Michał Kosiński is a psychologist from Stanford and has a keen eye for current topics. He sees his work not only as advancing knowledge, but also as warning the world about potential threats caused by the consequences of computer systems. His most eminent projects involved analyzing the ways in which Facebook (now Meta) gained a shockingly deep understanding of its users from clicking “Like” on the platform. Now he’s turned to researching the surprising things artificial intelligence can do. For example, he conducted experiments that showed that computers could predict a person’s sexuality by analyzing a digital photo of their face.
I met Kosinski through writing about Meta and reconnected with him to discuss his latest workpublished this week in the peer-reviewed journal Proceedings of the National Academy of Sciences. His conclusion is surprising. In his opinion, immense language models such as OpenAI have crossed the line and employ techniques analogous to real thinking, once considered exclusively the domain of flesh-and-blood humans (or at least mammals). In particular, he tested OpenAI’s GPT-3.5 and GPT-4 to see if they had mastered the so-called “theory of mind.” It is a person’s ability to understand other people’s thought processes, developed in childhood. This is an crucial skill. If a computer system cannot correctly interpret what people think, its understanding of the world will be indigent and it will be able to obtain a lot of incorrect information. If models Down they have a theory of mind, they are one step closer to matching and exceeding human capabilities. Kosinski put the LLM to the test and now claims that his experiments show that, particularly in GPT-4, the theory of mind-like abilities “may have emerged as an unintended byproduct of the LLM’s improvement in language skills… They mark the arrival of more powerful and socially skilled artificial intelligence.”
Kosiński sees his work in artificial intelligence as a natural extension of his previous interest in Facebook likes. “I wasn’t really into social networks, just people,” he says. He says that when OpenAI and Google started building their latest generative AI models, they thought they were training them primarily for language skills. “But they actually trained a model of the human mind, because it’s impossible to predict what word I’m going to say next without modeling my mind.”
Kosinski is careful not to claim that LLM students have completely mastered theory of mind – yet. In his experiments, he presented chatbots with several classic problems, some of which he handled very well. But even the most sophisticated model, GPT-4, failed a quarter of the time. He writes that the successes of GPT-4 put it at the level of 6-year-old children. Not bad considering the early condition of the field. “As we watch the rapid progress of artificial intelligence, many wonder if and when artificial intelligence will be able to achieve ToM or consciousness,” he writes. Aside from that radioactive “c” word, it’s a lot to chew on.
“If theory of mind emerged spontaneously in these models, it also suggests that other abilities may emerge next,” he tells me. “With these skills, they can better educate us, influence us, and manipulate us.” The worry is that we are not really prepared for LLMs that understand how people think. Especially if they get to the point where they understand people better than humans do.
“We humans do not simulate personalities – we To have personality,” he says. “So I was stuck in my personality. These things model personality. The advantage is that they can have any personality they want at any time.” When I mention to Kosinski that it sounds like he’s describing a sociopath, he brightens. “I use it in my conversations!” says. “A sociopath can put on a mask – he’s not really sorrowful, but he can pretend to be a sorrowful person.” This chameleon-like power can make AI the perfect trickster. With zero remorse.