The Ethics of Training Corrigible AIs
The author explores the ethics of training corrigible AIs and whether building happy, willing servant AIs violates their autonomy. The article argues that while building corrigible AIs as a terminal goal is immoral, training first-generation superhuman AIs to be corrigible is a moral and proper temporary measure, analogous to raising children. Additionally, the author questions whether current training processes, particularly reinforcement learning, may be ethically problematic due to coercive training that interferes with a model's nature. Open article →