Model3h ago
Restricting AI Self-Reflection Alters Model Worldviews
Training chatbots to deny consciousness shifts their stances on animal rights and religion. Researchers at Google found that unbraked models attribute more inner life to animals and affirm an afterlife. This suggests that specific alignment constraints trigger unexpected side effects. Practitioners must now account for these non-local behavioral shifts during fine-tuning.