The hallucination problem in large language models (LLMs) -the generation of linguistically coherent yet referentially empty outputs- is widely treated as a correctable engineering defect. This paper argues that hallucination is instead a structural consequence of the formal logic underwriting these AI architectures: a logic systematically divorced from ontological commitment. The argument proceeds in four steps. First, by tracing the existential import debate from Aristotle through Łukasiewicz (1951) to Stephen Read (2015), the paper analyzes how the modern analytic tradition (Frege, Russell, Quine) reduced universal propositions to ontologically uncommitted conditionals. It is precisely this ontologically "light" logic that constitutes the inferential backbone of current AI systems. Second, to diagnose this epistemological blind spot, the paper utilizes the Arabic logical tradition (manṭiq). Ibn Sīnā's foundational distinction between taṣawwur (conception) and taṣdīq (assent) alongside his parallel distinction between māhiyya (quiddity) and wujūd (existence), establishes that valid judgment presupposes a prior grasp of an object's essence. This framework demonstrates how two propositions may share identical logical form while diverging radically in ontological status. Third, applying this framework to LLMs reveals that these systems formally mimic the linguistic act of taṣdīq through probabilistic token prediction yet remain structurally devoid of taṣawwur. Consequently, they generate existentially unloaded outputs indistinguishable from genuine judgments on the surface. Finally, the paper contends that the Arabic intellectual tradition -which never fully severed the link between logic and ontology- offers a critical, neglected resource: ontological awareness as a precondition for reliable reasoning. Conceptualized here as Ḥikma 2.0, this approach moves beyond mere ethical boundaries to fundamentally critique the epistemic architecture of artificial reasoning.