Training approach where feedback labels are generated by AI systems instead of humans, often used to scale preference data for alignment.