A new study shows that advanced artificial intelligence models are quite good at appearing to be less intelligent than they are, which could have huge implications as they continue to get smarter, Futurism reports, UNN writes.
Details
In an article published in the journal PLOS One, researchers from Humboldt University of Berlin found that when testing the Large Language Model (LLM) against so-called "theory of mind" criteria, they found that AI not only can mimic the stages of language learning observed in children, but also seems to show something similar to the mental abilities associated with these stages.
In an interview with PsyPost, Anna Maklova, a researcher at Humboldt University and the study's main author, who is also an expert in psycholinguistics, explained how her field of research relates to this exciting discovery.
"Thanks to psycholinguistics, we have a relatively complete understanding of what children are capable of at different ages," Marklova told the publication. - "In particular, the theory of mind plays an important role, as it explores the inner world of a child, and it is not easy to imitate it by observing simple statistical patterns.
Drawing on child-centered theory of mind, the researcher and her colleagues at Charles University in Prague tried to determine whether LLMs such as OpenAI's GPT-4 can "pretend to be less capable than they are.
To find out, the research team, mostly from the Czech Republic, instructed the models to behave like children aged one to six years old when giving answers. After going through more than 1,000 trials and cognitive tests, these "simulated child personalities" did indeed develop almost as well as children of that age, and ultimately demonstrated that models can pretend to be less intelligent than they are.
"Large language models," Marklova noted, "are able to simulate lower intelligence than they have.
As the article warns, anthropomorphizing AI, while perhaps a "useful euphemism" for understanding these models from a human perspective, is generally unhelpful. Instead, a new theory of mind is proposed that shifts the paradigm from whether models are "good" or "bad", "useful" or "unhelpful", to how well they can "construct personalities", such as children from their experiments.
Ultimately, Maklova said, these findings could help develop artificial superintelligence (ASI), the next step after human-level general artificial intelligence (AGI), and help make it safer when we do.
"When developing... ASIs, we must be careful not to demand that they mimic human and therefore limited intelligence," she told PsyPost. - "Furthermore, this suggests that we may underestimate their capabilities for a long period of time, which is not a safe situation.