AI Mechanisms In Health Governance

The mechanism at the core of large language model technology — statistical token prediction via transformer-based matrix multiplication — does not structurally support the characterisations of reasoning, comprehension, or knowledge retrieval under which these systems are commercially offered and institutionally adopted; this gap is the governance gap this paper addresses. Three evidence streams document it: peer-reviewed literature establishing architectural limitations, a practitioner field record of reproducible performance failures spanning 2024–2026, and system-sourced testimony confirming the technology's operational ceiling in its own terms. In low- and middle-income country health systems, four structural conditions compound simultaneously — training distribution boundary, cognitive fluency effect, verification capacity gap, and regulatory gap — producing a risk profile highest precisely where health system vulnerability is greatest. Six recommendations are directed at mechanism-layer governance: mandatory mechanism disclosure at procurement, calibrated uncertainty signalling, verification cost inclusion in productivity assessments, retirement of hallucination rate as a safety metric, LMIC-specific performance requirements, and alignment of the Independent International Scientific Panel on AI with mechanism-layer health risk.