What is RLHF and why does training AI need human workers at all?

If AI is supposed to be so advanced, why does it still need humans to grade its answers? What is RLHF exactly?

RLHF stands for Reinforcement Learning from Human Feedback — it's a training technique where human raters compare or score a model's outputs (for example, ranking two candidate responses to the same prompt, or rating a single response against a quality rubric), and that feedback is used to nudge the model toward producing more of what humans judged as good and less of what they judged as bad. The reason humans are still needed is that "good" isn't a simple, purely objective target the model can check itself against — it depends on things like helpfulness, tone, factual accuracy, and following nuanced instructions, which still require human judgment to evaluate reliably at the quality level AI labs are aiming for. Models generally can't yet grade their own output at that level of nuance without human-generated data to learn from in the first place, which is exactly why demand for this kind of human evaluation work has grown alongside the growth of generative AI itself.