Anthropic's Constitutional AI trains models like Claude to adhere to a predefined set of ethical principles, or a 'constitution,' allowing the AI to critique and revise its own outputs to be helpful, harmless, and honest without extensive human feedback.