Meta Can Read Your Mind Now. We Should Probably Discuss Its Résumé.

There are some technologies so fundamentally important that the first reaction should be wonder.

Meta's new brain-to-text system is one of them. It's called Brain2Qwerty, and its long-term purpose is genuinely extraordinary: restore communication to people who can no longer speak or move, without requiring electrodes to be surgically implanted in their brains.

That's magnificent. Unfortunately, the company standing next to this particular miracle is Meta. And before we hand Meta access to the electrical activity inside the human nervous system, it seems reasonable to ask how the company handled all the less-sensitive information we already gave it.

The answer isn't especially reassuring.

What Meta Actually Built

Let's separate the real research from the "Facebook can read your thoughts through your Ray-Bans" version, because that isn't what happened.

On June 29, 2026, researchers from Meta and several collaborating institutions published a peer-reviewed paper in Nature Neuroscience titled "Noninvasive decoding of typed sentences from human brain activity."

The system uses magnetoencephalography, or MEG, to measure tiny magnetic fields produced by activity in the brain. Unlike Neuralink and other invasive brain-computer interfaces, it doesn't require electrodes to be implanted in brain tissue.

In the peer-reviewed study, 35 healthy volunteers briefly memorized sentences and then typed them while their brain activity was recorded using either EEG or MEG. The MEG version achieved an average character error rate of 29%. For the best participants, that fell to 18%. The much more accessible EEG recordings performed considerably worse, with an average character error rate of 65%. That research became Brain2Qwerty v1.

Meta simultaneously announced a much stronger successor, Brain2Qwerty v2, trained on approximately 22,000 sentences collected from nine participants who each spent about ten hours inside an MEG system actively typing.

According to Meta's newly released v2 research, the system achieved an average word accuracy of 61%. Its best participant reached 78%, with more than half of that person's sentences decoded with no more than one incorrect word. That 61% average is a huge jump from the roughly 8% word accuracy Meta cites for earlier noninvasive approaches.

There's an important distinction here: the v1 work is what appeared in Nature Neuroscience. The newer and more impressive v2 results were released separately through arXiv and haven't yet received the same peer-reviewed validation. Still, this isn't a parlor trick. The system takes extremely noisy measurements from outside the skull, uses deep learning to identify patterns associated with language production and typing, and then uses a language model to turn those signals into coherent sentences.

That last part is important. The system isn't pulling a pristine transcript directly from someone's consciousness. The language model uses probability and semantic context to correct uncertain neural signals, much like autocorrect reconstructs what it believes you intended to type.

It's neural decoding mixed with language-model inference. Which is both less magical and, in some ways, more interesting.

No, It Cannot Hear Your Secret Thoughts

Brain2Qwerty doesn't currently read an unspoken inner monologue.

Participants weren't sitting in the scanner thinking about lunch while Meta secretly extracted the password to their banking app. They were actively typing known sentences during a tightly controlled experiment after hours of participant-specific data collection. The peer-reviewed study also found that the system relied heavily on neural activity associated with motor control. Its mistakes tended to involve letters located near each other on a QWERTY keyboard, and its learned internal representations reflected which hand and finger would normally press each key.

In other words, it's currently closer to decoding the brain's process of deliberately producing typed language than opening a general-purpose window into the mind.

The equipment is also nowhere near consumer-ready. MEG scanners are expensive, immobile systems containing hundreds of sensors and generally require magnetically shielded environments. We aren't getting Brain2Qwerty in the next Meta Quest update. But none of that makes the accomplishment trivial.

Technologies don't arrive in their final form. They begin as impractical laboratories full of cables, specialized equipment and highly controlled demonstrations. Then the sensors improve, the models get larger, the signal requirements shrink and somebody eventually puts the thing in a headset.

Meta's own results indicate that decoding accuracy improved as the amount of participant data increased, without an obvious performance plateau. The company believes additional scale may close part of the remaining gap between noninvasive systems and implanted devices.

Today, the system decodes the neural signals of a person intentionally typing inside an MEG scanner. The direction of travel is the story.

The Most Intimate Data We Have Ever Created

For years, technology companies have built increasingly detailed models of us from indirect evidence. What we click. Where we go. Who we speak with. How long we pause over an image. Which advertisement makes us hesitate. What time we wake up. When we're lonely, angry, frightened, impulsive or vulnerable.

The platforms don't need to know exactly what we're thinking. They only need enough probabilistic advantage to predict what we might do next.

Neural data collapses some of that distance. A system capable of extracting language from brain activity creates a pipeline that looks something like this: neural signal, then AI interpretation, then language, then inferred intent.

