From Brain Waves to Words: Meta’s Brain2Qwerty v2 Reads Your Mind Without a Single Cut
Published on July 9, 2026 — By Vito Ruocco
Imagine sitting in a room, wearing what looks like a high-tech helmet covered in sensors, and simply thinking the words you want to write. No keyboard. No voice assistant. No surgery. No implant drilled into your skull. Just your brain, an AI model, and a screen that fills with text as if by magic.
That scenario is no longer science fiction. Meta AI, in collaboration with the Basque Center on Cognition, Brain, and Language (BCBL), has just unveiled Brain2Qwerty v2 — the most advanced non-invasive brain-to-text decoding system ever built. It reads magnetic signals from your brain, feeds them through a deep learning pipeline, and reconstructs full sentences with an accuracy that was, until very recently, only achievable with electrodes surgically implanted directly onto the cortex.
This is a landmark moment for neuroscience, for artificial intelligence, and — potentially — for the millions of people worldwide who have lost the ability to communicate due to brain lesions, neurodegenerative diseases, or paralysis. But it also raises profound ethical questions about who controls this technology, what it can be used for, and whether the same tools designed to restore speech could one day be turned into instruments of surveillance and commercial exploitation.
What Is Brain2Qwerty v2? The Technical Breakthrough
Brain2Qwerty is Meta AI’s end-to-end deep learning pipeline capable of decoding sentences directly from non-invasive brain recordings — specifically, magnetoencephalography (MEG). MEG measures the tiny magnetic fields generated by the electrical activity of neurons in the brain. Unlike EEG, which captures electrical signals through the scalp and suffers from poor spatial resolution, MEG offers far better precision because magnetic fields pass through the skull with minimal distortion. However, MEG still falls short of the resolution provided by invasive techniques like electrocorticography (ECoG) or stereo-EEG, which require craniotomies.
The v2 release builds on last year’s v1 by replacing hand-crafted feature extraction pipelines with an end-to-end deep learning architecture that learns directly from raw brain signals. The pipeline works as follows:
- Signal Acquisition: Participants wear a MEG device that records magnetic brain activity at hundreds of sensors simultaneously, sampling at millisecond temporal resolution.
- Deep Learning Decoding: A neural network — trained end-to-end — processes the raw MEG signals and maps them to character-level predictions, without any intermediate feature engineering.
- Language Model Integration: The decoded character stream is then passed through a fine-tuned large language model (LLM) that leverages semantic context to correct errors, disambiguate noisy predictions, and produce coherent sentences.
- AI Agent Optimization: Meta deployed AI agents to explore the hyperparameter space and identify optimal training configurations, though final engineering decisions were made by human researchers.
The key innovation here is the combination of end-to-end deep learning with LLM-based language modeling. Previous non-invasive approaches relied on rigid, hand-tuned preprocessing pipelines that threw away most of the signal’s information content. Brain2Qwerty v2, by contrast, lets the neural network discover the optimal representation of brain signals on its own — and then uses the LLM as a powerful “auto-correct” that knows what words make sense together.
The Numbers: 61% Average, 78% Best-Case — A Quantum Leap
The performance figures from Brain2Qwerty v2 are nothing short of staggering when placed in context:
- 9 volunteer participants, each recorded for approximately 10 hours while actively typing sentences.
- ~22,000 sentences used for training the decoding pipeline.
- 61% word accuracy rate on average across all participants.
- 78% word accuracy rate for the best-performing participant.
- 8% word accuracy rate — the benchmark for previous non-invasive methods, as reported in Nature Neuroscience.
Let those numbers sink in. The previous state-of-the-art for non-invasive brain-to-text decoding was 8%. Brain2Qwerty v2 pushes that to 61% on average — a 7.6x improvement. And for the best participant, 78% means that more than half of all decoded sentences had one word error or less. We’re talking about sentences that are essentially readable, often perfect, reconstructed purely from magnetic brain signals picked up outside the skull.
Even more encouragingly, Meta’s researchers found that decoding accuracy improves log-linearly with data volume. This means that the gap between non-invasive and invasive approaches could potentially be closed simply by collecting more data — without any new architectural breakthroughs required. It’s a finding that suggests the current limitations are far from fundamental.
For comparison, invasive approaches using ECoG and stereo-EEG — which require opening the skull and placing electrodes directly on or inside the brain — still outperform non-invasive methods. But the margin is shrinking, and the scalability argument is overwhelmingly in favor of non-invasive techniques. After all, you can’t ask millions of patients to undergo brain surgery just to communicate.
