Variety Deficits in AI Governance: A Structural Analysis

Introduction

The governance of artificial intelligence presents a distinctive regulatory challenge. AI systems are characterized by rapid capability growth, broad applicability across domains, and emergent behaviors that are difficult to anticipate in advance. These characteristics generate high disturbance variety — the range of outcomes that governance frameworks must be able to absorb.

Ashby's Law of Requisite Variety provides a precise criterion for evaluating regulatory adequacy: a regulator can reduce the variety of outcomes in a system only to the extent that it possesses at least as much variety as the disturbances it must absorb. Applied to AI governance, this criterion implies that regulatory frameworks must be able to model and respond to the full range of AI capabilities and applications.

This paper develops a formal methodology for measuring regulatory variety and applies it to six major AI governance frameworks: the EU AI Act, the US Executive Order on AI, the UK AI Safety Institute framework, the OECD AI Principles, the G7 Hiroshima AI Process, and the UN Advisory Body on AI recommendations.

Measurement Methodology

We measure regulatory variety across five dimensions: technical modeling capacity (the ability to model AI capabilities and behaviors), jurisdictional coverage (the range of actors and applications covered), enforcement authority (the ability to impose binding requirements), institutional independence (freedom from capture), and adaptive capacity (the ability to update in response to new information).

For each dimension, we construct a variety score ranging from 0 (no variety) to 1 (full variety relative to current AI capabilities). We then aggregate across dimensions using a geometric mean, which captures the multiplicative nature of regulatory capacity — a framework with high variety in four dimensions but zero variety in one dimension has zero effective variety.

We assess disturbance variety using a parallel methodology, measuring the range of AI capabilities and applications across the same five dimensions. The variety ratio for each framework is the ratio of regulatory variety to disturbance variety.

"The variety of the regulator must be at least as great as the variety of the disturbances it must absorb." — W. Ross Ashby

Findings

All six frameworks exhibit variety deficits across all five dimensions. The deficits are largest in the dimensions of technical modeling capacity and adaptive capacity — the two dimensions most directly relevant to governing rapidly evolving AI capabilities.

The EU AI Act achieves the highest overall variety ratio (0.41) among the frameworks examined, primarily due to its relatively strong enforcement authority and jurisdictional coverage. However, it falls well below the threshold of 1.0 required for effective governance of frontier AI systems.

The US Executive Order on AI achieves a variety ratio of 0.31, with particular weaknesses in institutional independence and adaptive capacity. The UK AI Safety Institute framework achieves a ratio of 0.28, with strengths in technical modeling capacity but weaknesses in enforcement authority and jurisdictional coverage.

Implications for Governance Reform

The findings suggest three priorities for governance reform. First, all major frameworks need significant investment in technical modeling capacity — the ability to model AI capabilities and behaviors in advance of their deployment. Second, adaptive capacity mechanisms — formal processes for updating regulatory requirements in response to new information — must be built into governance frameworks from the outset. Third, international coordination is needed to address the jurisdictional coverage deficits that affect all existing frameworks.

The Good Regulator Theorem provides a precise criterion for evaluating reform proposals: does the proposed reform increase the variety ratio of the regulatory framework? Reforms that increase enforcement authority without increasing technical modeling capacity will not produce effective governance of frontier AI systems.