Mind & behavior
A brain scanner reconstructed the meaning of stories people heard
The reconstructed sentences captured aspects of meaning, but the experiment depended on individual training and cooperation. It was not a transcript of unspoken thoughts.
A story reconstructed from a slower signal
A person listened to a story inside an MRI scanner. A computer then produced a sequence of words that sometimes recovered aspects of the story's meaning. Jerry Tang and colleagues reported this non-invasive semantic decoder in 2023. The striking feature was continuous language reconstruction, rather than simply choosing one item from a short menu of possible words.
The input was functional MRI, an indirect measure related to blood oxygenation, not a microphone inside the brain. The decoder used a language model together with a model connecting language to an individual's brain responses. Its fluent output therefore combined constraints from measured signals with expectations about plausible language.
Meaning and wording are different targets
Recovering an approximate idea is not the same achievement as recovering the exact sentence. A reconstruction may preserve who is doing something while changing the action, or preserve a general situation while inventing a detail. Fluency can conceal those errors because a well-formed sentence feels complete even when its evidential basis is incomplete.
A useful comparison is between an answer and the search process that produced it. The language model proposes possibilities; brain measurements help distinguish among them. Readers should ask how performance changes when the brain evidence is removed or mismatched, not judge the result solely by whether a displayed example sounds like English.
Cooperation was part of the result
The researchers explicitly tested mental privacy. They found cooperation necessary both when training and when applying this decoder. That finding sets a boundary around the demonstrated capability. The published examples included perceived language, imagined language and silent videos, but they did not demonstrate unrestricted access to any person's private thoughts.
The authors' code repository makes the computational workflow inspectable. It includes separate preparation, model-training and decoding steps. This is valuable because a demonstration becomes more assessable when others can examine its machinery. The repository is not a ready-made claim that the system works equally well on untrained people, arbitrary scanners or every kind of mental activity.
A promising interface and a measurement problem
One possible motivation is communication for people who cannot easily produce speech. Moving from a research demonstration to a dependable interface would require evaluating accuracy, practical burden, user control and the consequences of errors in the intended population. Those are questions for additional studies, not outcomes established by an impressive reconstruction.
The privacy discussion should be equally precise. This experiment supplies evidence of both capability and resistance under its tested conditions. It neither justifies dismissing future concerns nor supports a claim that remote mind reading already works. Its most useful lesson is that a generated sentence is an interpretation of a measured signal. Treating that sentence as a person's own words would discard the very uncertainty the study needs us to understand.
Sources and further reading
- Semantic reconstruction of continuous language from non-invasive brain recordings ↗
Author manuscript, PMID/PMCID record for Nature Neuroscience 26, 858–866
- Tang et al., Nature Neuroscience version of record (2023) ↗
Abstract; cooperation tests; data and code availability
- Huth Lab semantic-decoding source code ↗
README and preparation/training/decoding workflow