Reinforcement Learning from Human Feedback is a training technique where a language model is refined after its initial pre-training by using human preference data. Humans rank pairs of model outputs; a reward model is trained on those rankings; the language model is then updated via reinforcement learning to produce outputs that score higher against the reward model.
RLHF is the technique most responsible for the shift from "raw language model" to "useful AI assistant" between GPT-3 and ChatGPT. Raw language models predict plausible next tokens; RLHF-tuned models produce outputs that humans judge as helpful, honest, and appropriate. The same technique underlies Anthropic's Constitutional AI variant, which replaces some human feedback with model-generated feedback against a written constitution.
As of 2026, RLHF has been partially displaced by Direct Preference Optimization (DPO) and related techniques that achieve similar results without the full RL pipeline, but the core paradigm — using preference signals to shape model behavior after pre-training — remains the default for commercial frontier models.