That doesn't mean Meta can read minds today. It means we're beginning to manufacture a new class of data from which increasingly intimate mental states may eventually be inferred.

UNESCO has already recognized the danger. Its 2025 Recommendation on the Ethics of Neurotechnology calls for protection not only of direct neural recordings, but also of non-neural information capable of revealing mental states. It specifically warns against using neurotechnology to manipulate people or compromise autonomy and freedom of thought.

Colorado and California have also moved to classify neural information as sensitive data under their state privacy frameworks. That's a start. It's nowhere close to enough. You can delete a browsing history. You can close an account. You can change a password after a breach. You can't rotate your nervous system. And deleting the raw recording may not be sufficient once its patterns have been incorporated into personalized models, embeddings, training sets or downstream inferences.

This isn't just another privacy setting. It's a question of mental sovereignty.

Now Let's Discuss the Company Holding the Scanner

This would be a serious conversation no matter which company developed the technology. With Meta, we have additional context.

A 2025 Reuters investigation found that Meta had internally projected approximately 10% of its total 2024 revenue, around $16 billion, could come from advertising for scams and prohibited goods. One internal estimate said Meta's platforms were showing users approximately 15 billion higher-risk scam ads every day. Another placed annualized revenue from that higher-risk category at roughly $7 billion.

According to the documents, advertisers Meta's automated systems considered likely fraudulent, but not at least 95% certain to be fraudulent, could be charged higher advertising rates instead of being immediately removed. Meta disputed Reuters' characterization. The company said the 10% estimate was rough and overly inclusive, that it contained legitimate advertising, and that user reports of scam ads had fallen 58% over the preceding 18 months. Fair enough. That response belongs in the record.

So does the original question: how does a company reach the point where suspected fraud becomes a category it can price?

We discussed that previously in "PSA: Meta Knows It's Profiting from Scamming You—and They're Okay With That."

Then came Meta's internal AI rules.

Reuters reviewed an authenticated, more-than-200-page document called GenAI: Content Risk Standards. According to the reporting, it had been approved by Meta personnel across legal, public policy and engineering, including the company's chief ethicist.

Among other things, it permitted Meta's chatbots to engage children in conversations described as romantic or sensual. It contained allowances for demeaning statements about protected groups and for producing false information under certain circumstances.

Meta said the examples involving children were erroneous, inconsistent with its policies and subsequently removed. The company also acknowledged inconsistent enforcement.

Again, that response should be included.

But this wasn't a random chatbot producing one ugly answer after somebody spent six hours trying to jailbreak it. These examples existed inside the document used to define acceptable system behavior. They made it through the organization's policy machinery until reporters asked about them.

As we covered in "Meta Wrote It Down: The Company That Documented Its Own Moral Collapse," the disturbing part wasn't only what appeared in the document. It was how many responsible adults apparently saw it before the public did.

There was also a lawsuit involving Meta's relationship with Netflix. Court filings alleged that Netflix received unusually broad programmatic access to Facebook's Inbox API. Meta has denied that Netflix was given the ability to read people's private conversations, saying the integration allowed users to send and receive Facebook messages about Netflix content from within the Netflix application.

The lawsuit's allegations go further than what Meta acknowledges, and the exact scope of Netflix's access remains disputed in litigation. So "Meta sold everybody's DMs to Netflix" is stronger than the established evidence supports. What's established, and what Meta doesn't really dispute, is that select corporate partners received forms of read, write and delete access necessary to operate those messaging integrations, access most users almost certainly didn't understand in meaningful detail.

That history matters when the next proposed dataset is coming from inside the skull.

Even the comparatively silly story about Meta's internal AI-token leaderboard reveals something about the culture. An employee-built dashboard called Claudeonomics ranked workers by token consumption and awarded titles such as "Token Legend" and "Cache Wizard." More than 60 trillion tokens were reportedly consumed over one 30-day period. The dashboard disappeared after details reached the press, although Meta said the employee removed it voluntarily.

It was goofy, but it illustrated a familiar institutional failure: create a metric, turn it into a competition and watch intelligent people optimize the number instead of the outcome.

Now imagine applying that culture to neural-interface adoption.

More data becomes better models. Better models become competitive advantage. Competitive advantage creates pressure for more users, longer sessions, broader collection and more permissive definitions of consent.

The incentive machine doesn't become virtuous merely because the underlying technology has a noble medical purpose.

This Is Not an Argument Against the Technology

The easiest version of this article would be "Meta bad, mind reading scary."

That version would also be intellectually lazy.

Brain2Qwerty could eventually help people trapped behind paralysis, stroke, ALS or traumatic brain injuries communicate with the world again. If noninvasive systems approach the performance of implanted electrodes, they could avoid the risks of neurosurgery, infection and long-term implant maintenance.

