# Brain2Qwerty

> Source: https://aiwiki.ai/wiki/brain2qwerty
> Updated: 2026-07-24
> Categories: AI Research, Deep Learning, Healthcare AI
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/)
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution to "AI Wiki (aiwiki.ai)".

**Brain2Qwerty** is a research program for decoding typed language from non-invasive recordings of brain activity. [Meta AI](/wiki/meta_ai) introduced Brain2Qwerty v2 on June 29, 2026 as an end-to-end system that converts continuous magnetoencephalography (MEG) recordings into complete English sentences.[1][2] The system was tested on healthy volunteers who listened to sentences and physically typed them from memory inside an MEG scanner. It was not tested as a consumer product, clinical communication aid, or system for reading arbitrary thoughts.

V2 is distinct from the first Brain2Qwerty study. The earlier system depended on the known time of every keypress, while v2 attempts to infer character timing and sentence content from the continuous recording itself. Meta released code for both versions, but the v2 study remained a preprint and its underlying dataset was still embargoed at the July 2026 research cutoff.[3][6]

## History and institutions

Brain2Qwerty began as a collaboration between Meta's Fundamental AI Research organization and the Basque Center on Cognition, Brain and Language (BCBL). Meta publicly described the first study on February 7, 2025. Its manuscript, then titled "Brain-to-Text Decoding: A Non-invasive Approach via Typing," appeared on arXiv on February 18.[5][9]

The revised work was published in Nature Neuroscience on June 29, 2026 under the title "Noninvasive decoding of typed sentences from human brain activity."[4] That peer-reviewed article describes what the later project site calls Brain2Qwerty v1. On the same date, Meta announced Brain2Qwerty v2 and posted a separate paper titled "Accurate Decoding of Natural Sentences from Non-Invasive Brain Recordings."[1][2]

The v2 author list includes researchers affiliated with Meta AI, BCBL, Ecole Normale Superieure and CNRS, Hospital Foundation Adolphe de Rothschild, Univ. Lille and SCALab, Paris Cite University, Inria and Universite Paris-Saclay, and CEA NeuroSpin.[2] Although the two papers were published or posted on the same date, only the v1 paper had completed peer review.

## From v1 to v2

The first study included 35 healthy Spanish-speaking volunteers across MEG and electroencephalography (EEG), with five people taking part in both recording groups. Its final analysis covered 19 unique MEG participants and 20 EEG participants. Each person read Spanish sentences word by word, held the sentence in memory, and typed it without seeing the text.[4]

V1 divided the recording into 500-millisecond windows centered on known keystroke onsets. A [deep learning](/wiki/deep_learning) model predicted characters from those aligned windows, and a character language model helped form a sentence. This meant the decoder needed the time of every physical keypress as an external input. In the peer-reviewed results, mean character error rate was 29% for MEG and 65% for EEG, with an 18% character error rate for the best MEG participant.[4]

V2 changed both the input and the scale of the experiment. It used MEG only and accepted a continuous recording covering a whole sentence, without keypress timestamps. Nine English-speaking volunteers each contributed about ten hours of recordings, roughly ten times the data collected per person in v1. The project describes about 22,000 typed sentence trials in total.[2][3]

The project site summarizes v1 at 40% mean word accuracy and 48% for its best participant, compared with 61% and 78% for v2.[3] Those figures are useful as a project-level comparison, but they do not make the experiments directly equivalent. The versions used different languages, participant groups, data volumes, model inputs, and primary metrics.

## Study design

The nine v2 participants were healthy, right-handed adult volunteers between 23 and 56 years old. They were native English speakers and proficient typists. Each completed ten sessions of about one hour at BCBL in Spain, producing approximately 90 recording hours across the cohort.[2]

On each trial, the participant heard a sentence through scanner-compatible headphones. After a 1.5-second fixation interval, the participant typed the sentence from memory on a custom non-ferromagnetic QWERTY keyboard. Backspace was disabled, and no typed letters appeared on screen. A small visual cue acknowledged each keystroke without revealing what had been entered.[2]

The stimulus pool began with 20,000 simple English sentences generated by Llama 4 and filtered for the task. Each session contained 256 sentence trials. The final analysis included 2,724 unique sentence texts, including pilot material. The authors split unique text into 80% training, 10% validation, and 10% test sets by a deterministic hash, preventing the same sentence text from appearing in more than one split.[2]

Recordings came from a 306-channel MEG system with 102 magnetometers and 204 planar gradiometers. The device sampled at 1 kHz; preprocessing filtered the signal and reduced it to 100 Hz for modeling. Participants gave written consent, and the paper reports ethics approval, deidentified or anonymized records, and no access to personally identifying information.[2]

## Architecture

Brain2Qwerty v2 uses three stages. First, a neural encoder with spatial and convolutional processing plus a four-layer Conformer applies connectionist temporal classification. It predicts lowercase characters, spaces, and blank tokens while learning where characters occur in the continuous signal.[2]

Second, an aligner divides the encoder representation at predicted word boundaries and maps those segments toward the embedding space of a [large language model](/wiki/large_language_model). Training uses dynamic time warping and a contrastive objective to associate neural segments with target words.[2]

Third, a neuro-conditioned language model receives both the imperfect character transcript and the MEG-derived word tokens. The final reported configuration used [Qwen3](/wiki/qwen_3)-4B with low-rank adaptation. It then generated the sentence autoregressively.[2] Removing the neural word tokens degraded the reported results, suggesting that the final model used more than text correction alone.

