Brain2Qwerty v2: The 2026 Brain-to-Text Guide

The practical guide to non-invasive brain-to-text decoding, what Meta's Brain2Qwerty v2 actually does, and where the whole field really stands in 2026.

On June 29, 2026, Meta's research team published a system that decoded full sentences from a person's brain, in real time, with 61% word accuracy, and the only thing touching that person was a helmet of magnetic sensors. No surgery. No electrodes under the skull. No wires through the scalp. The model is called Brain2Qwerty v2, and Meta describes it as the highest-performing pipeline ever built for real-time sentence decoding from non-invasive recordings - AI at Meta.

That single number, 61%, is doing a lot of work, and most of the coverage you will read this week gets it wrong in one of two directions. One camp treats it as science fiction arriving early, the moment machines started reading minds. The other dismisses it as a lab toy that will never leave a magnetically shielded room. Both miss what is actually interesting, which is that a non-invasive method just went from 8% word accuracy to 61%, an eightfold jump that the field genuinely did not expect to see this decade - AI at Meta.

Here is the problem this guide solves. Brain-computer interfaces are simultaneously one of the most over-hyped and one of the most genuinely transformative areas in technology, and almost nobody can tell you where the real line sits. The implanted systems that already let a paralyzed man with ALS speak again at 97.5% accuracy are real - UC Davis Health. The consumer EEG headbands that claim to read your thoughts are mostly not. Meta's MEG result sits in a third category that deserves precise understanding, not hype.

This guide breaks down exactly what Brain2Qwerty v2 does, how it works, and why it matters. It then maps the entire 2026 brain-to-text landscape: the invasive implants from Neuralink, UC Davis, Synchron, and Precision Neuroscience that lead on raw accuracy, the non-invasive research from Meta and academic labs that leads on safety, the companies and the money behind all of it, the clinical reality for the patients who need it most, and the new wave of neural-privacy law that is racing to catch up. Read it and you will understand the field better than most of the people writing headlines about it.

Contents

1. The 2026 brain-to-text scorecard
2. What Brain2Qwerty v2 actually is
3. How we got here: Brain2Qwerty v1 and the typing paradigm
4. Why reading a brain through the skull is so hard
5. The real breakthrough is AI, not better sensors
6. The invasive competition: implants that already work
7. The non-invasive field beyond Meta
8. Who this is really for: the clinical reality
9. The money, the valuations, and the hype filter
10. Your brain as data: neurorights and privacy
11. The road ahead: wearable MEG, AI agents, and honest timelines
12. Conclusion: how to think about brain-to-text in 2026

1. The 2026 brain-to-text scorecard

Before going deep on any single system, it helps to see the whole field on one page, scored against the question that actually matters to a non-specialist: how close is this approach to giving a person who cannot speak fast, accurate, safe communication, and how soon? That question is different from "which result is the most scientifically novel," and the difference is the single most important thing to understand about this space. The most headline-grabbing research is often the furthest from a usable product, while the least glamorous engineering is often the closest.

The table below scores the leading approaches on four criteria weighted by what a patient and clinician would care about. Decoding performance (30%) captures raw accuracy, speed, and vocabulary. Safety and non-invasiveness (25%) rewards methods that avoid brain surgery. Portability and practicality (20%) rewards systems that could plausibly be used outside a research facility. Clinical readiness (25%) rewards systems that are actually being used by patients, with regulatory progress, rather than healthy volunteers in a lab. Each score runs 0 to 10, and the final column is the weighted average.

Read this table carefully and a counterintuitive picture emerges. Synchron tops it not because it is the most accurate (it is not, by a wide margin) but because it threads its electrode array up through a blood vessel and into the brain without ever opening the skull, and it is the furthest along the regulatory path toward a real product - Fierce Biotech. Meta's Brain2Qwerty v2, the system that prompted this entire guide, sits at number 7, because for all its scientific importance it has only ever been tested on healthy people who can already type, runs on a scanner that weighs roughly half a ton, and is explicitly described by its own creators as not being on a path to becoming a product - MIT Technology Review.

That gap between scientific significance and clinical readiness is the through-line of this guide. Brain2Qwerty v2 matters enormously as a proof that non-invasive decoding can get good, and almost not at all as something a patient will use next year. Holding both of those truths at once is the entire skill of reading this field correctly. The rest of the guide explains, system by system, why each one lands where it does, starting with the announcement that made this week interesting.

2. What Brain2Qwerty v2 actually is

Brain2Qwerty v2 is a deep-learning model that takes the raw magnetic signals coming off a person's head while they type, and turns those signals into the sentence they were typing, as it happens. The "v2" matters because there was a v1, published in early 2025, that did something similar but slower and less accurately, and the jump between the two versions is the real story. Meta's research team, the FAIR Brain and AI group led by Jean-Remi King in Paris, announced v2 on June 29, 2026, the same week the peer-reviewed version of the original work appeared in Nature Neuroscience - Nature Neuroscience.

The headline number is 61% average word accuracy, computed across nine participants, with the best single participant reaching 78% word accuracy and decoding more than half of all sentences with one word error or fewer - AI at Meta. Stated as an error rate, that is an average word error rate of 39%, which sounds high until you remember the comparison point. Meta frames the prior state of the art for non-invasive sentence decoding at roughly 8% word accuracy, meaning v2 represents close to an eightfold improvement over what other non-invasive methods could do - AI at Meta.

