Glossary
RLAIF (reinforcement learning from AI feedback)
An alignment method in which the preference judgments come from a judge model applying written principles, rather than from human raters.
A judge model compares or scores candidate responses, and those judgments train the target model the way human labels would in RLHF. Constitutional AI is the best-known variant, with the judge applying an explicit list of principles. RLAIF scales cheaply, but the judge’s biases become the trainee’s, so human preference data is still used to calibrate and evaluate it.
Why it matters
Even with AI feedback in the loop, human judgment stays the reference standard the AI judge is checked against.