Governance des Unbekannten

Governing the Unknown — Professional Regulation, AI Governance, and the Limits of Analogy How does a society regulate expertise it cannot fully comprehend? This question predates any debate about artificial intelligence. For over a century, professions such as medicine, aviation, nuclear engineering, and financial auditing have operated under governance arrangements designed to control decisions whose technical basis remains opaque to outsiders. When Anthropic published a constitution for Claude in January 2026 that codified judgment rather than obedience, this old question resurfaced — in a sharper form. If an AI system is expected to exercise something functionally equivalent to professional judgment, the governance problem is no longer hypothetical. It becomes structural. This essay approaches AI governance through two analytical lenses. The first draws on the sociology of professions, examining governance mechanisms across thirteen professions in five legal traditions (common law, civil law, Nordic, East Asian, hybrid). The comparative analysis identifies ten recurring governance mechanisms and five structural failure patterns — a repertoire that proves remarkably consistent across jurisdictions. The second lens is a four-stage model of cognitive architectures, distinguishing rule-based systems (Stage 1), text predictors such as large language models (Stage 2), systems with world models (Stage 3), and potentially conscious or sentient systems (Stage 4). The intersection of both frameworks constitutes the analytical structure of the essay. Three findings emerge. First, AI governance is not a novel problem but a sharpened variant of the century-old challenge of professional regulation. The mechanisms societies have developed — licensing, peer review, independent oversight, confidential incident reporting — remain structurally relevant. Second, these mechanisms encounter breakpoints that shift qualitatively across stages. Two structural breakpoints — the absence of intrinsic motivation and the absence of personal consequences — distinguish AI systems from human professionals at every stage. Paradoxically, these breakpoints appear to close at Stage 4 (conscious systems), yet the governance problem does not resolve; it transforms. Third, consciousness agnosticism — the position that consciousness status may remain permanently undeterminable — functions not as a hindrance to governance but as a design parameter. A governance architecture that operates under irreducible ignorance about the governed entity's inner states requires different structural principles than one that can assume shared phenomenology. The essay is a structural analysis. It identifies architectural principles — among them the separation principle (no actor should simultaneously promote and regulate the same activity), the governance trilemma (autonomy, alignment, and accountability cannot be simultaneously optimized), and the necessity of a layered governance system operating under conditions of ignorance. It does not offer policy recommendations or implementation frameworks. It ends where institutional construction begins.