Below is the official announcement from Meta's research account, which is the primary source for everything described in this section. There is no separate launch video, because Meta positions this as a scientific publication rather than a product, so the post itself is the canonical announcement.

To get to that number, Meta did something deceptively simple in description and very hard in practice. The team recorded nine volunteers as each one typed for roughly 10 hours while wearing a magnetoencephalography (MEG) device, collecting about 22,000 typed sentences of paired brain-and-keystroke data in total - AI at Meta. Then they trained an end-to-end neural network to predict, directly from the raw magnetic recordings, what was being typed. The phrase end-to-end is the architectural heart of the upgrade and deserves unpacking, because it is the difference between v1 and v2.

The word accuracy of v2 only becomes meaningful when you place it on the full spectrum of brain-to-text performance, from the old non-invasive baseline through Meta's result to what surgically implanted systems already achieve. The chart below does exactly that, and it is the single most important picture in this guide.

What this chart makes obvious is the thing the celebratory coverage tends to skip. Meta's v2 closed most of the gap between the old non-invasive floor and the invasive ceiling, but a meaningful gap to surgical systems remains, and the 97% from an implanted array is operating on a 125,000-word vocabulary in genuine real-time conversation, while Meta's number comes from typing memorized sentences in a lab - UC Davis Health. The achievement is real and the distance left to travel is also real.

There is one more detail from the v2 announcement that almost everyone overlooks, and it connects this neuroscience milestone directly to the broader AI moment. Meta states that AI agents were deployed to explore optimizations for the decoding pipeline, with engineers selecting the final training configurations manually - AI at Meta. In other words, autonomous AI systems helped design the model that reads the brain. That detail is a quiet signpost for where research itself is heading, a theme this guide returns to at the end.

It is worth being precise about what real-time and end-to-end actually mean here, because those two phrases carry most of the upgrade and are easy to nod past. In v1, decoding happened in separate hand-built stages and could not produce output until an entire sentence was finished, so the system was fundamentally retrospective. In v2, a single neural network maps raw magnetic signal straight to language as the signal arrives, which is what makes the pipeline real-time capable. The practical importance is that real-time decoding is the difference between a system that can transcribe a finished thought and a system that could, in principle, support a flowing conversation. Note the careful wording, though: real-time describes the decoding pipeline, not the hardware. The MEG scanner is exactly as immovable as before. Meta made the software live; the physical constraints that keep this in a lab did not change at all, and conflating the two is the most common mistake in the coverage of this result.

3. How we got here: Brain2Qwerty v1 and the typing paradigm

You cannot understand v2 without understanding v1, because v1 established the paradigm and exposed the limitations that v2 was built to overcome. The original Brain2Qwerty was introduced as a preprint in February 2025 and reached peer-reviewed publication in Nature Neuroscience, under the title that captures its method exactly: decoding text via typing - arXiv. Rather than trying to read abstract "thoughts," which is vague and nearly impossible to label, the researchers anchored the problem to a concrete, measurable behavior: pressing keys on a keyboard.

The experimental setup was careful and a little unusual. The team recruited 35 healthy, right-handed, skilled typists and recorded them at the Basque Center on Cognition, Brain and Language (BCBL) in Spain, with twenty taking part in MEG sessions, twenty in EEG sessions, and five doing both - arXiv. Each participant saw a sentence flashed one word at a time, briefly held it in memory, and then typed it on a custom keyboard built with non-magnetic silver springs so it would not interfere with the magnetic recordings. They typed without backspacing, so every keystroke was a clean training label paired with a slice of brain activity.

The model that learned from this data had three distinct stages, and the figure below from the original paper shows the whole pipeline, from a participant typing under the scanner to the decoded characters that come out the other end. This is the architecture that v2 would later collapse into a single end-to-end network.

The three stages each did a specific job. A convolutional module first cleaned and compressed the raw sensor data, using a spatial-attention mechanism to account for where each sensor sat on the head and a subject-specific layer to handle the fact that every brain is shaped slightly differently, before passing it through an eight-block dilated convolutional network - arXiv. A transformer module then looked across the whole sentence to model the relationships between keystrokes. Finally, a character-level language model, a nine-gram statistical model trained on Spanish Wikipedia, refined the raw predictions by nudging them toward letter sequences that actually occur in real language.

The results revealed the central truth of non-invasive decoding, the one fact you must hold onto through this entire guide: the choice of sensor matters more than almost anything else. With MEG, Brain2Qwerty reached an average character error rate of 32%, and for the best participant that dropped to 19%. With EEG, the much cheaper and more common technology, the error rate ballooned to 67%, more than double - AI at Meta. The same model, the same task, the same people in some cases, and the difference between a usable result and noise came down entirely to which machine was listening.

The figure below from the original paper shows this contrast directly, plotting the decoding error rates for MEG against EEG across the model variants the team tested. The visual gap between the two sensors is not subtle, and it is the empirical core of why Meta built v2 on MEG rather than the cheaper, more portable EEG that most of the consumer neurotech industry relies on.

This sensor gap has a strategic consequence that shapes the entire field. EEG is what fits in a headband, an earbud, or a pair of glasses, which is why almost every consumer brain-reading claim is built on it. MEG is what delivers the clean signal that makes real language decoding possible, but it lives inside a multi-million-dollar shielded room. The whole non-invasive industry is, in effect, caught between a sensor that is portable but too weak and a sensor that is powerful but immovable, and progress depends on whichever side of that gap closes first. Meta bet on the powerful side and is waiting for the hardware to shrink, a bet examined in detail later in this guide.

