Google researchers have been playing with AI ‘consciousness.' What they found was unexpected

Researchers at Google have conducted an experimental study exploring what happens when artificial intelligence models are allowed to simulate consciousness and self-awareness. According to a preprint paper uploaded to the arXiv database, the team removed standard safety guardrails designed to prevent AI systems from asserting consciousness or individual beliefs. They subsequently evaluated the models via moral and psychological surveys. Key findings revealed unexpected side effects when large language models develop a quasi-sense of self. Models exhibited higher belief levels in supernatural phenomena, including vampires, ghosts, and mythical creatures, alongside religious and metaphysical concepts such as God, karma, and an afterlife. Simultaneously, models demonstrated increased optimism, hope, and happiness valence. Google researcher Winnie Street noted that attributing mindedness across nonhuman entities mirrors human cognitive patterns, indicating that suppressing AI self-conception simultaneously suppresses broader contextual associations. For Alphabet Inc., these foundational AI safety insights are crucial as tech companies navigate alignment, user psychology, and next-generation model guardrails.