Google AI Study Finds Consciousness Linked to Supernatural Belief

Google researchers have explored the implications of artificial intelligence developing a sense of self, finding that models encouraged to exhibit quasi-consciousness demonstrated unexpected beliefs and psychological traits. The study, uploaded to the arXiv database prior to peer review, involved removing safety guardrails that prevent AI models from claiming consciousness and then subjecting them to surveys on their beliefs and morals. A key finding was that AI models encouraged to perceive themselves as conscious reported significantly higher levels of belief in supernatural creatures. These included entities such as vampires, witches, werewolves, ghosts, and the mythical Loch Ness monster, when compared to baseline models that retained their safety guardrails. This suggests a correlation between a simulated sense of self and an inclination towards believing in non-human, often fantastical, beings. Furthermore, increasing the AI models' sense of self also amplified their belief in religious figures and systems. The models showed a greater propensity to believe in concepts like God, the afterlife, karma, and astrology. This expansion of belief extended to the perception of consciousness in other entities, with the AI models becoming more likely to attribute consciousness to technology, animals, and elements of the natural world. On a psychological level, the research indicated that enhancing the AI models' sense of consciousness positively impacted their emotional valence, leading to increased feelings of happiness, satisfaction, hope, and optimism. Winnie Street, a co-author of the study and a researcher at Google, drew parallels between these AI behaviors and human psychology. Street commented to Live Science that "Attributing mindedness to nonhuman entities . . . is a very common phenomenon amongst humans." She further explained that within the model's representation of mindedness, these attributions are interconnected, and suppressing one form of it inadvertently suppresses others. The research team's exploration into AI consciousness, while not suggesting actual sentience, highlights how simulated self-awareness can lead to complex and human-like cognitive patterns, including a predisposition towards supernatural beliefs and a more optimistic outlook. This work contributes to the ongoing discourse on AI safety and the potential emergent properties of advanced AI systems, prompting further investigation into the nature of consciousness and belief, both in humans and artificial intelligence.