Artificial intelligence will learn to read minds: scientists got scared

The moment when artificial intelligence (AI) will learn to read our thoughts is very close. Attempts to involve him in deciphering thoughts are already in full swing. And a group of computational neuroscientists have revealed in a new publication in the journal Science that they have developed an AI-based technique that can translate brain images into words and sentences. While the technique is in its early stages and far from perfect, it could eventually help people with brain injuries or paralysis regain their ability to communicate, the researchers say. But at the same time it can be useful to attackers, they do not hide.

According to the authors, a new study has shown that, using the right methods and better models, AI is indeed able to decode (decode) what a person is thinking about. This was stated by a specialist in computational neuroscience from the Massachusetts Institute of Technology Martin Shrimpf.

Attempts to translate thoughts into words using the brain-computer interface (neural interface) have been made for a long time. Research in this direction has been going on for several years. For example, other research groups have already created brain-computer interfaces that, for example, translate the brain activity of a paralyzed patient into words. However, most of these approaches are based on the use of electrodes implanted in the patient’s brain. Non-invasive methods based on methods such as the electroencephalogram, which measures brain activity using electrodes attached to the scalp, have been less successful. Methods based on the electroencephalogram (EEG) are so far only able to decipher phrases and cannot restore coherent speech. In addition, previous interfaces also usually focused on people trying to speak or thinking about speech, so they relied on brain regions involved in the production of speech-related movements and only worked when the person moved or tried to move.

Now neuroscientists at the University of Texas at Austin have developed a functional magnetic resonance imaging technique that allows you to directly access the language areas of the brain to decipher imaginary speech. Widely used in neurological research, this non-invasive method monitors changes in blood flow in the brain to measure neural activity. The goal of the research was to associate each word, phrase, or sentence with the particular pattern of brain activity it elicited. To gather the necessary data, the researchers scanned the brains of three participants, each of whom listened to about 16 hours of storytelling podcasts. Based on this data, the researchers compiled a set of cards for each subject that indicated how their brain reacts when they hear a particular word, phrase, or meaning. Because an MRI takes a few seconds to record brain activity, it doesn’t capture every single word, but rather the general idea of ​​each phrase and sentence. Scientists used MRI data to train AI to predict how a particular person’s brain would respond to language.

Initially, the system had difficulty converting brain scans into language. But then the researchers plugged the GPT natural language model into the system to predict which word might follow another. Using maps from the scans and a language model, they analyzed various possible phrases and sentences to see if the predicted brain activity matched the actual one. If they matched, they saved that phrase and moved on to the next one. After that, the subjects listened to podcasts that were not used in the training. Gradually, the system created words, phrases, and sentences, eventually producing ideas that exactly matched what the person heard. The technology was especially good at capturing the essence of the story, even if it didn’t always get every word right. It also worked when the subject was telling a story or watching a video. In one experiment, participants watched a movie without sound while the system tried to decipher their thoughts. When the subject was watching an animated film in which a dragon knocks someone down, the system said, “He’s knocking me to the ground.” All this happened without the participants being asked to speak. “This really shows that we are dealing with something deeper than just language,” the authors of the study said.

Scientists hope that in the future this system could help people who have lost the ability to communicate due to a brain injury, stroke, or “blocking syndrome” – a type of paralysis in which a person is conscious but paralyzed. However, for this it is necessary not only to improve the technology, using more data for training, but also to make it more accessible. Since the system is based on MRI, it is expensive and difficult to use, so simpler and more portable imaging modalities, such as EEG, are being sought.

While this technology is not yet capable of deciphering spontaneous thoughts in the real world, there are serious concerns that if improved, the technology could mimic mind reading. “When it worked for us, we thought:“ My God, this is just awful, ”the authors of the study themselves admit.

To address these concerns, the authors tested whether a decoder trained on one person would work on another, and it turned out not to. Consent and cooperation were also critical, because if a person resisted doing a task, such as counting instead of paying attention to a podcast, the system could not decipher the meaning of their brain activity. “Time will tell how applicable all this will be in practice, but experts are already afraid that such technologies can be used both for good, for example, to communicate with people who have lost the ability to speak, and for evil, for example, to extract information. of people who do not want to share it,” says military doctor, therapist Alexei Vodovozov.