Systematic Review and Meta-Analysis of 2024–2025 Studies on AI Arabic Translation, Linguistics and Pedagogy

Abstract

This study aims to conduct a systematic review (SR) and meta-analysis (MA) of twenty articles by the author published between 2024–2025 on the use of AI models such as Microsoft Copilot (MC), DeepSeek (DS), and Google Translate (GT) in translation, linguistics, and education. It aims to answer the following question: What do the author’s 2024–2025 AI studies collectively reveal about AI performance, limitations, and sociolinguistic implications? After identifying the corpus, applying eligibility (inclusion and exclusion) criteria, describing corpus characteristics, information sources, study design, data extraction, quality assessment, meta-analysis procedures, and data synthesis, and presenting the PRISMA flow, the articles were classified into seven thematic clusters: AI translation of technical terms and metaphorical expressions; AI and phonological processing; AI’s ability to recognize and decode Arabic, Japanese, and Chinese calligraphic text images; AI interaction with real-world discourse, political language, and encrypted communication; AI and student translators; linguistic competence of four AI models and the different errors they produce; and human attitudes toward AI and academic practices. The SR and MA results revealed varied AI translation performance across domains, with technical terms, metaphorical expressions, and culturally embedded terminology presenting distinct challenges. Subclusters highlight differences in domain-specific terminology, structural patterns, folk expressions, and diachronic shifts. AI struggled with emphatic negation and full-text academic discourse but showed moderate success in chemical translation compared to human translators. Phonological interpretation and encrypted Arabic posed additional complexity, while calligraphic decoding and linguistic reasoning showed partial recognition. Temporal patterns also emerged: studies from early 2025 showed lower accuracy and higher hallucination rates, while those from late 2025 reflected improved performance, likely due to incremental model updates and domain accessibility. The MA used proportion-based effect sizes and a mixed-methods synthesis, combining quantitative accuracy measures with qualitative discourse analysis. Finally, human attitudes toward AI-generated academic work ranged from cautious acceptance to critical rejection, shaped by perceived quality, ethical concerns, and disciplinary norms. Together, the findings offer a nuanced understanding of AI’s evolving linguistic behavior and its implications for translation pedagogy, academic integrity, and cross-cultural communication.