From AI Dependence to AI Literacy

Addressing Neural Machine Translation Biases in English–Arabic Digital Media Translation

Abstract

This empirical study examines how 45 Moroccan undergraduate students in Applied Foreign Languages translated English digital media content into Modern Standard Arabic (MSA). Analysing 360 translations across eight genres; including social media posts, streaming content descriptions, and mobile application interfaces, two independent raters applied a seven-category taxonomy adapted from Multidimensional Quality Metrics (MQM) frameworks, with strong interrater agreement (Cohen’s κ = 0.84). Findings reveal systematic error patterns consistent with documented neural machine translation (NMT) biases: over-formalisation (73% of translations), inappropriate literal transfer of cultural references (68%), and platform constraint violations (62%). Drawing on an extended PACTE translation competence model, the study argues that students lack the critical AI literacy required to evaluate machine translation output; particularly regarding register, cultural adaptation, and platform-specific conventions, despite ready access to tools such as Google Translate and ChatGPT. Follow-up interviews indicate that students evaluated AI output primarily on surface grammatical correctness, demonstrating insufficient strategic competence to identify functionally inappropriate output. A pedagogical framework is proposed that embeds NMT bias awareness, post-editing skills, and AI tool evaluation as assessable competency milestones. The results call for translation programmes to reposition critical AI literacy from a peripheral concern to a core curricular component structured around human–AI complementarity.