Unwarranted beliefs—those pseudoscientific, conspiratorial notions we cling to despite evidence—are a pervasive human quirk. But what if AI could help us break free from these mental traps? A fascinating new study explores how large language models (LLMs) stack up against humans when it comes to recognizing logical fallacies and conspiratorial thinking. Researchers put several LLMs, including GPT-3.5, GPT-4, and Gemini, through a psychometric test called PEUBI, designed to assess belief in the paranormal, pseudoscience, and conspiracies. The results? AI outperformed humans, showing a stronger ability to identify unwarranted beliefs. This intriguing finding raises questions about the nature of rationality in both humans and machines. While LLMs sometimes display inconsistencies in their reasoning, their overall performance on the PEUBI test suggests a unique form of "unstable rationality." This means their reasoning abilities, while seemingly present, are easily disrupted by changes in language or contradictory information. The researchers propose that this "unstable rationality" might actually be a useful tool for persuasion. They experimented with using LLMs to challenge unwarranted beliefs by leveraging psychological principles like cognitive dissonance, the discomfort we feel when holding conflicting beliefs. Early results suggest that LLMs could potentially act as personalized guides, helping us navigate away from misinformation and towards a more rational understanding of the world. While more research is needed, this study opens exciting possibilities for using AI to enhance critical thinking and combat the spread of false beliefs. The future of fighting misinformation might just lie in embracing the unstable rationality of our AI companions.