Can AI Read Your Thoughts? Non-Invasive Brain Decoding in 2026

Every few months a headline says AI can now read minds. The research behind those headlines is real and often impressive, but it is also narrower than the headlines suggest. In 2026, decoders that work without surgery can reconstruct typed sentences from a magnetic brain scanner, recover the gist of a story someone is listening to, or describe a video someone remembers. None can read an unwilling person's private thoughts, and the best results still need a machine the size of a room. This article goes through the main systems, what each actually measures, and how to judge the next headline.

Three ways to see the brain from outside

Method	What it measures	Timing	Portable?	Cost
EEG	Voltages on the scalp	Milliseconds	Yes (caps, headbands)	Hundreds to thousands of dollars
MEG	Magnetic fields from neural currents	Milliseconds	No, needs a shielded room	Millions
fMRI	Blood-oxygen changes	Seconds	No, a hospital scanner	Millions

EEG is the only one you can wear, and it is also the blurriest, because the skull smears electrical signals. MEG sees fast activity much more clearly. fMRI sees where activity is with fine spatial detail, but blood flow responds several seconds after the neurons, so it captures meaning over a stretch of time rather than individual words.

Meta Brain2Qwerty: typed sentences from MEG

Meta's FAIR lab recorded volunteers in MEG and EEG while they typed sentences, then trained a decoder to reconstruct what was typed from the brain signals alone. The peer-reviewed result (Nature Neuroscience, June 2026): with MEG, a 29% character error rate on average and 18% for the best participant; with EEG, 65%. A second version released in June 2026 adds a large language model and decodes whole English sentences with a 39% word error rate, about 61% of words right.

The fine print: participants were healthy and actually typing, the signal came mostly from motor areas, and decoding is not real-time. Full details in Meta Brain2Qwerty explained.

UT Austin: the gist of a story from fMRI

In 2023, Jerry Tang and Alexander Huth at the University of Texas at Austin published a decoder in Nature Neuroscience that reconstructs continuous language from fMRI. Participants spent about 16 hours each in the scanner listening to podcasts. The decoder does not read words directly. A language model proposes candidate phrases, a second model predicts what brain activity each phrase would cause, and the phrases whose predictions best match the real scan are kept.

The output is paraphrase, not transcript. When a participant heard "I don't have my driver's license yet", the decoder produced "She has not even started to learn to drive yet". UT describes it as getting the meaning right about half the time. It worked on stories the person heard, stories they imagined telling, and even silent films.

The privacy findings matter as much as the decoding:

• A decoder trained on one person was barely above chance on another.
• Participants could defeat it by deliberately thinking of something else, such as counting by sevens or naming animals.
• Cooperation was needed both to train it and to use it.

A 2025 follow-up in Current Biology cut the training time for a new person from about 16 hours to about 1, by transferring a decoder from other people through alignment on shared stimuli, including silent Pixar shorts. The team is now working with aphasia researchers.

NTT mind captioning: describing what you see or remember

In November 2025, Tomoyasu Horikawa at NTT published "mind captioning" in Science Advances. Six participants watched short videos in an fMRI scanner for about 17 hours each. A decoder maps brain activity to the semantic features of a caption. A masked language model then starts from a blank and edits candidate text word by word until its features match.

It picked the right video out of 100 about 50% of the time while people watched, and about 30% while they only recalled it (chance is 1%). It also worked without relying on the brain's language network, so it seems to decode visual meaning rather than inner words. The author notes that consent and privacy questions will need answers before anything like this leaves the lab.

EEG to text: the field's cautionary tale

EEG is where consumer "thought typing" would have to happen, and it is where the most important critique sits. Early EEG-to-text papers, such as DeWave from the University of Technology Sydney (2023), reported encouraging scores on reading datasets. A 2024 paper, "Are EEG-to-Text Models Working?", found two problems in how such systems had been evaluated:

• Many fed the model the correct previous words during testing (teacher forcing), which inflates scores.
• When the EEG was replaced with random noise, some models scored about as well as with real brain data. The language model was writing plausible sentences on its own.

The takeaway applies to every brain decoder that ends in a language model, including the good ones: ask what happens when the brain signal is removed. Meta's version 2 paper answers exactly that question: without the brain tokens, its word error rises from 39% to 49%.

