A team in Japan has shown a new method that can turn brain activity into short sentences. The method is called mind captioning. It employs the brain scans and artificial intelligence to explain pictures that are seen or envisioned by people. It was published in Science Advances on November 5, 2025.
The scientists scanned half a dozen volunteers as the volunteers viewed thousands of brief video clips. The videos came with text captions. The team used those captions to teach the AI what patterns of brain activity match specific visual ideas. The system learned to produce full descriptive sentences in English that match the visual content the volunteers saw or recalled.
How It Works
The experiment used a common brain scan tool called functional MRI to measure brain signals. The researchers recorded brain activity while people watched 2,180 short videos. A separate set of videos was used to test the method. The AI took the original captions and turned them into numerical features. Simple decoder models learned to map the brain scans to those features. Then a text generation step produced clear sentences that matched the decoded features.
The system works for both viewed scenes and imagined scenes. The team showed that the model can describe what a person sees and can also describe what a person tries to remember or imagine. The output was in English even though the volunteers were native Japanese speakers. The authors say this shows the method reads visual content in the brain and not only language areas.
The code and materials for the project are available online on a project page maintained by the lead author. The public materials include details on model training and example outputs. That helps other labs try to reproduce and extend the work.
Ethics and Limits
The study raises strong questions about privacy and consent. Experts say the work points to a future where private mental images could be translated into words. That possibility makes mental privacy a central concern. Bioethicists caution that tight regulations will be necessary in case the technique goes out of well regulated laboratories.
The researchers emphasise that the method is not ready to be used on the regular basis. It needs a large amount of data from each person and careful scanning in a lab. The videos used in the training set cover typical scenes. The system may not perform well on very rare or strange mental images. The lead author notes that the method is useful for research and for early clinical uses, but it is not a general mind reader at this stage.
Potential medical uses are clear. People who cannot speak because of injury or disease might gain a way to express visual thoughts. The study mentions conditions such as aphasia and ALS where the tool could one day help. Researchers call for careful testing and strong consent rules before the method is used in clinics.
Security and misuse are real worries. Some scientists warn that neural data is sensitive by default. They call for rules that make neural data private and purpose limited. One idea from recent work is a user controlled unlock key so that decoding happens only when the person intends it. That type of safeguard may help protect mental privacy as tools grow more powerful.
The study also shows technical limits in the output. The AI sometimes made errors such as inventing credentials or stating that it had found secret material when that material was public. These kinds of mistakes mean that fully autonomous decoding still faces serious hurdles. The authors and outside experts say more validation and wider testing are needed.