Machine Translation (MT) of literary texts is a rich area of research that has not been investigated extensively. The conventional wisdom is that literary translation by nature is challenging even for the most competent translators as it requires preserving not only the meaning, but also the style of the author. This makes the MT output of any literary work viewed as inadequate no matter how accurate it is in conveying the meaning. However, while the standards of literary translation are still high for machines, the recent advances in MT seem promising and could open the door to an adequate post-edited MT for literary texts. This is a descriptive study that looks into the limitations and potential of MT in facilitating the task of literary translators. It analyzes the language errors in the Arabic MT of selected short English literary texts. The MT is generated from three systems: IBM Watson, Bing Microsoft, and Google Translate. The language errors in the raw MT output are detected by two professional post-editors, and classified by the researcher according to Temnikova's (2010) cognitive approach of MT post-editing. The findings of the study show that the highest number of errors fall under categories that are cognitively easy to correct by post-editors. It is hoped that this contribution can be useful to build MT tools with more adequate output for English-Arabic translation in general and literary translation, in particular.