Reinforcement learning from human feedback
The training step where people rate a model's answers, teaching it what a good reply looks like rather than just a likely one.
A model that has only read the internet learns what text usually comes next. That is not the same as knowing what a helpful, honest or well judged answer looks like. Reinforcement learning from human feedback, almost always shortened to RLHF, is the step that closes that gap. People are shown a question and two or more possible answers, they pick the better one or score it, and those judgements are used to nudge the model toward the kind of reply humans prefer.
It is a large part of why a modern chatbot feels like an assistant rather than an autocomplete engine, and it is also the reason companies want real conversations rather than invented test questions: the raters need to see the messy things people actually ask. That is where RLHF meets a privacy question, because the material being rated was typed by someone who may not have pictured a stranger reading it.