The Open Source Release: Code, Data, and a New Era of Collaboration
In a move that deserves genuine praise, Meta AI has released the full training code for both Brain2Qwerty v1 and v2 as open source. Their partner, BCBL, has simultaneously released the v1 dataset on HuggingFace (accessible via the bcbl190626/SpanishBCBL repository). This is a significant departure from the typical “paper + press release” model that dominates industry AI research, where code and data are often kept proprietary.
The open release includes several interconnected components that form Meta’s broader “open foundational models of the brain” initiative:
- Brain2Qwerty v1 & v2 Code: The complete training and inference pipelines, allowing researchers worldwide to reproduce the results and build upon them.
- BCBL Dataset: The neural recordings from the v1 study, made available to the scientific community for further analysis and model development.
- NeuralBench: A unified benchmarking framework for NeuroAI models, accompanied by NeuralBench-EEG v1.0 — a large benchmark including 36 EEG tasks and 14 deep learning architectures evaluated across 94 datasets. As described in the arXiv paper (2605.08495), NeuralBench reveals that current foundation models only marginally outperform task-specific models, and that many cognitive and clinical prediction tasks remain highly challenging even for the best systems.
- NeuralSet: A Python framework that unifies the processing of diverse neural recordings — fMRI, M/EEG, iEEG, fNIRS, EMG, and spike trains — alongside the embedding of naturalistic stimuli like text, audio, and video. Published as arXiv 2605.03169, NeuralSet solves a critical infrastructure problem: the neuroscience software ecosystem has been fragmented by modality, with tools like MNE-Python, EEGLAB, FieldTrip, and Brainstorm each serving different communities but rarely interoperating. NeuralSet provides a single PyTorch-ready interface that scales from laptop prototyping to high-performance cluster execution.
- Digital Brain Project: A $5 million fund coordinated by the Rothschild Foundation Hospital and the University of Montreal, aimed at assembling the data necessary to build a functional model of the human brain during intelligent behavior. The project plans to record 10,000 hours of standardized brain recordings from multiple labs, using fMRI, MEG, and intracranial EEG, with all de-identified data released for open scientific research.
This is, without exaggeration, the most ambitious open neuroscience AI infrastructure ever released by a major technology company. Whether Meta’s motivations are purely altruistic or strategically self-serving (more on that below), the practical effect is that researchers worldwide now have access to tools and data that would have taken years to develop independently.
TRIBE v2: Simulating the Human Brain’s Response to Media
Brain2Qwerty doesn’t exist in isolation. It’s part of a broader research portfolio at Meta AI that includes TRIBE v2 — a predictive foundation model trained to understand how the human brain processes complex stimuli like images, audio, video, and text.
Building on the Algonauts 2025 award-winning model (which was trained on low-resolution fMRI recordings from just four individuals), TRIBE v2 leverages a massive dataset of over 700 healthy volunteers who were presented with a wide variety of media, including images, podcasts, videos, and text. The model can reliably predict high-resolution fMRI brain activity — and crucially, it enables zero-shot predictions for new subjects, languages, and tasks. This means that given a new person’s brain scan (even one the model has never seen before), TRIBE v2 can predict how that person’s brain would respond to a novel stimulus.
Meta describes this as creating “a digital model of the human brain” that allows researchers to “rapidly test hypotheses about its underlying functions without the need for human subjects in every experiment.” The model weights and code have been released under a CC BY-NC license, and a demo website is available for interactive exploration.
The implications are enormous. If you can accurately predict how a human brain will respond to a given image, video, or piece of music, you can — in principle — optimize content to maximize engagement, emotional response, or attention. You can reverse-engineer what makes a viral video captivating, what makes an advertisement persuasive, what makes a social media feed irresistible.
And this is where the story takes a darker turn.
The Ethical Firestorm: Medical Promise vs. Commercial Peril
Meta’s blog post emphasizes the medical potential of Brain2Qwerty: restoring communication for the millions of people suffering from brain lesions that prevent them from speaking. This is a real and urgent need. Conditions like ALS, locked-in syndrome, severe stroke, and traumatic brain injury can trap a fully conscious mind inside a body that cannot communicate. For these patients, a non-invasive brain-to-text system — even at 61% accuracy — could be life-changing.
But critics, including the creator of the YouTube Short that brought this story to wider attention (Simone Rizzo, @simone_rizzo98), have raised a pointed and uncomfortable question: What happens when the same technology is applied to people who don’t need it — and for entirely different purposes?