Meta's Neural Band is not mind reading

Meta's Ray-Ban Display glasses ship with the Meta Neural Band, a wristband that since January 2026 lets US users in early access write messages by tracing letters with a finger on any surface. It is often described as reading thoughts. It does not. It uses surface EMG, the electrical activity of the muscles in the forearm. The research behind it, published in Nature in 2025, trained on thousands of people so it works without per-person calibration, and reached a median of 20.9 words per minute for handwriting.

It reads the command your brain sends to your muscles as you move, even very slightly. That makes it a clever input device, not a brain decoder.

The startup claims

Several companies are promising wearable thought-to-text. Sabi, for example, has described an EEG beanie with a "brain foundation model" and a "thought typing" speed of around 30 words per minute. As of September 2026 there is no peer-reviewed paper or independent evaluation behind that figure. Given the noise-baseline problem above, treat such numbers as claims until someone publishes the method and the controls. Other ventures, such as Merge Labs, which launched in January 2026 with backing from OpenAI, are exploring ultrasound and molecular approaches, with no decoding results yet.

How to read the next "mind reading" headline

Five questions sort most claims:

1. What is the recording? EEG, MEG, fMRI, EMG, or an implant. MEG and fMRI results do not transfer to a headband.
2. What did the person do? Typing, listening, watching, attempting to speak, or imagining. "Thought" is usually one of the first four.
3. Open text or picking from a list? Choosing the right sentence among 100 candidates is much easier than writing it from scratch.
4. What does the language model add? Is there a noise or shuffled-signal baseline?
5. How much training per person? Hours in a scanner, with cooperation, is the norm.

For what implanted systems can do, which is much more, see brain-to-text in 2026.

Frequently asked questions

Can AI read your mind?

Not in the everyday sense. Non-invasive decoders can reconstruct typed sentences from MEG, the gist of a story from fMRI, or a description of a video someone recalls, but only in a lab, after hours of training on that person, and with their cooperation. A 2023 study showed people could defeat a decoder simply by thinking about something else.

Can EEG headsets turn thoughts into text?

Not reliably. In Meta's direct comparison, EEG had a 65% character error rate on typed sentences, against 29% for MEG. A 2024 study found some published EEG-to-text models performed about as well on random noise as on brain data. No consumer EEG device has published peer-reviewed evidence of open thought-to-text.

Does Meta's Neural Band read your thoughts?

No. It measures electrical activity in your forearm muscles (EMG) as you move your fingers, even slightly. Its handwriting feature recognises letters you trace, at about 21 words per minute in Meta's research.

Is mind reading a privacy risk?

Researchers take it seriously. Current non-invasive decoders need the person's cooperation to train and to use, and fail on anyone they were not trained on. Both UT Austin and NTT have called for rules on consent and mental privacy before the technology improves.

What is the most accurate non-invasive brain-to-text system?

As of September 2026, Meta's Brain2Qwerty version 2, which decodes English sentences from MEG with a 39% average word error rate and 22% for the best participant. It needs a MEG scanner in a shielded room and does not run in real time.

Sources

• Lévy et al., "Noninvasive decoding of typed sentences from human brain activity", Nature Neuroscience (2026)
• Brain2Qwerty v2 paper and code (Meta FAIR, 2026)
• Tang et al., "Semantic reconstruction of continuous language from non-invasive brain recordings", Nature Neuroscience (2023)
• Tang & Huth, "Semantic language decoding across participants and stimulus modalities", Current Biology (2025)
• UT Austin news: decoder transfer with less training (2025)
• Horikawa, "Mind captioning: Evolving descriptive text of mental content from human brain activity", Science Advances (2025)
• NTT press release: mind captioning
• Duan et al., "DeWave: Discrete EEG Waves Encoding for Brain Dynamics to Text Translation", arXiv:2309.14030
• Jo et al., "Are EEG-to-Text Models Working?", arXiv:2405.06459
• Défossez et al., "Decoding speech perception from non-invasive brain recordings", Nature Machine Intelligence (2023)
• CTRL-labs at Reality Labs, "A generic non-invasive neuromotor interface for human-computer interaction", Nature (2025)
• Meta: neural handwriting early access (CES 2026)