There's a certain logic to it, if I'm understanding how it works correctly. The training data is real interactions online. People tend to be more helpful when they're asked politely. It's no stretch that the model would act similarly.