The original paper was unusually honest about what it could not do, and those admissions are exactly the limitations v2 attacked. The authors stated plainly that the system did not operate in real time, because the transformer and language model needed the entire sentence to finish before producing output - arXiv. They noted the study used only healthy participants who could already type, which is the population that least needs the technology. And they acknowledged that the MEG hardware was not wearable and required controlled lab conditions. Brain2Qwerty v2 directly answered the first of those three: it made decoding real-time and end-to-end. The other two limitations, the healthy-typist paradigm and the immovable hardware, remain wide open, which is why the scorecard places even v2 far from clinical use.

4. Why reading a brain through the skull is so hard

To understand why Meta chose MEG over EEG, why the machine weighs half a ton, and why non-invasive decoding lagged invasive methods for so long, you have to confront the brutal physics of the problem. This is the section most coverage skips, and skipping it is why so much commentary about brain-reading is nonsense. The fundamental obstacle is not artificial intelligence and never was. It is signal strength.

The electrical activity of your neurons produces magnetic fields that are almost unimaginably faint. Cortical activity generates fields on the order of 10 femtotesla, and even the brain's strongest rhythms reach only about 100 femtotesla - Wikipedia. For scale, the magnetic noise of a normal urban environment is roughly 100 million femtotesla, eight to nine orders of magnitude larger than the signal you are trying to detect. Reading the brain non-invasively is like trying to hear a single whisper from across a stadium during a rock concert, except the concert is a billion times louder than the whisper.

This is why both major non-invasive methods fight an uphill battle, and why they make opposite trade-offs. The two technologies that dominate the field each measure a different physical consequence of neural firing, and the differences explain everything downstream.

• EEG (electroencephalography) measures tiny voltages on the scalp using electrodes. It is cheap, portable, and decades old, but the skull and scalp smear and weaken the electrical signal badly.
• MEG (magnetoencephalography) measures the magnetic fields instead, which pass through skull and tissue almost undistorted, giving far cleaner spatial information.

The practical gap between them is large. MEG can localize the source of activity to roughly 2 to 3 millimeters, while EEG manages only 7 to 10 millimeters, and that resolution difference is precisely why Meta's MEG error rates were less than half its EEG error rates - PMC. One widely cited estimate puts the signal fidelity of scalp EEG at only about 5% of the original brain signal after the skull has had its way with it - Frontiers in Neuroscience. The skull is, electrically speaking, a thick blurry curtain, and it is magnetically far more transparent, which is the entire reason MEG wins.

There is a second reason MEG suits language decoding specifically, and it is about time rather than space. Both MEG and EEG capture brain activity with sub-millisecond temporal resolution, far faster than functional MRI, which measures blood flow and lags neural events by several hundred milliseconds - PMC. That speed matters enormously for decoding typing or speech, because the brain produces language as a rapid sequence of events, letters and syllables and words unfolding in tens of milliseconds, and a sensor that blurs time the way fMRI does cannot follow that sequence. This is also why the UT Austin fMRI decoder discussed later recovers only the slow-moving gist of language rather than exact words: its sensor is fast enough to catch meaning but far too slow to catch the precise sequence of characters. MEG sits in the rare sweet spot of carrying both clean spatial information and fast timing, which is exactly the combination that brain-to-text needs, and exactly why Meta's results outrun every other non-invasive approach.

The cost of MEG's clean signal is its monstrous hardware. Conventional MEG uses superconducting sensors called SQUIDs that must be chilled to around 4 kelvin with liquid helium, fixed rigidly in a helmet-shaped vessel a few centimeters from the scalp, inside a room lined with multiple layers of magnetic shielding metal. A full MEG installation including that shielded room costs roughly 2 to 3 million dollars - QuSpin. The subject cannot move, because the moment the head shifts relative to the fixed sensors, the signal is lost. This is the hardware Jean-Remi King was describing when he said the second a head moves, the signal is gone, and why he concluded there is no product path because it is simply too difficult.

There is a genuine escape route from this hardware prison, and it is the most important thing to watch in non-invasive neurotech: optically pumped magnetometers, or OPMs. These are room-temperature quantum sensors, each about the size of a Lego brick, that need no liquid helium and can sit directly against the scalp in a lightweight wearable helmet that moves with the wearer - Nature Scientific Reports. Companies including Cerca Magnetics, FieldLine, and QuSpin are commercializing OPM-MEG, with Cerca's wearable helmet weighing under 3 kilograms against the half-ton of a SQUID system - Cerca Magnetics. OPM-MEG still needs a shielded room, and current systems are sold as research devices without medical approval, but it is the only credible bridge from Meta's lab result toward anything a person could one day wear. If you want to know whether non-invasive brain-to-text will ever escape the lab, the OPM roadmap is the thing to follow, not the next accuracy headline.

5. The real breakthrough is AI, not better sensors

Here is a question worth sitting with from first principles. If the physics of non-invasive recording has not fundamentally changed, the skull is just as opaque as it was a decade ago, the magnetic signals are just as faint, then why did decoding accuracy suddenly leap from 8% to 61%? The answer reframes this entire field. The breakthrough is not in the hardware. It is in the artificial intelligence that learned to extract signal from noise that classical methods threw away. This is the deepest insight in the guide, and it changes how you should read every brain-decoding announcement.

