Partly, under conditions that are worth stating precisely. In May 2023 Jerry Tang and Alexander Huth published a semantic decoder in Nature Neuroscience that reconstructed continuous language from non-invasive brain recordings. A person lay in an fMRI scanner; a transformer model turned the blood-flow signal into text that captured the gist of what they were hearing or imagining, matching the intended meaning roughly half the time. It is a real result and it was replicated in the paper across participants. The conditions are as important as the finding. The decoder required about fifteen hours of scanner time per person to train, and it worked only for the individual it was trained on — run against an untrained person, it produced unintelligible output. It worked only with willing participants. And when a trained subject deliberately resisted, by counting, naming animals or telling themselves a different story, the decoder failed entirely. The researchers tested that on purpose and reported it. So the honest answer is that meaning can be partially reconstructed from a cooperative, individually-trained person lying still inside a superconducting magnet the size of a small room. That is a genuine advance in decoding, and it is a long way from reading a mind that does not wish to be read.
- Tang & Huth et al., 'Semantic reconstruction of continuous language from non-invasive brain recordings', Nature Neuroscience, 1 May 2023 (DOI 10.1038/s41593-023-01304-9)
- ~15 hours of fMRI training per individual; decoder is subject-specific
- Applied to an untrained individual it produced unintelligible output
- Trained subjects who deliberately resisted defeated it completely — the authors tested resistance and published the failure
- Reconstruction captures gist, matching intended meaning about half the time, not word for word
Open questions
- The result is bounded by fMRI. It says nothing about what any other modality can or cannot do.
- That a cooperative, trained subject can be partially decoded does not establish that an unwilling, untrained person can be decoded by anything.
- The resistance finding is a property of this decoder tested in this way. It is not a general guarantee about future systems.
References
- Semantic reconstruction of continuous language — Nature Neuroscience (2023)
- What would it actually take to do this without consent?
- Why did legislatures write laws for neural data?
The paper is cited for its own published result, including the limits its authors reported.
Not independently verified
This concept reports one peer-reviewed study and the constraints its authors documented. It establishes no capability beyond what that study demonstrated.
Topics: Technology · Surveillance & memory access