Michal Kosinski is a research psychologist at Stanford University who is known for his timely and relevant research topics. He not only aims to advance knowledge, but also to raise awareness about potential dangers that may arise from computer systems. One of his most well-known projects involved analyzing how Facebook (now Meta) was able to gain a deep understanding of its users through their likes on the platform. Recently, he has shifted his focus to studying the surprising capabilities of artificial intelligence (AI).
In his latest paper, published in the peer-reviewed Proceedings of the National Academy of Sciences, Kosinski discusses his experiments with OpenAI’s GPT-3.5 and GPT-4 models. He claims that these large language models (LLMs) have crossed a boundary and are now using techniques that were once thought to be exclusive to human thought processes. Specifically, he tested their ability to understand “theory of mind,” which is the ability to understand the thoughts of others. This is an important skill, as it allows humans to better understand and interact with each other. If AI models can also master this skill, they may be one step closer to surpassing human capabilities.
Kosinski’s interest in AI stems from his previous research on Facebook Likes. He believes that he was not studying social networks, but rather, humans themselves. When OpenAI and Google began developing their latest generative AI models, they thought they were training them to handle language. However, Kosinski argues that they were actually training a model of the human mind, as predicting the next word in a sentence requires understanding the thought process behind it.
While Kosinski does not claim that LLMs have fully mastered theory of mind, his experiments show promising results. GPT-4, the most sophisticated model, performed on par with 6-year-old children in some tasks. This is impressive considering the early stage of AI development. Kosinski acknowledges that many people are curious about whether and when AI will achieve consciousness, but he believes that the emergence of theory of mind in these models is a significant step towards more socially skilled and powerful AI.