For decades, neuroscientists analyzed brain signals with hand-built statistical methods: averaging many repetitions of the same stimulus, filtering specific frequency bands, fitting linear models. Those methods work when the signal is strong and the structure is simple. They fail catastrophically when the signal is buried under noise a billion times stronger and the structure is the full complexity of human language. What deep learning brought was the ability to learn the right representation directly from data, discovering noise-robust patterns that no human engineer would have thought to specify. This is the same shift that transformed speech recognition when neural networks replaced older statistical pipelines, and the same shift that let AlphaFold crack protein structure prediction by learning evolutionary patterns rather than simulating physics - DeepMind. When intelligence becomes cheap and powerful enough, problems that looked like hardware problems turn out to be inference problems.

Meta's own research history shows this principle compounding over years. In 2022, the team showed it could identify which speech a person was hearing from their brain activity by training a model with a contrastive objective to align brain recordings to the internal representations of a self-supervised speech model, reaching up to 72.5% top-10 accuracy with MEG - Nature Machine Intelligence. In 2023, the same alignment trick let them reconstruct the rough content of images a person was looking at, by matching MEG signals to the embeddings of a self-supervised vision model - AI at Meta. The consistent thread is that the AI does the heavy lifting by mapping noisy brain data into a representation space that a powerful pretrained model already understands.

This contrastive-alignment idea deserves a beat of explanation, because it is the most underappreciated trick in the field. Rather than asking a model to reconstruct language from scratch out of brain noise, which is nearly hopeless, the team aligns the brain recording to the embeddings of an AI model that has already learned a rich understanding of speech, vision, or text from enormous unlabeled datasets. The brain decoder then only has to learn the comparatively easier mapping into that pre-existing space. In effect, the hard part of understanding is offloaded to a model trained on millions of hours of ordinary data, and the brain decoder borrows that understanding. This is why progress in non-invasive decoding tracks progress in self-supervised AI so closely: every time the underlying speech, vision, or language models get better, the brain decoders that lean on them get better too, with no change to the sensors at all. The brain-reading frontier is being pulled forward by the general AI frontier, which is the deepest structural reason to expect non-invasive decoding to keep improving rather than plateau.

The role of language models specifically is the subtle key to brain-to-text. A raw decode of brain activity is garbled, full of plausible-but-wrong characters, exactly the kind of noisy guesswork that a language model is built to clean up. In v1 this was a humble nine-gram statistical model correcting letter sequences. In v2, Meta went further and fine-tuned large language models on neural data, letting the system use semantic context to bridge noisy neural input and coherent language - AI at Meta. The same family of models that powers modern AI assistants, the technology we cover in our guide to building AI agents, is now doing error correction on signals read straight from the cortex. The brain provides a noisy prior about what was meant; the language model provides a strong prior about what is sayable; and the decode lives where those two priors agree.

This AI-first framing also comes with a crucial warning, and it is the cleanest cautionary tale in the field. In 2024, a paper bluntly titled "Are EEG-to-Text Models Working?" examined a string of celebrated results claiming to decode text from cheap EEG and found that many of them were an illusion - arXiv. The models had been evaluated with implicit teacher-forcing, meaning they were fed the correct answer token by token during testing, which inflated their scores by roughly threefold. Worse, when tested properly, several models performed about as well on pure noise as on real brain data, which means they were not decoding the brain at all. They were leaning entirely on the language model's ability to generate fluent text, and the brain signal was nearly irrelevant.

That finding should permanently calibrate how you read this space. A language model is so good at producing plausible sentences that it can create the convincing illusion of mind-reading while contributing almost nothing from the actual brain. The reason Meta's MEG work is credible where many EEG-to-text claims are not is that MEG carries enough genuine signal that the brain is doing real work in the decode, and Meta's evaluation does not hand the model the answer. When you see the next viral claim that someone decoded thoughts from a consumer headband, the question to ask is not "how accurate" but "how much of that came from the brain versus the language model filling in the blanks." For the discredited EEG-to-text results, the honest answer was: almost none.

6. The invasive competition: implants that already work

While Meta pushes the non-invasive frontier, a parallel field has spent years proving that if you are willing to put electrodes inside or against the brain, you can already restore communication at a level that genuinely changes lives. These invasive and minimally invasive systems are the true benchmark against which Brain2Qwerty must be measured, and they are why the scorecard rewards clinical readiness so heavily. The trade-off they embody, raw performance bought with surgical risk, is the central tension of the entire BCI field.

The clearest demonstration of what invasive BCIs can do comes from UC Davis Health and the BrainGate consortium. In 2024, they reported in the New England Journal of Medicine that a man named Casey Harrell, who has ALS and had lost intelligible speech, could communicate again through a 256-electrode array implanted in the speech motor cortex - UC Davis Health. The system reached 99.6% accuracy on a small vocabulary and 97.5% sustained accuracy on a 125,000-word vocabulary, decoding his attempted speech in near real time. A 2026 follow-up reported him using the system across more than 3,800 hours at over 99% accuracy and around 56 words per minute - UC Davis Health. This is not a research demo. It is a man holding conversations with his family through his own brain.

