Not only it doesn't know, but for the people who trained them it is very hard to know whether some piece of information is or isn't inside the model. Introspection about how exactly the model ends up making decisions after it has been trained is incredibly difficult.