Generative pre-trained transducers (GPT), such as those used in OpenAI’s ChatGPT chatbot and Dall-E image generator, are the foundation of modern technology in artificial intelligence research. Everyone wants to apply GPT models to almost everything, but these implementations bring controversy. To these discussions, the new GPT-1 model, which can read the human mind, was added.Artificial intelligence that reads thoughts
The developed GPT model is not different from ChatGPT in general. The main difference is that the command is human brain activity. The team, led by Jerry Tang of the University of Texas, published their work in Nature Neuroscience. The method uses images from an fMRI machine to interpret what the subject “heard, said, or imagined.” Therefore, this method is the only successful method that does not require electrodes connected to the subject’s brain. I say successful because the accuracy of the predictions can reach up to 82 percent. The model, called GPT-1, is the only method that interprets brain activity in a continuous language format. Other techniques can extract a word or a short sentence, but GPT-1 can explain what the subject is thinking.
For example, one participant listened to a recording of someone saying, “I don’t have a driver’s license yet.” The language model interpreted the fMRI images as “He hasn’t even started learning to drive.” There is a key difference between this new technique and existing techniques that use invasive electrodes placed in the brain. While electrode-based platforms typically predict text from motor activity in a person’s brain, Tang’s team focuses on brain blood flow captured in fMRI machines. Therefore, in this model, thoughts are not reflected verbatim, but as a summary of the idea.reflects the essence of thought
“Our system operates on a very different level. Instead of looking at this low-level motor thing, our system really works at the level of ideas, semantics, and semantics. That’s what we’re trying to achieve.” In experiments, GPT-1 highly accurately predicted the thoughts of people fed with the data they imagined, as well as auditory and visual data.- Perceived speech (subjects listened to a recording): 72-82 percent decoding accuracy
- Imaginary speech (subjects mentally told a one-minute story): 41-74 percent accuracy
- Silent movies (subjects watched silent Pixar movie clips): 21-45 percent accuracy in deciphering subjects’ comments about the movie
In one instance, the subject imagined, “I drove down a dirt road to a wheat field, over a stream, and past some log buildings.” The model commented that “To get to the other side, he had to cross a bridge and a very large building in the distance.” Thus, he missed some of the key details and key context that was arguable, but still managed to grasp the general elements of one’s thinking.
Warning about the dangers of technology
Machines that can read minds may be the most controversial form of GPT technology ever. While the team envisions the technology helping people with ALS or aphasia speak, it also acknowledges the potential for abuse. The study emphasizes the critical importance of raising awareness of the risks of brain analysis technology and enforcing policies that protect the mental privacy of each individual.
45821
