Creeping Towards Mordor
Salvo Magazine | August 22, 2023 | Daniel Witt
If you were worried about some benevolent world government lacking full access to our inmost thoughts, I have good news! In a new study, experimenters claim to have made a leap in mind-reading technology.
The study “Semantic reconstruction of continuous language from non-invasive brain recordings” demonstrates some ability to translate a human subject’s thoughts into words and sentences using fMRI (functional MRI) scans decoded by the GPT-1 language model and other software.
The technique was clever. Experiment subjects listened to hours of podcasts while undergoing brain scans. The AI model had access to both the brain scan data and the podcast data, and it was directed to find the correlations between the words said in the podcast and the activities of the recipient brain.
After hours of training, the AI model was then cut off from the podcast data and forced to infer sentences based on the brain data alone while the subjects listened to new podcast stories.
It worked. The AI was able to construct fairly accurate sentences based on the brain scans. For example, the podcast input:
i didn’t know whether to scream cry or run away instead i said leave me alone i don’t need your help
resulted in brain activity that was decoded as:
started to scream and cry and then she just said i told you to leave me alone you can’t hurt me anymore i’m sorry.
Next, they tested the ability of the AI to interpret imagined speech, absent any external stimulus. Again, it was successful, albeit with less accuracy than when it was analyzing perceived speech from podcasts.
Finally, they tested the ability of the AI to decode non-language-based thoughts by having the subject watch silent videos. Sure enough, the program was able to describe some of the events of the videos based only on the brain activity of the viewer.

