Other invasive milestones fill in the picture. At UCSF, Edward Chang's team gave a woman left without speech after a brainstem stroke a digital avatar that spoke at around 80 words per minute from a 253-electrode surface array, the first system to produce speech and facial expression together - ScienceDaily. And the lineage traces back to a landmark 2021 Stanford result, where a paralyzed participant typed at 90 characters per minute at over 94% accuracy simply by imagining handwriting, with the BCI decoding the intended pen strokes - PMC. These results, achieved years before Meta's MEG headline, are why specialists view non-invasive brain-to-text as promising rather than revolutionary: the bar was already high.

The most famous name in the field, of course, is Neuralink, though its public results are about motor control rather than the speech decoding that UC Davis leads. Neuralink's N1 implant carries 1,024 electrodes on threads thinner than a human hair, inserted by a surgical robot. Its first patient, Noland Arbaugh, paralyzed below the shoulders, has used the device to control a cursor, play chess and video games, and operate his devices by thought, roughly 10 hours a day - Wikipedia. Neuralink's trials have scaled quickly, and the company reported 21 participants enrolled worldwide as of late January 2026, up from around a dozen the previous autumn - US News.

Neuralink also won an FDA Breakthrough Device Designation for speech restoration in May 2025, signaling its ambition to move from motor control into the territory UC Davis already occupies - Neuralink. It is worth noting that a Breakthrough designation is an expedited-review pathway, not an approval or a clearance, a distinction the hype often blurs. The three regulatory milestones mean very different things: a Breakthrough designation is the FDA agreeing to prioritize review of a promising device, a 510(k) clearance like the one Precision earned means a specific component may be marketed for a defined use, and a full premarket approval for a complete implanted speech system is something no company in this field has yet achieved. When a headline says a brain implant was "approved," it is almost always describing one of the earlier, weaker milestones, and reading those terms precisely is the difference between understanding where the field is and being swept up in the marketing.

The deeper reason invasive systems lead, and will keep leading on raw performance for the foreseeable future, comes down to a hard physical number: bandwidth. An electrode sitting directly in cortical tissue can resolve activity at roughly millimeter scale, sample hundreds of times per second, and pick up signals on the order of a hundred microvolts, yielding an information channel that can exceed 200 bits per second - IOPscience. A non-invasive sensor reading through the skull captures a blurred, attenuated shadow of that, with a peak information rate roughly an order of magnitude lower. That bandwidth gap is not a software problem that better AI can fully erase; it is a consequence of where the sensor sits relative to the neurons. AI has narrowed the gap dramatically by extracting more from each non-invasive bit, which is exactly what Brain2Qwerty v2 demonstrates, but the physics floor underneath invasive recording is simply higher, and that is why the people who need the fastest, most accurate communication today still get it from an implant.

The most strategically interesting invasive players may be the ones avoiding open-brain surgery. Synchron threads its Stentrode array up through the jugular vein into a blood vessel adjacent to the motor cortex, achieving a brain interface with no craniotomy at all, and it has implanted around ten patients while preparing a 2026 pivotal trial - Fierce Biotech. Precision Neuroscience took a different minimally invasive route with a 1,024-electrode thin-film array that rests on the cortical surface, and it earned FDA 510(k) clearance for that array in spring 2025, a genuine regulatory clearance rather than a designation - CNBC. Paradromics performed its first human recording with a high-channel-count invasive implant in 2025 - Michigan Medicine. The pattern across these companies is a deliberate retreat from maximum invasiveness toward something a hospital might actually adopt, which is exactly the dimension the scorecard rewards. We trace the broader pattern of AI moving into clinical settings in our guide to applied AI in medicine.

7. The non-invasive field beyond Meta

Meta is the most visible non-invasive decoder, but it is not alone, and the company it keeps tells you a lot about which non-invasive claims to trust. The honest picture splits cleanly into a credible academic frontier, where careful labs report coarse but real results, and a noisy applied layer, where consumer devices make claims that mostly do not survive scrutiny. Knowing which side of that line a given product sits on is the practical skill this section builds.

The strongest non-invasive result outside Meta is the UT Austin semantic decoder, published in 2023 by Jerry Tang and Alexander Huth. Using fMRI rather than MEG, it reconstructs the gist of what a person is hearing, imagining, or even watching in a silent film, producing text that captures the meaning rather than the exact words - NIH. About half the time, the decoded text matched the intended meaning closely - Nature Neuroscience. Crucially, the authors were transparent about its limits and its built-in privacy protection: it required many hours of per-person training in a scanner, worked only on cooperative subjects, and could be defeated entirely by a person who simply thought about something else - ScienceDaily. It is genuine science, and it is also tethered to a room-sized fMRI machine, which is why it sits near the bottom of the scorecard on practicality.

A genuinely new non-invasive modality worth watching sits between the safe-but-weak EEG world and the powerful-but-buried implant world: functional ultrasound. The nonprofit Forest Neurotech began a multi-year safety study of an ultrasound-based interface in 2025, backed by the UK research agency ARIA, on the premise that sound waves can map blood-flow changes deep in the brain at finer resolution than scalp electrodes without penetrating tissue - Digital Health. It is the same general bet that Merge Labs, the Sam Altman-backed startup, is making with its early ultrasound and molecular research. Ultrasound is years behind MEG and EEG in demonstrated language decoding and may never catch them, but it is the clearest example of how the non-invasive field is still searching for the right sensor rather than treating the problem as solved. The honest reading is that no non-invasive modality has yet found the sweet spot of deep, clean, fast, and wearable all at once, and the company that does will reshape the scorecard overnight.