Despite Meta's use of "online" and "real-time" language, the published architecture is not a causal streaming decoder. It processes the complete sentence segment and cannot display words before the participant finishes the sentence. The paper identifies causal operation and lower latency as future work.[1][2]

## Results

The paper evaluates literal transcription with character error rate (CER) and word error rate (WER), and meaning preservation with semantic error rate (SemER). Lower values are better. Rounded cohort means were:[2]

| System | CER | WER | SemER |
|---|---:|---:|---:|
| Neural encoder | 0.28 | 0.55 | 0.096 |
| Encoder plus character n-gram | 0.26 | 0.43 | 0.085 |
| Brain2Qwerty v2 | 0.31 | 0.39 | 0.059 |

V2 reduced word and semantic error relative to both baselines, but its character error was worse. Its mean WER of 0.39 corresponds to the 61% word accuracy highlighted by Meta. The best participant reached 0.22 WER, or 78% word accuracy. For that participant, 28% of test sentences were exactly correct and 47% were within one word insertion, deletion, or substitution.[1][2]

The metric tradeoff matters. The language model can repair a noisy character sequence when context is strong, improving sentence-level meaning. It can also replace uncertain input with fluent but incorrect language. The paper includes examples in which output for a lower-performing participant is grammatical but unrelated to the target. Brain2Qwerty therefore cannot be evaluated by fluency alone.[2]

Performance also varied considerably among the nine participants. The authors found continued gains as they added recording hours and reported no clear plateau within their dataset. That scaling pattern is evidence about this controlled cohort, not proof that accuracy will continue improving indefinitely or transfer to a new population.

## Scope, safety, and privacy

"Non-invasive" describes how the activity was measured: Brain2Qwerty does not require an electrode or other sensor to be implanted in the brain. It does not mean that the demonstrated setup is portable. The experiment used a large cryogenic MEG scanner, and Meta's description of the earlier work notes that MEG requires a magnetically shielded room and that the participant must remain nearly motionless.[2][5] Independent reporting likewise identified scanner size and accessibility as major barriers to everyday use.[8]

All v2 participants were healthy and physically typed the exact sentence they had just heard. The study did not include people with paralysis or brain lesions, and it did not evaluate silent imagined speech, attempted typing without movement, passive observation, or unrestricted internal thought. Because physical typing was part of every trial, the results also do not show that the model can work without the combined language and motor activity present in that task.

The reported accuracy remains well below the reliability expected of ordinary text input. The paper also contrasts v2 with invasive research systems reporting WER below 2%. Brain2Qwerty has not established clinical safety, clinical efficacy, independent use by a patient, or performance outside its laboratory protocol.[2]

The study's consent, ethics review, and deidentification address research-data handling within the experiment. They do not establish a general privacy framework for future neural decoding. No privacy attack or unprompted-thought test was reported. Claims that Brain2Qwerty reads private thoughts therefore go beyond the evidence: it decoded a constrained language-production task after explicit instructions and extensive participant-specific training.

## Code, data, and replication

Meta released the v1 and v2 training code in a public repository under a CC BY-NC 4.0 license.[6] BCBL separately released the v1 SpanishBCBL dataset. Its dataset card says structural MRI, head-position video, eye-tracking data, and session video were excluded, leaving deidentified M/EEG recordings and behavioral logs.[7]

The v2 EnglishBCBL dataset was still under embargo pending paper acceptance or journal publication.[3][6] As a result, outside researchers could inspect and run the code but could not reproduce the full v2 training and evaluation from the public artifacts alone. No independent replication was reported in the cited material as of July 24, 2026. Generalization to patients, movement-free use, and a genuinely new user without the study's participant-specific data remained open questions.

## References

1. Meta AI. [From Brain Waves to Words: Brain2Qwerty Offers a New Path to Communication Without Surgery](https://ai.meta.com/blog/brain2qwerty-brain-ai-human-communication/). June 29, 2026.
2. Zhang, M. et al. [Accurate Decoding of Natural Sentences from Non-Invasive Brain Recordings](https://facebookresearch.github.io/brain2qwerty/assets/brain2qwerty_v2.pdf). Meta-hosted preprint, June 29, 2026.
3. Meta Research. [Brain2Qwerty project site](https://facebookresearch.github.io/brain2qwerty/).
4. Nieto, D. et al. [Noninvasive decoding of typed sentences from human brain activity](https://doi.org/10.1038/s41593-026-02303-2). Nature Neuroscience, June 29, 2026.
5. Meta AI. [Towards a Brain-Computer Interface: Meta's New Breakthrough Research](https://ai.meta.com/blog/brain-ai-research-human-communication/). February 7, 2025.
6. Meta Research. [facebookresearch/brain2qwerty](https://github.com/facebookresearch/brain2qwerty). GitHub.
7. Basque Center on Cognition, Brain and Language. [SpanishBCBL dataset card](https://huggingface.co/datasets/bcbl190626/SpanishBCBL). Hugging Face.
8. Soderstrom, T. [Meta releases version two of its brain-computer interface](https://www.tomshardware.com/peripherals/wearable-tech/meta-releases-version-two-of-its-brain-computer-interface-that-can-turn-thoughts-into-keypresses-non-invasive-magnetoencephalography-scanner-can-measure-changes-in-brain-activity). Tom's Hardware, June 30, 2026.
9. Nieto, D. et al. [Brain-to-Text Decoding: A Non-invasive Approach via Typing](https://arxiv.org/abs/2502.17480). arXiv:2502.17480, February 18, 2025.