An AI has succeeded in converting brain activity into text: this is the first time that a tool can read a person’s mind in a non-invasive way!
Using artificial intelligence coupled with a brain imaging system, neuroscientist Alexander Huth and researcher Jerry Tang of the University of Austin in Texas have been able to translate a person’s thoughts and turn them into text.
To do this, volunteers had to lie down to undergo the longest MRI of their lives: no less than 16 hours of scanning while listening to a podcast or silently making up stories to train the AI ​​in their learning. Depending on the words perceived by the brain, the blood flowed differently. The computer’s role was then to decode this data and learn to match key brain activities to words, eventually transcribing these stories into a continuous flow of text. Note that this model was developed “custom” because if the AI ​​were trained on one participant, it would poorly transcribe the thoughts of a second patient that it has not yet learned to “decode”.
To interpret a match between brain activity and meaning, the tool used a large AI language model based on OpenAI’s GPT-1, an ancestor of ChatGPT. Here the computer is not attempting to read the activity “word by word” but is attempting to match a type of neural activity to a sequence of words with a specific meaning. So what is transcribed is the meaning of the thought and not the exact words.
The machine has made some mistakes, particularly in relation to the pronouns or gender to be transcribed, and it happens that the first and third person are confused or the initial thought is transcribed slightly differently. For example, if one of the participants thought “leave me alone,” the machine would transcribe “I told you to leave me alone.” In another case, a participant heard “I don’t have a driver’s license yet” and the machine transcribed “She hasn’t even started learning to drive yet.” However, the meaning remains close.
The researchers were surprised by the accuracy and precision of this semantic decoder. This is a major advance over previously developed technologies, which can only sense single words or short phrases and typically require surgical sensors implanted in the skull.
They admit to being both shocked and excited by the results. How well will a machine be able to read a human’s mind? Nobody knows, but the result of this research gives hope to patients who have lost speech.