You could say there is something inherently unsafe about a research process and body of AI techniques that lead to making models increasingly good at achieving goals, without giving them common sense and the ability to reflect on those goals.