Anthropic introduced an alignment method called 'Constitutional AI' that uses a predefined set of value principles to constrain model behavior.