Executive Abstract: This paper examines multi-dialect stress-testing of enterprise language systems, the cross-border privacy boundarybetween Saudi PDPL, Moroccan Law 09-08 (CNDP), and European processing, and the human-oversight duties that EU AI Act Article 4 and Annex III (Section 3) place on institutional deployers. Standard English-to-Modern Standard Arabic (MSA) benchmarks mask severe operational failures when conversational agents and Retrieval-Augmented Generation (RAG) pipelines encounter authentic intra-prompt code-switching (Moroccan Darija + French, Saudi Najdi/Gulf + English, Egyptian, and Levantine). Evaluating 7 foundation models against the Arabic AI Observatory v2.0 Master Corpus (521 deterministic adversarial scenarios), we document a 37.0% accuracy degradation under dialectal negation shifts, a 2.41x token fertility infrastructure tax (1.18 vs. 2.85 tokens/word), and critical cross-border PII exfiltration vulnerabilities. Finally, we specify a zero-knowledge SHA-256 provenance protocol anchored on WaqfLedger.tech.
No takes yet. Share an insight, caveat, or question.
Adel EL Bouzagaoui (2026) studied this question.