It was revealed that artificial intelligence can “pretend to be stupid”

in berlin from Humboldt University researchers, large language models (LLM) In their studies on “theory of mind”, they revealed that artificial intelligence can act as if they were much more naive and stupid than they are. In this study, it was revealed that when artificial intelligence was asked to act like a child, artificial intelligence could do this.

Research Assistant and Psycholinguistics Expert from Humboldt University, who is the lead author of the study Anna Maklova, “Thanks to psycholinguistics, we have a relatively more comprehensive understanding of what children can do at various ages. Theory of mind, in particular, plays an important role because it explores the child’s inner world and cannot be easily imitated by simple statistical patterns.” he said.

Artificial intelligence can act stupider than it is

Researchers, on GPT-4 “less intelligent than he is”” To see if it could behave, he asked the artificial intelligence to respond like children aged between one and six. In more than 1000 experiments, child personas created by GPT-4 were also cognitively like real kids It turned out that it “developed”. This means that artificial intelligence can make itself seem less intelligent than it is.

Maklova and her team, in their article artificial to intelligence While he said that giving human characteristics is useful for understanding models, but not useful in general, the criteria for creating models are “good-bad”, “helpful-unable”” instead of focusing on the formed personality. Maklova stated that in this way, AGI (General artificial intelligence) and ASI (super artificial intelligence) He stated that his work would be safer.

The research was published in PLOS ONE.