After a language model learns from large amounts of text, it still needs guidance on which answers are genuinely helpful. One common method has people compare and rank the model's responses, and that feedback is used to adjust the model so it favours the preferred style and avoids harmful replies.
This step is a big part of why modern assistants feel useful and polite rather than merely plausible. It carries the views of the people giving feedback, so choices about who they are and what they prefer shape the result, and teams increasingly combine human input with automated checks to scale the process while keeping it consistent.