His point is that it's only overfitting if the model won't return new content when you clarify you're not just asking for the obvious answer from the context.