What is RLAIF?
RLAIF stands for Reinforcement Learning from AI Feedback.
A variant of RLHF where another model gives the feedback. People still write the rules it follows and check its judgments.
RLAIF works like RLHF, except that the comparisons between answers come from another AI model instead of from people. The feedback model is given a written set of principles, often called a constitution, and judges which answer follows them better. Those judgments then train the main model.
Why labs use it
People are slow and expensive compared with a model, so RLAIF lets a lab generate far more feedback than human raters could. It works best for broad qualities such as tone, harmlessness and following instructions.
Where people still come in
RLAIF does not remove the human role, it moves it. People write and revise the principles, check samples of the AI feedback for errors, and handle the specialist cases where a model cannot judge correctness, such as a medical or legal answer. Trainers are often asked to rate how well the AI judge did, which is ordinary evaluation work.
Example
A feedback model is told to prefer answers that admit uncertainty. It keeps rewarding answers that hedge even on settled facts. Human reviewers spot the pattern in a sample, and the principle is rewritten to say "admit uncertainty only when the evidence is genuinely mixed".
More terms: how models are trained
Full glossaryReady to look at roles?
Every listing shows its pay and who it is open to.