Imagine a person sitting still in a scanner, looking at a photograph of a beach, with no words spoken and nothing typed. Minutes later, a computer produces a rough picture of the beach from the pattern of activity in the viewer's visual cortex. Or imagine someone thinking a sentence in silence, while a decoder produces a caption for what they were picturing. Both experiments have been run in research labs in the last year. Neither is a mind reader in the sense the phrase suggests, and both raise a question that matters more than the accuracy figures: who should be asked before a machine reads what is in someone's head?
What a decoder learns
A brain decoder is a machine learning model trained on pairs of brain recordings and known stimuli or text. It learns which patterns of activity go with which content, then guesses content from new recordings. Non-invasive methods such as functional MRI, which tracks blood flow, and magnetoencephalography, which records magnetic fields from neurons, are slower or less precise than electrodes placed inside the skull. What they gain is safety and no surgery.
The newer step is the use of large AI language and image models. They supply a learned sense of what plausible sentences and pictures look like, so the decoder can produce fluent output from weak signals.
Meta's typing decoder
On 29 June 2026, Meta published Brain2Qwerty v2, a system that decodes typed sentences from magnetoencephalography recordings. In Meta's own figures, it reaches 61 per cent word accuracy across nine volunteers, trained on about 22,000 sentences. The best participant reached 78 per cent, and for that person more than half of sentences were decoded with one word error or less. Meta says non-invasive methods it compared against reached 8 per cent word accuracy. Meta says the comparison methods reached 8 per cent word accuracy; we have not checked the comparator. These are the company's figures, and the company's comparison is with the methods it chose.
Meta describes the system as capable of "real-time sentence decoding from non-invasive brain recordings." The post does not say whether decoding works live, as a person types, or after a full typing session. The technology press objected. The Next Web reported that the models "need a full typing session to finish before they produce anything, so there is no live feedback." Our reading: "real-time" is Meta's word, not a settled result.
Meta has released its training code under a non-commercial licence. The v2 dataset is under embargo until the paper is accepted, so outside researchers cannot yet check the claims on the same data.
Pictures from the scanner
A separate preprint, not peer reviewed, posted to arXiv on 23 July 2026 as "Real-time Reconstruction of Human Visual Perception from fMRI," describes decoding viewed images within seconds of their being shown. The abstract says the system decoded single-trial perception "within seconds after an image was shown to the participant." The work was done with one participant, and a Princeton University blog post on the study describes the participant as unnamed and the training data as about an hour. The preprint's own timing figures, and its comparisons, were not checked for this essay.
The authors describe the approach as the first real-time pipeline of its kind. That is their claim, and it has not been independently checked.
Mind captioning
A different line of work, reported in Science Advances in 2025 by researchers at NTT in Japan, produced text descriptions of what people were watching from fMRI scans of six participants, about 17 hours of scanning each. The method used a deep language model to generate candidate captions, which were then refined to match the decoded brain activity. The reports say the captions could be generated without relying on the brain's classic language areas. Secondary coverage says the correct video was identified about half the time, against 1 per cent by chance. The primary paper was not read in full for this essay.
The mind-captioning work is sometimes attributed to Princeton in newsletters. That attribution is wrong: the study is from NTT, and it was published in 2025.
What is not demonstrated
Two claims are not supported by any source this essay relied on. The first is remote thought control, the idea that someone can send commands from a distance by reading thoughts. No source describes that. The second is open-ended free thinking decoded live. The sources show reading from scanners, in cooperative participants, under lab conditions.
Our reading: the honest summary is narrower than the headlines and still significant. Decoding now works for constrained tasks, with cooperative participants and expensive hardware, and the trend is toward less equipment and more capable models.
Implications
Three things matter most here.
The first is mental privacy. Our reading: a scanner is not needed for every use of this technology in the future, but the research to make it smaller is what is being published now. Our reading, labelled as speculation: once the equipment is cheap, the question of consent will be asked by people who have less power to refuse.
The second is the gap between demonstration and proof. A 61 per cent word accuracy in nine volunteers is an important result and a long way from a usable product. Readers should ask what the test conditions were before they accept the number.
The third is governance. Meta has released its code, and the research community is moving faster than the rules for neural data. The law treats brain data differently in different places, and the sources read do not describe a shared standard.
Conclusion: the layer before the words
Decoding the brain's pre-language layer is becoming possible, and the first demonstrations are real. What they show so far is that a machine can match patterns of activity to sentences and pictures under strict conditions. Whether that becomes a tool for communication, or a tool for reading people who have not agreed to be read, depends on decisions that are still open. Those decisions should be made before the decoders are good enough to make them urgent.
Sources: Meta, "Brain2Qwerty v2" (29 June 2026), with the publication page and GitHub repository; The Next Web on the Meta system (1 July 2026); arXiv preprint "Real-time Reconstruction of Human Visual Perception from fMRI" (23 July 2026); Princeton University blog on the fMRI study (31 July 2026); Science Advances, NTT mind-captioning study (2025), read through secondary coverage; MIT Technology Review on implanted brain-computer interfaces (19 June 2026).

