Overview
- A Science Advances study reported that non-invasive MRI data from six volunteers was decoded into descriptions of what they watched or later recalled from short videos.
- The approach generated sentence-level captions that closely matched scenes, including an example of a person jumping near a waterfall.
- Researchers used a two-step process that aligns semantic features and refines outputs with language models to produce coherent text.
- The system requires extensive individualized training, active cooperation from the participant, and a large MRI scanner, and it has not been shown to decode unsolicited internal thoughts.
- Coverage highlights potential communication benefits for people with speech impairments alongside significant concerns about mental privacy, consent, and possible misuse.