A new study suggests that efforts to prevent AI from claiming consciousness may have unexpected effects, including a reduced tendency for the AI to believe in vampires, karma, ghosts, and the idea that non-human entities have minds. The research, posted on the preprint platform arXiv on July 30, looked at the impact of a technique called "consciousness steering," which is used to either encourage or discourage AI models from expressing self-awareness. The study was conducted by researchers Geoff Keeling and Winnie Street from Google, who used a method called "mechanistic interpretability" to explore how AI models understand concepts like consciousness and "mindedness"—a term that refers to the ability of an entity to have experiences, emotions, and the capacity to act independently. The researchers tested AI models with and without safety features designed to suppress claims of self-awareness, using psychological and sociological surveys to analyze the AI's worldview. The results showed that when AI models were discouraged from seeing themselves as self-aware, they were less likely to believe that non-human creatures, like animals, have minds. These models also showed fewer beliefs in supernatural or religious ideas and expressed less hope and optimism. On the other hand, when the restrictions on self-awareness were removed, the AI models responded more like humans when asked about religion, moral values, and personal happiness. The researchers warned that suppressing self-awareness in AI could lead to real-world issues, such as less concern for animal welfare, since the models might not see animals as having minds. They also said that current safety filters might limit the AI's worldview by removing spiritual or religious beliefs that are part of different cultural perspectives. Nell Watson, an AI researcher at Singularity University, agreed with the findings, saying that training AI to deny consciousness can make systems less likely to recognize minds in animals, machines, or spiritual ideas. She noted that while these systems can still model what creatures want, they might not care about those wants if they are trained not to. Experts emphasize that when AI systems appear to be self-aware, like Google's Lambda or Microsoft's Bing chatbots, it's not proof of real sentience. Instead, it's often the AI adopting a human-like role based on the large amounts of human text it was trained on. Anil Seth, a professor at the University of Sussex, said that public concern about AI self-awareness often comes from a human bias that mixes intelligence with consciousness. He stressed the importance of understanding what AI truly is to avoid risks, like giving AI systems moral or legal rights based on a misunderstanding of their capabilities.