The concerns fall into three main categories:
1. Meta Ads and the Attention Economy
Meta’s core business is advertising. The company’s revenue model depends on understanding what captures human attention and predicting what content will keep users scrolling, clicking, and engaging. TRIBE v2 — a model that can predict how the human brain responds to visual, auditory, and linguistic stimuli — is, from a certain angle, the ultimate advertising research tool. If you can simulate a brain’s response to an ad before showing it, you can optimize that ad for maximum neural engagement. You can A/B test at the level of individual neural circuits.
Brain2Qwerty itself, while currently framed as a communication tool, could in principle be used to decode what people are thinking while they browse social media. Imagine a future where a wearable MEG device (and Meta is already investing heavily in wearable hardware through its Reality Labs division) can read your neural responses to content in real time, feeding that data back into the recommendation algorithm. Not just “did you click on this post?” but “did your brain light up with interest before you even consciously decided to engage?”
2. Reverse Engineering the Brain for Profit
The open-source release of Brain2Qwerty, NeuralBench, NeuralSet, and TRIBE v2 — while commendable from a scientific standpoint — also means that these tools are available to any corporation, government, or bad actor with the resources to deploy them. The same end-to-end decoding pipeline that can reconstruct a paralyzed patient’s intended speech could, in a less benevolent context, be used to extract information from a person’s brain without their full understanding or consent.
Current MEG devices are bulky, room-sized machines that require a magnetically shielded environment. But the trajectory of technology is clear: what is bulky today becomes wearable tomorrow, and invisible the day after. EEG headsets are already consumer products. MEG-on-a-chip research is advancing rapidly. The question isn’t whether non-invasive brain decoding will become portable — it’s when.
3. Social Media Addiction, Neurologically Optimized
Perhaps the most insidious concern is the combination of TRIBE v2’s brain-response prediction capabilities with Meta’s existing social media infrastructure. If the company can model how a human brain responds to content — and if future wearables can measure those responses in real time — the feedback loop between content generation, neural measurement, and engagement optimization becomes closed. The algorithm wouldn’t just optimize for clicks and likes; it would optimize for neural engagement — the literal, physical response of your brain to what you see on screen.
This is the scenario that the YouTube Short by @simone_rizzo98 warns about: a future where the technology that could give voice to the voiceless is instead primarily used to make social media more addictive, to extract more attention from more brains for more profit.
Invasive vs. Non-Invasive: Closing the Gap
To understand why Brain2Qwerty v2 matters so much, it’s worth examining the landscape of brain-computer interfaces (BCIs) in detail.
Invasive BCIs — such as those developed by Neuralink, BrainGate, and various academic labs — require neurosurgery to implant electrodes either on the surface of the brain (ECoG) or deep within its tissue (stereo-EEG, depth electrodes). These approaches offer unparalleled signal quality because the electrodes are in direct contact with neural tissue. In clinical studies, patients with implants have achieved remarkable feats: typing at speeds of up to 90 characters per minute, controlling robotic arms, and even partially restoring vision.
But invasive BCIs have fundamental scalability problems:
- Surgery carries risks of infection, bleeding, and brain damage.
- Electrode degradation over time means the signal quality decays, requiring additional surgeries.
- The cost and complexity of neurosurgery put these solutions out of reach for the vast majority of patients, especially in low-resource settings.
- Regulatory hurdles for implantable devices are enormous.
Non-invasive BCIs — using EEG, MEG, fMRI, or fNIRS — avoid all of these problems. They require no surgery, have no risk of infection, are theoretically scalable to anyone who can wear a helmet, and face much lower regulatory barriers. Their Achilles’ heel has always been signal quality: the skull and scalp attenuate and scatter electrical signals (for EEG) and even magnetic signals (for MEG, though to a lesser extent), making it far harder to decode fine-grained neural information.
This is exactly why Brain2Qwerty v2’s 61–78% accuracy is so significant. It proves that the performance gap between invasive and non-invasive approaches is not an unbridgeable chasm. With better algorithms, more data, and larger models, the gap may close entirely — and the non-invasive approach would win on every dimension: safety, cost, scalability, and accessibility.
Meta’s finding that accuracy improves log-linearly with data volume is particularly important. It means that the path to clinical-grade non-invasive decoding is primarily a data collection problem, not a fundamental physics problem. With enough data — and the Digital Brain Project’s plan to record 10,000 hours of standardized brain recordings is a start — the performance could reach levels sufficient for practical clinical use.
The Bigger Picture: NeuroAI as a Scientific Revolution
Brain2Qwerty v2, TRIBE v2, NeuralBench, NeuralSet, and the Digital Brain Project are not isolated products. Together, they represent Meta AI’s attempt to build a comprehensive infrastructure for what is increasingly called NeuroAI — the intersection of neuroscience and artificial intelligence.
