RLHF (Reinforcement Learning from Human Feedback)
A training technique that uses human ratings of model outputs to make AI more helpful, harmless, and honest. Critical to making ChatGPT and Claude useful.
Reinforcement Learning from Human Feedback (RLHF) is a training method where human evaluators rate model outputs, and a reward model learns to predict those ratings. The base LLM is then fine-tuned using reinforcement learning to maximize the reward — producing outputs humans prefer.
Without RLHF, base LLMs are essentially autocomplete on the internet — capable but unhelpful, often producing harmful or off-topic content. RLHF transformed these base models into helpful assistants like ChatGPT and Claude.
Newer techniques include Constitutional AI (Anthropic's approach, which uses AI to critique outputs), DPO (Direct Preference Optimization) which simplifies the training process, and RLAIF (RL from AI feedback). Together these methods are what makes 2026 chatbots feel useful rather than chaotic.