The applied layer is where caution becomes essential, because here the language separates sharply from the reality. Several well-funded companies build non-invasive neurotech, but almost none of them decode language, and the ones that imply they do should be read skeptically. The most honest of them are clear that they measure cognitive state, not thoughts.

• Kernel, founded by Bryan Johnson, sells fNIRS neuroimaging helmets that measure blood-oxygenation as a proxy for activity, useful for research, not thought-reading - Wikipedia.
• OpenBCI ships open-source EEG hardware and the Galea mixed-reality headset for researchers and developers - OpenBCI.
• Neurable licenses EEG built into headphones that measure focus and attention, explicitly not language or text - TechCrunch.

The one applied non-invasive system genuinely aimed at communication is Cognixion, whose Axon headset combines EEG with augmented reality and AI-assisted prediction to help people with ALS and severe motor impairment select words and phrases. It received an FDA Breakthrough Device Designation in 2023 and launched a small clinical trial in 2025 - PR Newswire. Even Cognixion, though, works by helping a user select among options rather than freely decoding sentences, which is why its decoding-performance score is modest even as its safety and portability are high. The takeaway for the whole non-invasive applied layer is consistent: the safe, wearable devices are real, but what they actually do is far narrower than the mind-reading framing suggests, and the gap between those two things is where most public confusion lives.

8. Who this is really for: the clinical reality

It is easy to discuss brain-to-text as a technical contest and forget who the technology exists to serve. Strip away the company rivalries and the funding rounds, and the entire field is ultimately about a specific, devastating human problem: people whose minds are fully intact but who have lost the ability to speak or move. Understanding that population, and how badly current tools serve them, is what makes the stakes legible and what separates a genuine advance from a lab curiosity.

The clearest case is amyotrophic lateral sclerosis (ALS), a progressive disease that destroys the motor neurons controlling voluntary movement while typically leaving cognition untouched. In the United States, roughly 33,000 people were living with ALS as of 2022, with around 5,000 new diagnoses each year - CDC. As the disease advances, many patients lose speech entirely, ending in some cases in locked-in syndrome, where a fully conscious person retains almost no way to signal the outside world. Beyond ALS, brainstem stroke and severe spinal cord injury can produce similar states, and the United States alone has between 257,000 and 388,000 people living with spinal cord injuries, more than half with some degree of tetraplegia - PMC.

The reason brain-to-text matters so much for these patients is that the tools they have today are agonizingly slow. The contrast with natural speech is the whole argument, and it is best seen as a chart.

The numbers behind that chart are stark. People with ALS using eye-gaze typing average under 10 words per minute, and EEG-based P300 spellers often manage only a few characters per minute, while a person in complete locked-in syndrome may be limited to roughly one character per minute - Nature Communications. Natural conversational speech runs 120 to 160 words per minute. That is the chasm the field is trying to cross, and it is why the invasive systems that already hit 56 words per minute feel miraculous to the people using them, and why a non-invasive system, even at 61% word accuracy, is celebrated despite its imperfections.

This clinical lens also exposes the single most important limitation of all the Meta results, the one that the scorecard encodes and that most coverage omits. Every Brain2Qwerty experiment, v1 and v2 alike, was conducted with healthy participants who typed with their own hands. The model was decoding the brain activity that accompanies actual finger movements, which is the very ability that ALS and paralysis destroy. It is an open and genuinely hard scientific question whether a model trained on real typing will work for a person who can only attempt or imagine typing, with no movement at all. Prior attempts to train BCIs for completely locked-in patients have a sobering history of failure. Until non-invasive decoding is demonstrated in the patients who need it, rather than the volunteers who do not, the technology remains a promise rather than a treatment, however impressive the lab numbers look.

The distinction at the heart of this limitation is between motor execution and motor intent, and it is worth making concrete because it is where lab demos and clinical reality most often diverge. When a healthy volunteer types, their brain sends the commands and the fingers actually move, producing a strong, consistent, repeatable neural signature that a model can learn from tens of thousands of examples. A person with advanced ALS or a high spinal cord injury can form the intent to type but generates no movement, and the neural signal that accompanies pure imagined or attempted movement is weaker, more variable, and harder to label, since there is no keystroke to pair it with. The invasive speech systems solved a version of this by decoding attempted speech directly from the speech motor cortex, where even a patient who cannot make a sound still generates rich, learnable activity through an implanted array. Whether a non-invasive sensor reading through the skull can capture enough of that attempted-movement signal to be useful is the single unproven assumption on which the entire clinical promise of Brain2Qwerty rests. It is a tractable research question, not a fantasy, but it has not been answered, and no amount of accuracy on healthy typists answers it.

9. The money, the valuations, and the hype filter

Follow the money in brain-computer interfaces and you learn two things at once: the sector is attracting serious capital, and a large fraction of the numbers floating around it are unreliable. Separating the real signal from the inflated noise is a practical skill for anyone trying to assess this field, and it follows the same discipline we apply to any frothy category, including in our analysis of AI market power consolidation. The rule of thumb is simple: trust disclosed funding rounds, distrust market-size forecasts, and be most skeptical of valuations attached to companies with no product.

