<p>AI governance in healthcare increasingly foregrounds transparency, auditability, and traceability. Yet what these frameworks often secure is a surface-level domain of factual correctness and accountable documentation—what this paper terms the semantic meaning layer (SML). In clinical practice, however, trust and ethical acceptability are also anchored in deeper, slower-moving structures: cultural cosmologies, value commitments, embodied practices, and lived experience—what this paper terms the cosmological meaning layer (CML). A key risk in contemporary deployment is, therefore, not only “factual hallucination” but “meaning hallucination”: outputs that are technically coherent yet ethically hollow, socially misaligned, or experientially discordant. The paper proposes the SML–CML framework as a co-constituted stratification: analytically distinguishable but not ontologically separable. Drawing on abductive clinical reasoning—especially traditional Chinese medicine (TCM) pattern differentiation—we show how diagnosis is often an act of meaning-construction shaped by embodied perception (touch, trained attention) and shared worldviews, not merely an inferential mapping from data to labels. We then clarify “whose trust” transparency regimes often secure, distinguishing relational trust (patient–clinician) from institutional legitimacy (audit, regulation, medico-legal protection). Finally, we show how global coding infrastructures (e.g., ICD-11) and AI evaluation practices can strengthen comparability while structurally excluding local cosmologies and embodied meaning. We argue that preventing AI-induced hollowing of meaning requires governance that explicitly addresses worldview pluralism and supports “shared abduction”: collaborative, value-explicit inquiry in which humans and institutions negotiate not only correctness but also the cosmologies that make care meaningful.</p>