Reinforcement Learning from Human Feedback (RLHF)
A training technique that uses people's judgements to teach a model which responses are helpful, honest, and appropriate. It is a large part of why modern chatbots feel polite and useful rather than merely fluent.
A raw language model, freshly trained to predict the next word, is fluent but unruly. It has no particular sense of what makes an answer helpful, safe, or polite. Reinforcement learning from human feedback is the step that tames it. People are shown different responses the model might give and asked which is better; those judgements are distilled into a reward signal, and the model is then trained to produce the kinds of answers people preferred.
This is why the assistants most of us use feel cooperative and reasonably well-mannered rather than like an autocomplete that has read too much of the internet. The human preferences steer the model away from unhelpful, offensive, or evasive replies and towards ones a person would actually want.
RLHF is one of the main tools in AI alignment, the broader effort to make systems behave in line with human intentions. It is powerful but not perfect: the model only learns from the feedback it is given, so blind spots and biases in that feedback can carry through into how it behaves.