The funding numbers are the trustworthy part, because they reflect actual transactions. The standout is Neuralink, which raised roughly 650 million dollars at a 9 billion dollar valuation in mid-2025, more than doubling its prior valuation - SiliconANGLE. Synchron raised a 200 million dollar Series D in late 2025 to fund its pivotal trial - Fierce Biotech. Precision Neuroscience raised around 102 million dollars alongside its FDA clearance. Across the whole sector, Crunchbase put total neurotech venture funding at roughly 896 million dollars in 2024, on track for about 1.4 billion in 2025, driven overwhelmingly by Neuralink's round - Crunchbase News.

The clearest hype flag in the sector right now is Merge Labs, the brain-interface startup co-founded by Sam Altman, which raised a 250 million dollar seed round at an 850 million dollar valuation led by OpenAI in early 2026, despite being an early-stage research lab with no product - TechCrunch. Even sympathetic coverage noted that the company is still in an early research phase, which makes that valuation a bet on the founders and the narrative rather than on anything built - Tom's Hardware. When a pre-product company is valued near a billion dollars on the strength of who is involved, the discipline we describe in our note on the honest truth about AI's impact applies directly: separate the demonstrated capability from the story being sold around it.

The least trustworthy numbers of all are the market-size forecasts, and it is worth saying so plainly because they get repeated everywhere. Estimates for the 2024 BCI market range from about 262 million dollars to 5.2 billion dollars depending on which research firm you ask, a roughly twentyfold spread that signals these are not measurements but guesses dressed up as data - Precedence Research. The only piece of the market-size consensus worth carrying forward is the growth rate, which clusters in the mid-teens percent annually across firms. Any specific dollar figure for the total market should be treated as decoration, not information. This is the kind of inflated, single-sourced claim that the hype filter exists to catch, and the brain-interface sector is full of it.

10. Your brain as data: neurorights and privacy

The moment a machine can read intention from neural signals, even imperfectly, a question appears that no previous technology raised: who owns the data inside your head, and what stops it from being used against you? This is not a distant philosophical worry. It is a fast-moving area of law that went from nonexistent to a patchwork of statutes in roughly four years, and it is the part of the brain-to-text story most likely to affect ordinary people, because the consumer EEG devices collecting brain data already exist and are largely unregulated. Understanding this dimension is part of thinking clearly about data sovereignty in general, a theme we develop in our AI sovereignty guide.

The core reason brain data is different from other personal data is worth stating from first principles. Other sensitive data, your location, your fingerprints, your medical records, describes what you have done or who you are. Neural data can reveal involuntary inner states, inferences about emotion, attention, and reaction that a person cannot consciously control and may not even be aware of revealing. As legal scholar Nita Farahany argues, this threatens a kind of freedom previous law never had to name, a right to cognitive liberty and mental privacy - Duke Today. The danger is not only surveillance but manipulation, the use of inferred mental states to shape preferences and behavior.

The legal response has moved faster than almost any other area of data privacy. Chile led the world, amending its constitution in 2021 to protect mental integrity, and in 2023 its Supreme Court became the first court anywhere to rule on a neuroprivacy case, ordering the EEG company Emotiv to delete a user's brain data - Stanford Law School. In the United States, the action moved to the states. Colorado became the first to fold neural data into a comprehensive privacy law in 2024 - Hunton. California followed with SB 1223, classifying neural data as sensitive personal information alongside genetic data and Social Security numbers, effective January 2025 - California Legislature. Montana went further still in 2025, adding a warrant requirement for law enforcement access to neural data - Cooley.

The regulatory gap these laws are racing to close is alarming when you see it laid out. A 2024 audit of thirty consumer neurotechnology companies by the Neurorights Foundation found that 29 of the 30 allowed essentially unlimited access to users' neural data, with the vast majority reserving the right to transfer brain data to third parties - Neurorights Foundation. Because HIPAA medical-privacy protections do not apply to consumer devices, a roughly 500 dollar EEG headband sold for meditation or focus typically comes with almost no legal safeguards on the brain data it collects - STAT News. Globally, UNESCO adopted the first international ethical framework for neurotechnology in November 2025, though it is non-binding - UNESCO. The practical implication is that the law is being written right now, in real time, while the devices are already on people's heads, and where you live determines how much protection your brain data has.

11. The road ahead: wearable MEG, AI agents, and honest timelines

So where does all of this actually go, and how fast? This is where first-principles reasoning matters most, because the field generates two opposite kinds of bad prediction. The optimists extrapolate the accuracy curve and imagine thought-typing consumer devices within a few years. The cynics point at the half-ton scanner and conclude it will never matter. Both are wrong, and seeing why is the payoff of everything above.

Start with the single most important fact Meta reported about v2's future, because it reframes the optimism correctly. The team found that decoding accuracy improves log-linearly with the amount of training data, which suggests the remaining gap to surgical systems could be partly closed simply by collecting more brain recordings - AI at Meta. This is the same scaling dynamic that drove the rest of modern AI, and it means the accuracy ceiling for non-invasive decoding is probably higher than today's 61% implies. The bottleneck is no longer the core algorithm. It is data and hardware.

The hardware bottleneck is the harder one, and it is where the honest timeline lives. For non-invasive brain-to-text to ever reach a patient's home, the half-ton MEG scanner has to become something wearable, and the only credible path is OPM-MEG. The wearable OPM helmets from Cerca and others are real and improving, but they still require a shielded room, still cost on the order of a million dollars for a full system, and have not yet been shown to do high-quality language decoding outside controlled conditions - Sarah Constantin. A realistic reading is that wearable MEG language decoding is a research program measured in years, not a product measured in quarters, and Meta's own scientists are the first to say so.