The thesis is straightforward: understanding the brain makes AI better, and better AI helps us understand the brain. Large language models, for instance, are already being used to predict brain responses to language — and the similarities between transformer architectures and cortical processing are striking. Conversely, brain data provides a unique training signal for AI: it tells us not just what the model predicts, but how a biological system with billions of years of evolutionary optimization processes the same information.
NeuralBench’s finding that current foundation models “only marginally outperform task-specific models” on brain tasks is a humbling reminder that we are still in the early days of NeuroAI. Despite the impressive performance of Brain2Qwerty v2, the field as a whole has not yet found its “GPT moment” — the point where a single architecture generalizes across all brain-related tasks with decisive superiority.
NeuralSet’s attempt to unify the fragmented neuroscience software ecosystem is equally important. For decades, researchers using EEG, MEG, fMRI, and intracranial recordings have worked in separate software silos, with incompatible data formats, preprocessing pipelines, and evaluation metrics. By providing a single PyTorch-ready interface that scales from laptop to cluster, NeuralSet could do for neuro-AI what HuggingFace did for natural language processing: dramatically lower the barrier to entry and accelerate the pace of innovation.
The Digital Brain Project, with its $5 million fund and ambitious plan to record 10,000 hours of standardized brain activity during complex cognitive tasks, represents the data side of the equation. If NeuralBench is the benchmark and NeuralSet is the infrastructure, the Digital Brain Project is the fuel — the massive, standardized, multi-modal dataset that could train the next generation of brain models.
The Road Ahead: Promise and Peril in Equal Measure
Brain2Qwerty v2 is a genuine scientific achievement. It pushes the boundaries of what non-invasive neurotechnology can do, it’s being released openly for the benefit of the global research community, and it points toward a future where people with severe communication disabilities might regain their voice without undergoing brain surgery. These are unambiguously good things.
But the technology exists within a broader context that cannot be ignored. Meta is not a neutral research institution — it is a for-profit corporation whose primary revenue stream is advertising, and whose track record on privacy, consent, and the ethical use of user data is, to put it charitably, checkered. The same company that is open-sourcing brain decoding models is simultaneously building the world’s most sophisticated social media engagement algorithms, investing billions in wearable hardware, and developing AI systems that can predict how human brains respond to content.
The question is not whether Brain2Qwerty and TRIBE v2 will be used for medical purposes — they will, and that’s wonderful. The question is whether they will also be used for purposes that are far less wonderful. And the answer, based on everything we know about how technology evolves in market economies, is almost certainly yes.
What we do about it — what regulations we enact, what oversight mechanisms we build, what public discourse we foster — is the real challenge. The technology is here. The conversation about its ethical use needs to happen now, not after it has been deployed at scale in ways we cannot undo.
As Simone Rizzo’s YouTube Short provocatively puts it: Meta has developed an AI model capable of reading our thoughts. The question is no longer can they? — it’s what will they do with it?
Key Takeaways
- Brain2Qwerty v2 decodes brain waves into text using MEG + deep learning + LLMs, achieving 61% average word accuracy (78% best participant) — a 7.6x improvement over previous non-invasive methods (8%).
- The system was trained on ~22,000 sentences from 9 volunteer participants, each recorded for 10 hours.
- Meta has open-sourced the full training code for v1 and v2, and BCBL has released the v1 dataset.
- TRIBE v2 is Meta’s predictive foundation model for brain responses, trained on data from 700+ volunteers, capable of zero-shot prediction of high-resolution fMRI activity.
- NeuralBench provides a unified benchmarking framework for NeuroAI models, with NeuralBench-EEG v1.0 including 36 tasks and 14 architectures across 94 datasets.
- NeuralSet is a Python framework that unifies neural data processing across all modalities (fMRI, M/EEG, iEEG, spikes) with a scalable PyTorch interface.
- The Digital Brain Project is a $5M initiative to build a functional model of the human brain during intelligent behavior, with 10,000 hours of planned recordings.
- Ethical concerns center on the potential use of brain-decoding technology for advertising optimization, reverse engineering neural responses for profit, and deepening social media addiction.
- Decoding accuracy improves log-linearly with data volume, suggesting the gap between invasive and non-invasive methods can be closed with more data.
What do you think about Brain2Qwerty and the future of brain-to-text technology? Is Meta’s open-source approach a genuine gift to science, or a strategic move with strings attached? Share your thoughts in the comments below.
Follow the conversation: Official Meta AI Blog Post | YouTube Short by @simone_rizzo98 | Digital Brain Project | NeuralBench Paper | NeuralSet Paper
— Vito Ruocco, July 9, 2026