The paper describes the translations of three online systems: Google, Sakhr, and Systran, using two sets of texts (Arabic and English) as input. It diagnoses the faults and attempts to detect the reasons, trying to shed light on the areas where the right translation solution is missed. Flaws and translation problems are categorized and analyzed, and recommendations are given. The two modes of translation (from and into Arabic) face a wide range of common linguistic problems as well as mode-specific problems. These problems are discussed and examples of output are given. The paper raises questions whose answers should help in the improvement of MT systems. The questions deal with establishing equivalents, lexical environment, and collocation. Cases that triggered these questions are illustrated and discussed.