It doesn't try to do anything. It doesn't work like that. It regurgitates the most likely tokens found in the training set.
It is a first principle though so it helps to “stir the context windows pot” by having it pull in research and other shit on the web that will help ground it and not just tell you exactly what you prompt it to say.