standards.

Brain2Qwerty sits in a Brain-Computer Interface research landscape that has accelerated significantly in recent years. Neuralink, the most publicized project led by Elon Musk, has developed an invasive brain implant with thousands of electrodes surgically inserted into brain tissue, offering enormously higher signal resolution than non-invasive technologies: the first human implants announced in 2024 showed impressive results in speed and accuracy, but require neurosurgery and involve clinical risks that limit adoption outside serious medical scenarios. Synchron takes an intermediate approach, with an endovascular device inserted through cerebral blood vessels without craniotomy, with a much lower risk profile than Neuralink but lower signal quality. Blackrock Neurotech, a longtime maker of Utah arrays used in academic research, remains the reference in invasive clinical and experimental settings. In parallel, academic labs like UCSF, Stanford and Duke have shown systems in the last two years capable of decoding imagined speech with growing accuracy. In this context Meta's bet is clear: go all-in on the non-invasive path, sacrificing signal quality in exchange for potential scalability to hundreds of millions of users, a market no surgical solution will ever reach.

For a professional or company working today with AI, agents and automations, what concretely changes with Brain2Qwerty? In the short term, little or nothing: the system is still in the research phase, requires clinical-grade MEG hardware to reach peak performance and there is no consumer product on the market. In the medium term, however, the trajectory is clear. If Meta manages to bring Brain2Qwerty to consumer EEG hardware integrated into AR glasses or light headsets, and if EEG performance keeps improving at the current pace, in the next five to seven years we may face input interfaces designed as an alternative or complement to voice and touch. For software designers this means starting today to think about UX that no longer assumes keyboard and mouse as mandatory inputs, but plans for multiple channels, from speech to gesture and, one day, to thought. For those designing AI agents the picture is even more interesting: an input based on thought drastically reduces the friction between the user's intent and the agent's action, and agent architectures designed to interpret minimal, ambiguous intents will be enormously advantaged.

For the Italian and European context, Brain2Qwerty raises at least three operational questions worth addressing right away. The first is regulatory: the AI Act and GDPR will have to be extended or interpreted to handle a new category of personal data, neural data, which today has no dedicated classification and which in a few years will be the subject of mass collection. Legislators and Data Protection Authorities have a window of time to anticipate the issue, and compliance professionals will do well to include neural data now in forward-looking risk maps. The second is clinical: Italian hospitals and rehabilitation centers working with ALS, stroke and spinal cord injury patients should start evaluating experimental pathways with non-invasive BCIs, both to avoid arriving late compared to other European health systems and to build clinical expertise that will be needed when these tools enter standard practice. The third is industrial: the Italian biomedical instrumentation supply chain has a strong tradition and SMEs in the sector can carve out significant space in next-generation EEG acquisition components, preprocessing software and calibration services, areas where demand will grow rapidly in the coming years.

Brain2Qwerty is a research project, not a commercial product, and it is important not to confuse the two planes. But it is also a concrete demonstration that a question long considered science-fictional has received its first serious answer: yes, it is possible to read thought and turn it into text with non-invasive technology and in a sufficiently accurate way to have functional applications. From this point on, the trajectory is engineering, not discovery, and the timelines shorten. The public conversation on what we want to allow, what we want to prohibit and how we want to protect the intimacy of individual thought has to start now, before technical choices consolidate into widely deployed products and changing the rules becomes much harder. As with every frontier technology, the window between what is possible and what becomes normal is short, and it is closed by whoever gets there first. It is worth deciding together where we want that window to close.