Constitutional AI

Using written principles and AI feedback to align models.

In Constitutional AI, how does the model learn to behave more safely?