This can happen because of training data. Imagine you have thousands of legal documents rejecting things to Iran.

eventually, model generalizes it and rejects other topics