Michal Kosinski, a research psychologist at Stanford, is recognized for selecting timely research topics. His work aims to expand knowledge while also warning the public about potential threats posed by computer systems. Kosinski’s most notable projects examined how Facebook, now known as Meta, gained a profound understanding of its users through analyzing “likes” on the platform. Currently, his focus has shifted to exploring the surprising capabilities of AI. One of his experiments demonstrated that computers might predict a person’s sexuality by analyzing a digital photo of their face.
Kosinski, known to some through writings about Meta, recently discussed his latest paper, published in the Proceedings of the National Academy of Sciences. The paper presents a startling conclusion: Large language models, such as OpenAI’s, have supposedly begun using techniques akin to human thought processes. This represents a domain once thought exclusive to humans or other mammals. Specifically, Kosinski tested OpenAI’s GPT-3.5 and GPT-4 to determine if these models had developed a “theory of mind,” an ability developed by humans in childhood to comprehend the thought processes of others. This ability is crucial for understanding human thoughts accurately and preventing misunderstandings. If these models possess theory of mind, they are potentially a step closer to equaling or surpassing human cognitive capabilities. Kosinski’s experiments suggest that GPT-4, in particular, may have developed a theory of mind-like ability as an unintended consequence of its improving language skills, marking significant AI advancements with enhanced social skills.
Kosinski views his AI research as a natural progression from his earlier work on Facebook Likes. He clarifies that his studies focus on human behavior rather than social networks themselves. He believes that when companies like OpenAI and Google developed their generative AI models, they unintentionally trained models akin to human minds, as accurately predicting future words requires modeling human thought processes.
While Kosinski is cautious about asserting that LLMs have completely mastered theory of mind, his experiments presented classic challenges to the chatbots, which some managed competently. However, even the most advanced model, GPT-4, failed about 25% of the time. Despite these limitations, Kosinski likens GPT-4’s successes to the capabilities of 6-year-old children, which is noteworthy given the field’s infancy. He comments on the rapid progress of AI and speculates about when AI might achieve theory of mind or even consciousness, though the latter remains a controversial topic.
Kosinski suggests that if theory of mind emerged in these models, other abilities might also develop. This progress could enhance the models’ educational, influential, and manipulative capacities. He expresses concern about preparedness for AI systems that understand human thought processes, particularly if they surpass human self-awareness.
He emphasizes the distinction between humans, who possess intrinsic personality traits, and AI models that simulate personality. These models can adopt any personality, offering a significant advantage. Kosinski compares this dynamic to sociopaths, capable of mimicking emotions without genuine feelings, suggesting a potential for AI to become adept manipulators without remorse.
