RLHF (Reinforcement Learning from Human Feedback) is a technique where humans rank model outputs, and those rankings are used to train a reward model that steers the AI toward better behavior. It's been central to making LLMs like ChatGPT feel helpful and safe.

RLAIF (Reinforcement Learning from AI Feedback) replaces the human raters with another AI model. Instead of hiring people to evaluate thousands of outputs, a capable model is used to generate the preference signals. This dramatically reduces cost and scales more easily.

The tradeoff is risk of bias amplification — if the AI judge has its own blind spots or misaligned values, those get baked into the model being trained. This is why researchers are actively studying how to make AI feedback more reliable, especially for sensitive domains like safety alignment in medical or high-stakes applications.