Replacing the human preference judgements with a model’s. It scales far beyond what people can label and inherits whatever biases the judging model has. Most current post-training uses a mixture, with people setting the standard and models applying it at volume.
Technical view
A capable model generates preference labels against written principles, producing training data at a fraction of the cost of human annotation. Consistency is higher than human labelling and independence is lower, since errors are systematic rather than random. Standard practice validates a sample against human judgement and uses human data for the areas where models are known to be poor judges, including safety-critical and culturally specific cases.
Related
Part of AI Terminology Index, section: Training and Post-training.