Between the Quoted, the Computed, and the Generated: Protecting Classical Arabic Heritage from AI Hallucination

When a reader today asks a frontier large language model about a Qur’anic verse, the authenticity and transmission chain (isnād) of a prophetic tradition (ḥadīth), or a pivotal event in early Islamic history, they encounter a hazard unprecedented in fourteen centuries of Arabic scholarship: eloquent hallucination. Because an autoregressive language model is a probabilistic next-token predictor that makes no structural distinction between verified historical record and plausible linguistic sequence, it can fabricate a prophetic tradition that was never narrated, splice together a transmission chain between two scholars who lived a century apart, or attribute a commentary to Al-Ṭabarī or Ibn Kathīr that appears nowhere in their works—all delivered in polished classical prose that easily deceives a non-specialist.

In software prototyping or casual copywriting, a model’s fabrication is quickly spotted and cheaply fixed. In canonical scripture, classical philology, and historical transmission, blurring the line between verified text and statistical guesswork destroys the very foundation of scholarly trust (amānah ʿilmiyyah) upon which Islamic historiography and transmission science (ʿilm al-isnād) were built.

When we set out to engineer the Mishkāt Epistemic Trilogy—spanning the mishkat.qa Qur’anic knowledge graph, the Al-Jāmiʿ Hadith encyclopedia (indexing 715,790 traditions across 425 primary collections), and the Early Islamic History Corpus—we established an uncompromising architectural constitution: *a strict structural and visual separation between three epistemic layers: the Quoted (al-Manqūl), the Computed (al-Maḥsūb), and the Generated (al-Muwallad).*