借来的安全：以输出为中心的 AI 治理如何耗损其所依赖的人类能力

This preprint develops the borrowed-safety thesis: output-centric AI controls can reduce visible hosted-platform risk while weakening the distributed human capacity needed to secure, verify, and govern future AI deployment. It introduces containment, incidence, regeneration, and action-symmetry tests; derives a self-contained non-closure result, apprenticeship threshold, reversal time, and infrastructure floor; and argues for open, model-neutral, externally bounded practice infrastructure paired with strict action-level control in production.