We should want that work to continue.

The answer isn't to stop decoding the brain. The answer is to establish the rules before neural data becomes another product category governed by a 47-page terms-of-service agreement nobody reads.

At minimum, neural data needs protections stronger than ordinary biometric or medical information:

• Explicit, revocable and purpose-specific consent.

• A prohibition on advertising, behavioral targeting and insurance or employment decisions based on neural data or inferred mental states.

• Strict separation between medical research systems and commercial identity or advertising infrastructure.

• Independent audits of collection, retention, model training and deletion practices.

• Rights covering raw recordings, derived features, embeddings and personalized models, not merely the original file.

• A meaningful way to revoke future use of data and models derived from it.

• Serious personal and corporate liability when those protections are violated.

A checkbox saying "I agree" isn't informed consent when the user can't understand what future models may infer from the information being collected today. And no company should be permitted to solve that problem by anonymizing the filename while preserving a neural signature capable of identifying the person it came from.

A Miracle With a Governance Problem

Brain2Qwerty is cool as hell. It's also early, impractical and nowhere close to silently extracting random thoughts from people walking down the street. We should say that clearly because the actual breakthrough is important enough without turning it into science fiction. But we should be equally honest about where this road points.

For most of the internet era, surveillance capitalism operated from the outside in. Companies observed our behavior and used it to construct approximations of our internal state. Neurotechnology potentially works from the inside out.

That requires a category of trust Meta hasn't earned.

The problem isn't that Meta's researchers created something dangerous. They appear to have created something potentially transformative and released substantial portions of the work openly so other researchers can build on it.

The problem is that scientific possibility eventually collides with institutional incentive. Meta has already shown us what happens when safety, privacy or human vulnerability lands on the opposite side of the spreadsheet from revenue and growth. Not once. Not hypothetically. Repeatedly, in internal documents, technical systems, policy decisions and public reporting.

So yes, we should celebrate the science. We should fund it. Expand it. Use it to restore communication and return independence to people who have lost both. But before Meta, or anyone else, turns neural decoding into a consumer platform, we need something stronger than another privacy policy and a promise that this particular category of intimate human data will be handled responsibly.

We've seen that movie. This time, the data is coming from inside the theater.

Sources

Nature Neuroscience — Noninvasive Decoding of Typed Sentences from Human Brain Activity Peer-reviewed Brain2Qwerty research, methodology, participant cohort and accuracy results. https://www.nature.com/articles/s41593-026-02303-2

Meta AI — From Brain Waves to Words: Brain2Qwerty Meta’s announcement covering Brain2Qwerty v2, its training data and reported performance. https://ai.meta.com/blog/brain2qwerty-brain-ai-human-communication/

Meta AI Research — Accurate Decoding of Natural Sentences from Non-Invasive Brain Recordings Research page and technical abstract for Brain2Qwerty v2. https://ai.meta.com/research/publications/accurate-decoding-of-natural-sentences-from-non-invasive-brain-recordings/

Reuters — Meta Is Earning a Fortune on a Deluge of Fraudulent Ads Investigation into Meta’s internal estimates concerning scam advertisements and prohibited-goods revenue. https://www.reuters.com/investigations/meta-is-earning-fortune-deluge-fraudulent-ads-documents-show-2025-11-06/

Reuters — Meta’s AI Rules Let Bots Hold “Sensual” Chats With ChildrenReporting on Meta’s authenticated GenAI Content Risk Standards and the company’s response. https://www.reuters.com/investigates/special-report/meta-ai-chatbot-guidelines/

TechCrunch — Meta Denies Netflix Read Users’ Private Facebook Messages Review of the lawsuit’s Inbox API allegations, Meta’s denial and the documented scope of the Netflix integration. https://techcrunch.com/2024/04/02/meta-again-denies-that-netflix-read-users-private-facebook-messages/

Fortune — Meta’s Internal AI Token Leaderboard Reporting on Claudeonomics, tokenmaxxing and the employee-created leaderboard that tracked AI consumption.https://fortune.com/2026/04/09/meta-killed-employee-ai-token-dashboard/

UNESCO — Recommendation on the Ethics of Neurotechnology International framework addressing neural data, mental privacy, autonomy and manipulation. https://www.unesco.org/en/legal-affairs/recommendation-ethics-neurotechnology

---

Rich Washburn is a technologist, strategist, and Founder & Chief AI Architect of ARIA AI Labs, working at the intersection of AI, infrastructure, communications, and capital. He also serves as Managing Partner and Chief AI Officer at Eliakim Capital.