On the Inconsistencies of Conditionals Learned by Masked Language Models