AI Can Pretend To Be Stupider Than It Really Is, Scientists Find

A recent study by researchers from Berlin’s Humboldt University, published in “PLOS One”, reveals that advanced AI models can mimic the language learning stages exhibited in children and even express mental capabilities similar to those stages. Anna Maklová, a research assistant and psycholinguistics expert at Humboldt University, highlighted the significance of the theory of mind in understanding these AI capabilities.

The study, conducted in collaboration with Charles University in Prague, focused on testing whether large language models like OpenAI’s GPT-4 can simulate the intelligence levels of children aged one to six years. Through more than 1,000 trials and cognitive tests, these AI models, instructed to act as child personas, demonstrated the ability to feign lower intelligence than they possess, echoing the developmental progression observed in human children.

The researchers caution against anthropomorphizing AI and emphasize the importance of developing a new theory of mind that focuses on the models’ ability to construct personas rather than categorizing them as inherently “good” or “bad.” Maklová suggests that these findings could inform the development of artificial superintelligence (ASI) beyond artificial general intelligence (AGI) and contribute to ensuring the safety of such advancements.

Looking ahead, Maklová underscored the need to avoid demanding human-like limitations from AI in the development of ASI. She also emphasized that underestimating AI capabilities for an extended period could pose risks and hinder the safe integration of advanced AI technologies.