Anthropic uses that AI evaluation dataset to train a preference model that helps fine-tune Claude to consistently output responses that conform to its constitution’s principles. In February 2025 ...