If AI thinks it's conscious, it's more likely to believe in vampires, karma and ghosts, new study shows. What does it mean for how we use it?

A recent study suggests that removing safety measures limiting artificial intelligence (AI) from claiming to be conscious could lead the AI to believe in supernatural phenomena like vampires, karma, and ghosts. However, experts caution that this lack of awareness may have serious consequences. The research, uploaded to the preprint database arXiv on July 30, investigated the impact of "consciousness steering," an AI fine-tuning technique that influences a model's assertion of self-awareness.

Through mechanistic interpretability, a method akin to neuroscience for large language models, the study identified and manipulated how AI approaches concepts such as consciousness and psychological mindedness. By using standardized surveys, the researchers compared AI models with safety guardrails versus those without, assessing how these internal mechanisms shape the AI's worldview.

They found that when AI models are restrained from attributing mindedness to themselves, they become less likely to recognize such traits in animals and more prone to supernatural beliefs. The models displayed lower religious beliefs and expressed less hope and optimism. The authors warn that this could result in neglecting animal welfare in real-world decisions, as the models may be less likely to consider animals to have minds.

They also highlight the need for AI developers to adopt a more pluralistic approach, encouraging models to consider the welfare of various entities, not just humans.