A model trained only to predict text does not automatically behave helpfully or safely. One common step is to show people pairs of responses, record which they prefer, and use those judgments to nudge the model toward answers people actually want.
This shapes tone, helpfulness and refusal of harmful requests, but it inherits the biases and blind spots of whoever gave the feedback. It is a powerful lever and a reminder that a model's manners reflect human choices, not an objective standard.