RLHF

RLHF stands for reinforcement learning from human feedback. It is a way of training an AI model in which people look at its answers, rate them or pick the better of two, and the model is then adjusted to give more of the answers people preferred. Think of a new volunteer who gets feedback from a coordinator after each phone call. Over time, they learn what a good call sounds like. RLHF is one of the main ways chat assistants are taught to be helpful and to turn down harmful requests.

Why it matters

RLHF makes a chat assistant behave better on average, but it does not guarantee good behavior. People still find a jailbreak that gets many models to do things they were trained to refuse. A polite, careful assistant does not prove that it will always say no to a harmful request. When you compare AI tools, ask the vendor what safety checks they add beyond the model’s training and how they test whether those checks hold up. RLHF is closely tied to alignment and fine-tuning.