It helps to name the specific milestones that would have to fall for non-invasive brain-to-text to become a real treatment, because they form an honest checklist against which to judge future announcements. First, decoding would need to be demonstrated in paralyzed patients attempting movement, not healthy people executing it, closing the motor-intent gap described earlier. Second, the hardware would need to shrink from a shielded-room MEG into a wearable OPM system that works in something closer to ordinary conditions. Third, accuracy would need to climb from today's roughly 61% word accuracy to a level where the output is reliable enough for daily communication, which the log-linear scaling result suggests is reachable with more data but has not been shown. None of these is science fiction, and all three are being actively worked on, but each is a multi-year effort in its own right, and they have to land roughly together for a product to exist. When you see a claim that thought-typing is almost here, check it against this list, and you will almost always find that at least two of the three milestones are still wide open.

It is also worth being precise about what Meta itself wants here, because it is widely misread. Meta's brain-decoding work is basic neuroscience, not a product roadmap. Jean-Remi King has been explicit that there is no product path because the hardware makes it impractical, and the deeper goal is to understand how the brain produces language in order to inform the design of more capable AI - MIT Technology Review. Meta's consumer neural-input ambitions live in a completely separate effort, the wrist-based sEMG band shipped with its AR glasses, which reads muscle signals at the wrist rather than activity in the brain - Meta Reality Labs. Conflating the two is one of the most common errors in coverage of this announcement. The brain work is about understanding minds; the product work is about reading wrists.

The most forward-looking detail of all loops back to where this guide began. Meta noted that AI agents helped optimize the Brain2Qwerty v2 pipeline, and that small admission points at the larger shift reshaping research itself - AI at Meta. Autonomous AI systems are increasingly doing the experimental search, configuration, and iteration that used to consume human researchers, a pattern we explore in our guide to self-improving AI agents and in our look at AI for scientific discovery. The same trend is moving into ordinary work, where platforms such as O-mega let businesses run autonomous agents that learn a tool stack and carry out multi-step processes on their own, the operational sibling of the research agents Meta is now pointing at neuroscience. The interface between human intention and machine action is being rebuilt from both ends at once: agents are learning to act on our behalf from software, while neuroscience inches toward reading intention directly from the brain. This frontier is exactly the kind of territory Yuma Heymans (@yumahey), who built O-mega after the autonomous AI recruiter HeroHunt.ai, spends his time mapping, where cheap intelligence meets the messy physical and biological reality it is trying to interpret.

The likeliest shape of the next few years, reasoning from the structure rather than the hype, is therefore this. Invasive systems will keep leading on accuracy and will reach the first regulatory approvals, with minimally invasive approaches like Synchron's probably crossing the clinical finish line first because they trade a little bandwidth for a lot of safety. Non-invasive decoding will keep improving on the strength of more data and better AI, and will remain stuck in the lab until OPM hardware matures, serving science long before it serves patients. And the neural-privacy framework will keep racing to define the rules before consumer devices make the question urgent. None of that is as thrilling as "Meta can read your mind," but it is what the evidence actually supports, and it is a far more interesting story than either the hype or the dismissal.

12. Conclusion: how to think about brain-to-text in 2026

If you take one decision framework away from this guide, make it this: judge every brain-to-text claim against three axes, and never let one of them stand in for the others. The first axis is accuracy and speed, the raw performance. The second is invasiveness and safety, the cost of getting that performance. The third is clinical readiness, whether real patients are actually being helped. Almost all the confusion in this field comes from collapsing those three into one, treating a high accuracy number as if it meant clinical readiness, or treating non-invasive safety as if it meant the system worked.

Held to that framework, Brain2Qwerty v2 is a landmark on exactly one axis. It is the best non-invasive sentence decoder ever built, real-time and end-to-end, and it proved that AI can pull genuine language out of the faint magnetic whisper of a brain through an intact skull at 61% word accuracy - AI at Meta. On the other two axes it is early-stage science: maximally safe but tested only on healthy typists, and tethered to hardware that cannot leave a shielded room. That is not a criticism. It is what a basic-science result looks like, and pretending otherwise does the work a disservice.

The systems actually restoring communication today are the invasive ones, where a man with ALS speaks to his family at 97.5% accuracy through electrodes in his cortex, and where minimally invasive players like Synchron are closest to turning that capability into an approved product - UC Davis Health. The non-invasive field, Meta included, is where the future safety and accessibility of the technology will be won, on the strength of more data, better AI, and wearable OPM hardware that does not exist yet. Both halves of the field are necessary, and watching only the headline-grabbing one gives you a distorted map.

The honest bottom line is that brain-to-text in 2026 is neither the mind-reading dystopia nor the empty hype its loudest commentators claim. It is a real, fast-moving science with a clear clinical purpose, a genuine breakthrough this week in making the safe approach finally work, a long road of hardware and patient-generalization problems still ahead, and a privacy reckoning arriving in parallel. Read it on those terms and you will not be fooled by the next breathless headline, in either direction, which is the most useful thing this guide can leave you with. For the broader arc of how autonomous AI is reshaping the boundary between human intention and machine action, our guide to the future of the autonomous agent workforce picks up where this one leaves off.

This guide reflects the brain-computer interface and brain-to-text landscape as of June 2026. The field moves quickly, model results and clinical trials change month to month, and several figures here come from announcements and preprints rather than final peer-reviewed publications. Verify current details before relying on any specific number.