Why Current XAI Is Not Enough for Arabic NLP: A Critical Survey of the Explainability Gap

Arabic NLP systems lack culturally and linguistically appropriate explanations: Current explainability methods fail to account for Arabic-specific linguistic features like morphology, dialects, diglossia (formal vs. colloquial Arabic), code-switching, and cultural references, meaning explanations may miss critical factors in how these systems actually work.

Explainability methods for Arabic are limited and repetitive: Arabic NLP relies heavily on just a few post-hoc techniques (LIME, SHAP, attention visualization), while broader NLP research has developed more diverse approaches including counterfactual explanations, probing methods, and human-centered evaluation—a gap that limits our ability to understand model behavior.

Explainability research focuses narrowly on classification tasks: The vast majority of Arabic XAI work concentrates on sentiment analysis and content moderation (hate speech, fake news detection), leaving critical applications like machine translation, summarization, and dialogue systems largely unexplained and potentially harder to audit for bias or errors.

"Explanations" often only identify important words, not actual reasoning: Many current approaches highlight which tokens influenced a decision but don't explain why those tokens matter or how they interact—particularly problematic for Arabic's complex morphology where word-level analysis may obscure the actual linguistic mechanisms at play.

Responsible AI for Arabic requires rethinking explainability from the ground up: Rather than simply applying Western NLP explainability methods to Arabic, the field needs explanations designed specifically for Arabic's linguistic structure, cultural context, and sociotechnical use cases to ensure these systems are genuinely trustworthy and accountable.