They have many ways to manipulate the LLM's results, for example they can use a lot of the same mechanisms that are used to block or filter out inappropriate material.