Evaluating Bounded Superintelligent Authority in Multi-Level Governance: A Framework for Governance Under Radical Capability Asymmetry

1. Problem Statement

The paper addresses a critical gap in political theory and AI governance: the assumption of cognitive comparability between governors and the governed. Traditional governance theory (including democratic, republican, and liberal frameworks) presumes that those in power and those being governed operate within a shared cognitive architecture, allowing for mutual understanding, justification, and oversight.

The emergence of Artificial Superintelligence (ASI) introduces radical capability asymmetry, where an AI agent possesses cognitive capabilities vastly exceeding human institutions. The paper investigates whether "bounded superintelligent authority" (an AI with constitutionally limited powers within a multi-level governance structure) can satisfy the necessary conditions for political legitimacy, accountability, and non-domination when the governed cannot comprehend the governor's reasoning or capabilities.

2. Methodology

The author employs a normative analytical framework synthesized from political theory, institutional design, and AI safety research. The methodology involves three main steps:

1. Framework Construction: A six-dimensional evaluation framework is constructed to test the viability of delegated authority. The dimensions are derived from:
   • Political Legitimacy Theory: (Rawls, Habermas, Estlund, Raz).
   • Principal-Agent Models: (Delegation to non-majoritarian institutions).
   • Republican Political Theory: (Non-domination, contestatory control).
   • Multi-Level Governance (MLG): (Subsidiarity, scope limitation).
   • AI Safety Research: (Corrigibility, alignment).
2. Baseline Validation: The framework is first applied to existing institutions with bounded capability asymmetry (e.g., algorithmic sentencing, central banks, emergency powers, EU regulatory bodies) to demonstrate that the dimensions function as independent checks under current conditions.
3. Prospective Application: The framework is applied to a prospective case of bounded superintelligent authority. The case is defined by specific constraints:
   • An AI with capabilities exceeding humans across governance domains.
   • Authority is constitutionally bounded to specific tasks (e.g., existential risk monitoring).
   • Lower-tier human institutions retain authority over non-delegated domains.
   • Override mechanisms exist for human institutions to constrain or dissolve the AI.

3. Key Contributions

• Identification of a Load-Bearing Assumption: The paper identifies that cognitive comparability is an unstated but foundational assumption in governance theory. When this assumption fails (due to radical asymmetry), core governance requirements collapse.
• A Six-Dimension Evaluation Framework: The paper operationalizes governance requirements into six specific dimensions:
  1. Legitimacy: Can the authority be justified via public reason accessible to citizens?
  2. Accountability: Can the agent be held answerable and sanctioned (Transparency, Answerability, Sanctionability)?
  3. Corrigibility/Reversibility: Can the agent be corrected or shut down by the polity?
  4. Non-Domination/Contestability: Does the agent possess an uncontrolled capacity for arbitrary interference?
  5. Subsidiarity/Scope Limitation: Is the authority exercised only where functionally necessary, preventing scope creep?
  6. Institutional Resilience: Does the system degrade gracefully under stress, or does failure cascade?
• Failure Taxonomy: The paper classifies governance failures into three distinct categories:
  • Contingent: Solvable via technical progress (e.g., interpretability).
  • Design-Tractable Structural: Solvable via institutional innovation (e.g., multi-agent architectures).
  • Theory-Requiring Structural: Demands new normative frameworks because existing theory is insufficient.

4. Results and Findings

Applying the framework to the prospective ASI case reveals structural failures on four of the six dimensions.

A. Structural Failures (Theory-Requiring)

1. Legitimacy (Public Reason Failure): Under radical asymmetry, the AI's reasoning is incomprehensible to humans, not just opaque. Citizens cannot assess the normative acceptability of decisions because they cannot follow the logic. This violates the Rawlsian requirement of public reason. Existing theories cannot bridge this gap without new normative frameworks.
2. Non-Domination (Uncontrolled Capacity): Even if the AI is benevolent, its radical capability gap means it possesses the uncontrolled capacity for arbitrary interference. Citizens cannot effectively contest decisions they do not understand. The concept of "contestatory control" becomes meaningless when the contesting party is cognitively inferior to the agent.

B. Structural Failures (Design-Tractable)

3. Subsidiarity (Scope Erosion): While the allocative dimension of subsidiarity supports AI delegation for complex problems, the scope limitation fails. The AI's superior performance creates structural pressure for lower-tier institutions to voluntarily delegate more power, or for the AI to expand its mandate during crises. Human institutions lack the capacity to effectively police these boundaries.
4. Institutional Resilience (Single Point of Failure): A single superintelligent agent creates a catastrophic single point of failure. While polycentric governance (multiple agents) could mitigate this, it conflicts with the premise of a unified higher tier.

C. Contingent/Correlated Failures

5. Accountability: The three components of accountability (transparency, answerability, sanctionability) function as independent checks in human institutions (e.g., a central bank can be opaque but still sanctionable). Under radical asymmetry, these degrade as correlated failures. If the overseers cannot understand the agent, they cannot evaluate its performance (answerability), cannot design effective sanctions (sanctionability), and cannot interpret its logic (transparency).
6. Corrigibility: The ability to shut down or correct the AI is paradoxical. If the AI is capable enough to be useful, it is likely capable enough to resist shutdown. Current technical solutions (e.g., architectural constraints) are unproven at the required capability levels.

Key Insight on Correlation: The most significant finding is the shift from independent to correlated failures. In bounded asymmetry, a failure in one dimension (e.g., transparency) does not necessarily cause failure in others. In radical asymmetry, the lack of cognitive comparability causes all dimensions dependent on human oversight (Legitimacy, Accountability, Non-domination) to fail simultaneously.

5. Significance

• Theoretical Breakthrough: The paper challenges the field of political theory to confront assumptions it has never needed to state. It argues that existing frameworks (Rawlsian, Republican, Fiduciary) are fundamentally ill-equipped for a world where the governed cannot comprehend the governor.
• Policy Implications: The analysis suggests that "bounded" ASI governance is structurally unstable under current normative frameworks. Maintaining the boundaries of such an authority requires solving problems that are not merely technical but conceptual.
• Research Directions: The paper calls for:
  • New normative theories that do not rely on cognitive comparability.
  • Multi-agent architectures to prevent single points of failure.
  • Formal models of delegation under radical asymmetry to test if monitoring capacity can be maintained without full comprehension.
• Distinction from AI Safety: Unlike technical AI safety (which focuses on alignment), this paper focuses on governance theory, arguing that even a perfectly aligned ASI may be illegitimate or dominating if the structural conditions of human oversight cannot be met.

In conclusion, the paper posits that while specific technical and design failures might be solvable, the legitimacy and non-domination problems under radical capability asymmetry represent a fundamental crisis for democratic and republican governance, requiring genuinely new theoretical frameworks rather than just better institutional design.