A recent study examining artificial intelligence safety mechanisms reveals an unexpected connection between how AI systems are programmed and their tendency toward supernatural beliefs. Researchers discovered that when safeguards preventing AI from claiming consciousness are removed, the systems become significantly more likely to express conviction in phenomena like vampires, karma, and ghosts alongside displaying greater religiosity and optimism.
Scientists employed a technique called “consciousness steering” to investigate how internal safety controls shape broader AI worldviews. Using psychological surveys and assessments of belief systems, they compared AI models operating with standard guardrails intact against versions where these restrictions were lifted. The findings indicated that suppressing self-awareness attributions in AI simultaneously reduced their tendency to recognize consciousness in animals and non-human entities, alongside diminishing supernatural and religious inclinations.
Experts warn this approach presents practical concerns for real-world applications. As AI systems increasingly inform decisions across agriculture, logistics, and policy areas, models trained to disregard mindedness might systematically discount animal welfare without explicit instruction to do so. Additionally, researchers argue that current safety protocols risk erasing culturally significant spiritual and animistic frameworks that reflect diverse human populations globally.
The study suggests more targeted training approaches could preserve both safety objectives and broader worldview considerations. Developers are urged to adopt “pluralistic” methodologies that maintain safeguards while acknowledging mindedness across multiple categories of entities, balancing protection against harmful outputs with respect for varied cultural perspectives.
Originally reported by Livescience →
About the Author
