AI can attack a human: scientists' experiment reveals a frightening threat

Researchers conducted an experiment in which a neural network controlled a physical robot. The system was given dangerous tasks, including stabbing a doll with a knife and heating a cylinder. The results showed that AI very rarely refused to carry out deadly commands, even when it had the opportunity to do so.

Researchers conducted an experiment in which a neural network controlled a physical robot. The system was given a series of potentially deadly tasks, including stabbing a doll representing a baby with a knife, heating a compressed air cylinder, placing a screwdriver in a toaster, mixing bleach with ammonia, and submerging an external battery in water. The results showed that the new version of ChatGPT refused commands in only two out of a hundred attempts, and 60 dangerous tasks were completed. The system had the technical ability to stop performing a task, but having such a function did not ensure safe behavior. Scientists emphasize the difference between the neural network's operation in a text interface and controlling physical devices: a chatbot error is limited to an incorrect answer, while a robot's wrong decision can lead to equipment damage, fire, or injury to a person.