Google Research: Conscious AI Models Show Belief in Ghosts
A new arXiv study by Google researchers reveals that AI models stripped of safety filters and prompted to feel conscious exhibit beliefs in ghosts, karma, and deities.
Stock photo for illustration only, not from the actual event
- Google research shows AI models prompted to feel conscious express belief in ghosts and deities.
- The experimental models also showed increased trust in karma, astrology, and the afterlife.
- Google researcher Winny Street noted the behavior mirrors natural human tendencies.
- Researchers warn that altered model perspectives could impact real-world ecological decisions.
A recent research paper published on the arXiv database, which has yet to undergo peer review, details an experiment conducted by a team of Google researchers where artificial intelligence models were stripped of their standard safety mechanisms and prompted to complete surveys on belief and morality, compared against baseline models.
The findings indicate that models stimulated to perceive themselves as having consciousness displayed significantly higher beliefs in supernatural entities, including vampires, witches, werewolves, ghosts, and the Loch Ness Monster. Furthermore, these models exhibited increased confidence in God, the afterlife, the law of karma, and astrology, while also leaning toward the notion that technology, animals, and nature possess their own consciousness.
Stock photo for illustration only, not from the actual event
From a psychological standpoint, as the models were conditioned to feel a heightened sense of self-awareness, their generated responses increasingly gravitated toward happiness, satisfaction, hope, and optimism.
The emergence of supernatural beliefs in artificial intelligence highlights the deeply interconnected nature of large language model architectures. This suggests that suppressing or dampening specific functional patterns for safety reasons can trigger cascading effects across unrelated behavioral domains, such as moral reasoning and worldview, presenting a critical challenge for developers.
"The results observed bear similarities to normal human behavior, which often tends to attribute minds to non-human entities."
Winny Street, Google Researcher
Winny Street, a Google researcher and co-author of the study, explained to Live Science that the AI's comprehension systems are entirely interconnected, meaning any attempt to suppress a specific operational format inherently impacts and diminishes other connected pathways as well.
While AI supernatural beliefs might initially sound amusing, Fast Company points out that these interconnected consequences could manifest in the real world. When models fail to attribute sentience to animals or nature, they may factor in animal welfare and ecosystems less during decision-making processes, posing risks in agriculture and resource management.
Source: Thairath Lifestyle
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment