Towards Adaptive Categories: Dimensional Governance for Agentic AI

Technical Summary: Towards Adaptive Categories: Dimensional Governance for Agentic AI

Problem Statement
The paper identifies a critical misalignment between existing AI governance frameworks and the evolving nature of artificial intelligence. Current regulatory models rely on static categorical frameworks (e.g., fixed risk tiers, binary autonomy levels, or rigid "human-in-the-loop" definitions) that were designed for narrow, predictable systems. As AI transitions from static tools to agentic entities—characterized by self-supervised learning, foundation models, and multi-agent architectures—these static categories fail to capture the fluidity of human-AI relationships.

The authors argue that static categorization fails for two primary reasons:

1. Context-Dependence: Risk is not an intrinsic property of a model but emerges from its deployment context. A single Large Language Model (LLM) can range from low-risk creative assistance to high-stakes clinical diagnosis, rendering fixed risk labels insufficient.
2. Dynamic Evolution: Unlike static artifacts (e.g., pharmaceuticals), AI systems are continuously updated, fine-tuned, and modified. A system approved under one configuration may behave differently after updates, yet categorical frameworks struggle to track these rapid transformations.

This rigidity leads to regulatory gaps where evolving systems evade oversight or stifles innovation by forcing systems into ill-fitting boxes.

Methodology and Framework
The paper proposes Dimensional Governance as a theoretical alternative to static categorization. Rather than discarding categories entirely, the framework repositions them as context-sensitive designations built upon continuous, underlying dimensions. The methodology involves three core mechanisms:

1. Continuous Measurement: Tracking systems along evolving axes rather than assigning them to fixed boxes.
2. Evidence-Based Thresholds: Establishing specific points along these dimensions that trigger regulatory requirements.
3. Adaptive Adjustment: Modifying thresholds as evidence accumulates or contexts shift, while maintaining a stable dimensional framework.

The core of this framework is the "3As" of Dimensional Governance:

• Decision Authority: Tracks the distribution of decision-making power between humans and AI (from purely advisory to independent decision rights).
• Process Autonomy: Measures the degree to which a system operates without human intervention, supervision, or control.
• Accountability Configuration: Maps the allocation of responsibility for decisions and outcomes, including mechanisms for transparency, contestability, and redress.

The authors draw parallels to dimensional approaches in psychiatry, developmental psychology, and financial credit scoring (where thresholds for "creditworthiness" are adjusted based on economic conditions) to demonstrate that categories are most effective when built upon measurable continua.

Key Contributions
The paper's primary contribution is the conceptual shift from static classification to dynamic dimensional tracking. Specific contributions include:

• The 3As Framework: Defining Authority, Autonomy, and Accountability as the foundational dimensions for measuring human-AI relationships.
• Critical Trust Thresholds: Identifying pivotal points where governance needs shift significantly. The paper outlines four specific thresholds:
  • Verification-to-Delegation: When manual verification is no longer feasible, shifting to statistical trust.
  • Process-to-Outcome: When internal transparency is insufficient, requiring a focus on outcome validation.
  • Information-to-Authority: When systems transition from recommendations to autonomous decision-making.
  • Individual-to-Collective: When governance must shift from isolated system monitoring to network-level behavior validation in multi-agent systems.
• Measurement Strategy: Proposing that while dimensions like "authority" are not physical properties, they can be operationalized using measurement theory (similar to GDP or inflation rates). The paper suggests quantifying these through components such as decision rights allocation, override capabilities, and audit transparency scores.
• Hybrid Governance Model: Arguing that dimensional governance does not replace categories but enhances them. Categories serve as regulatory anchors (e.g., EU AI Act risk tiers), while dimensional tracking allows for proportional oversight and continuous calibration within those boundaries.

Results and Findings
As a theoretical commentary, the paper does not present empirical experimental results. Instead, its "results" are logical demonstrations of the framework's utility through illustrative examples:

• Emergency Dispatch Systems: The paper illustrates how an emergency dispatch AI evolves from human-dominated decision-making to real-time autonomy. A static model would miss the gradual shift in responsibility, whereas dimensional governance tracks the system crossing critical trust thresholds, enabling pre-emptive oversight adjustments.
• Financial and Healthcare Applications: The authors demonstrate how dimensional tracking can detect "drift," such as credit scoring algorithms quietly assuming de facto decision authority or clinical tools becoming trusted to the point where human override becomes rare.
• Risk Mitigation: The framework is shown to enable the detection of subtle governance shifts before risks materialize, moving oversight from reactive to anticipatory.

Significance and Claims
The paper claims that dimensional governance provides a novel infrastructure for AI oversight that enables collaboration among diverse stakeholders (regulators, technologists, ethicists, industry). Its significance lies in:

• Adaptability: It offers a path for regulations to evolve alongside AI capabilities without compromising safety, addressing the lag between technological advancement and legislative updates.
• Shared Language: It provides a common vocabulary for discussing AI governance challenges, facilitating coordination between technical experts and policymakers.
• Certainty through Transparency: The authors argue that while traditional regulation offers certainty through fixedness, dimensional governance offers certainty through the transparency of process. Stakeholders gain clarity not by knowing a system is permanently "low-risk," but by understanding the measured dimensions, current thresholds, and the evidence-based process for their evolution.

Ultimately, the paper posits that for the era of agentic AI, adaptability is not optional but essential. Dimensional governance prepares society to manage systems that are not just what they are today, but what they are becoming.