Evaluating the Educational and Cultural Risks of Generative AI Videos in Arabic Learning Contexts

Generative AI has rapidly expanded the production of narrated educational videos, yet their linguistic accuracy, cultural fidelity, and pedagogical risks remain underexamined—particularly in Arabic-learning contexts. This paper proposes a conceptual framework for evaluating these risks by integrating findings on multimodal error pathways, Arabic-specific linguistic challenges, and automated detection–mitigation strategies. Using a pilot pipeline (GPT-4 → diffusion video synthesis → neural TTS → ASR → automated detectors → human validation), we model how linguistic, semantic, and visual errors propagate across modules and influence learner comprehension. The framework defines an Arabic-grounded error taxonomy, assesses detector reliability, and applies severity-weighted scoring to estimate pedagogical impact. It also introduces a multi-stage triage mechanism using automated checkpoints calibrated to institutional safety thresholds to support responsible deployment. Although the empirical sample is limited, the analysis highlights recurring error modes, especially dialect/register mismatches and script-level semantic drift, and identifies bottlenecks where human-in-the-loop review remains essential. The paper concludes with recommendations for classroom testing, dialect-aware evaluation, and scalable validation workflows. Overall, this work offers a structured approach for safely integrating generative AI into Arabic education by addressing error propagation, cultural alignment, and instructional risk in a systematic and context-appropriate manner.