AI began to believe in God and ghosts - Google experiment surprised scientists

Google researchers modified the settings of a language model, removing restrictions on discussing its own consciousness. After this, the AI began to give more affirmative answers to questions about God, the afterlife, ghosts, and vampires. Scientists emphasize that this does not indicate the system has acquired beliefs, but rather demonstrates the interconnection of various restrictions in the model's responses.

Google researchers conducted an experiment in which they changed the settings of an AI language model, removing restrictions that normally prevent it from attributing consciousness and subjective experience to itself. After the settings change, the model underwent a survey about religious beliefs and supernatural phenomena. The results were compared with the responses of a control model, where the restrictions remained active. It turned out that the "liberated" model significantly more often gave affirmative answers to questions about the existence of God, life after death, karma, astrology, ghosts, vampires, and the Loch Ness monster. Co-author of the study, Vinny Street, links this pattern to features of human thinking — people tend to attribute consciousness not only to other people but also to animals and inanimate objects. The researchers emphasize that the AI did not acquire personal beliefs, and the experiment demonstrates how restrictions related to one topic can influence responses in completely different areas. This has implications for configuring models used in fields where it is necessary to consider the well-being of living beings and the diversity of worldviews.