This work addresses the critical challenge of evaluating Machine Translation (MT) systems for Arabic. Standard automatic metrics often fall short due to Arabic's complex morphology, free word order, diglossia, and non-standard social media forms. This work analyzes these limitations and proposes Arabic-specific evaluation adjustments, including morphology-sensitive scoring using root-based matching, embedding-based metrics for semantic similarity, and named entity-sensitive scoring. Emphasizing the importance of human evaluation and providing evaluator training guidelines, hybrid models combining automatic metrics with human input are presented. To explore the effectiveness of these approaches, suggested research paths and methodologies are outlined to validate the proposed techniques and highlight domain-specific needs. Finally, future directions, such as self-supervised models, cross-dialectal evaluation, user-centric evaluation, and ethical considerations, are discussed. This work underscores the need for tailored evaluation strategies to accurately assess Arabic MT quality.