走向实证：以有界主张服务公共善的 AI 治理框架

My Analysis of Responsible AI and Institutional Failures

I'm struck by how Artificial Intelligence transcends mere governance concerns; it exposes systemic institutional weaknesses. AI deployment becomes an intervention, potentially repairing, exacerbating, substituting for, or even concealing institutional failures. This necessitates a dual evaluation: the system itself and the institutional rupture it intersects. The shift from principles to protocols, such as the EU AI Act, NIST AI RMF, and ISO/IEC 42001, is underway, translating commitments into concrete roles, requirements, oversight, and evaluation practices. However, critical questions arise: what do these protocols truly establish, whose power remains unchecked, and where should measurement cease? The broader moral framework offered by Pope Leo XIV's Magnifica Humanitas, centered on dignity, technological power, and the common good, informs my development of a "rupture test" linking institutional baselines to system evaluation. I distinguish between "evidence-bounded deployment," restricting claims to evaluated elements, and "measurement-bounded governance," which establishes overriding constraints, even in the presence of favorable evidence. Within these boundaries, RISE AI offers an architecture for formulating bounded, evidence-based claims regarding Responsibility, Inclusivity, Safety, and Empowerment. Ultimately, Responsible AI demands better engineering, institutional repair, and continuous ethical and political judgment.

人工智能所带来的不仅是一个治理难题。它还能揭示出各类制度在哪些方面早已未能提供响应性、归属感、关怀和问责。人工智能一旦部署，便成为对这些现实状况的一种介入。它可以修复、加剧、替代或掩盖它所遭遇的失灵。因此，负责任的人工智能必须同时评估该系统以及它所切入的制度性断裂。从原则走向规程的转变已经在进行之中。《欧盟人工智能法案》、NIST AI RMF、ISO/IEC 42001 以及各项保障实践，正在将承诺转化为角色、要求、记录、监督和评估。更难回答的问题在于，这些规程究竟确立了什么、未曾触动谁的权力，以及度量必须在何处止步。教宗利奥十四世的《Magnifica Humanitas》提供了一个以尊严、技术权力和共同福祉为核心的更广阔的道德框架。借鉴这一框架，我们开发了一种“断裂测试”，将制度基线与系统评估联系起来。我们区分了“证据约束型部署”（将主张严格限定在实际经过评估的范围之内）与“度量约束型治理”（记录下有利证据也无法逾越的约束条件）。在这些限制之内，RISE AI 提供了一个架构，用于就责任（Responsibility）、包容（Inclusivity）、安全（Safety）和赋能（Empowerment）提出有边界且基于证据的主张。负责任的人工智能需要更完善的工程技术、制度修复，以及持续的道德与政治判断。