It can happen because training data contains lots of rejections to groups (Iran sanctioned, don't do business with Iran and so on). Then model might be generalizing 'rejection' to other types of responses