Active Inference: A method for Phenotyping Agency in AI systems?

Technical Summary: Active Inference: A Method for Phenotyping Agency in AI Systems

1. The Problem: The Conceptual Gap in AI Agency

As AI shifts from passive predictive models (Generative AI) toward "agentic workflows" and Artificial General Intelligence (AGI), the field lacks a rigorous, measurable framework to characterize agency. Current definitions primarily rely on "autonomy" and "goal-directedness," which the authors argue are insufficient. These linear definitions fail to capture the multi-dimensional nature of agency (e.g., a robot may be autonomous in navigation but not in goal-setting) and do not provide a way to distinguish between different "levels" or "types" of agentic behavior in a way that is useful for both technical development and societal governance.

2. Methodology: Active Inference as a Framework

The authors propose using Active Inference—a framework derived from theoretical neurobiology—to define and measure agency. They move away from "weak" definitions of agency toward a tripartite philosophical definition:

1. Intentionality: Actions grounded in internal mental states (beliefs and desires).
2. Rationality: Actions that are normatively coherent and follow logically from a world model.
3. Explainability: Actions that are causally traceable to those internal states.

Computational Implementation:
The authors instantiate these criteria within a Partially Observable Markov Decision Process (POMDP) under a variational framework. In this model:

• Beliefs are represented by posterior distributions over hidden states.
• Desires are represented by prior preferences over observations.
• Action Selection is achieved by minimizing Expected Free Energy (EFE), which naturally balances exploitation (satisfying preferences) and exploration (reducing uncertainty).

3. Key Contributions: Computational Phenotyping and Empowerment

The paper introduces the concept of "computational phenotyping" for AI. Just as psychiatry uses models to identify behavioral symptoms of disorders, the authors suggest using active inference to identify "phenotypes" of agency in AI.

To operationalize this, they utilize Empowerment as a metric. Empowerment is defined as the channel capacity (maximum mutual information) between an agent's actions and its future observations. It measures the degree of control an agent has over its environment. The authors distinguish between:

• Objective Empowerment: Control possible in the environment.
• Subjective Empowerment: Control the agent believes it has based on its model.
• Potential vs. Actual Empowerment: The difference between the capacity to act and the capacity remaining after preferences constrain choices.

4. Results: The T-Maze Experiment

The authors validate their framework using a minimal T-maze paradigm. An agent must choose between a "Cue" (which provides information) and "Left/Right" (which are terminal "traps").

• Zero-Agency Phenotype (0 bits): Occurs when the agent is in a "trap" state. All actions lead to the same observation, meaning the agent has no control over its future.
• Intermediate-Agency Phenotype (~1 bit): Occurs at the start of the task. The agent has some influence, but environmental ambiguity (not knowing where the cheese is) limits its effective control.
• High-Agency Phenotype (~1.585 bits): Occurs after the agent performs an epistemic action (visiting the Cue). By resolving uncertainty, the agent increases its empowerment, transitioning from a redundant action space to a fully differentiated one where it can rationally align actions with preferences.

Key Finding: Exploratory actions (driven by the need to reduce uncertainty) directly increase an agent's empowerment. Knowledge is not just a byproduct; it is a prerequisite for agency.

5. Significance and Implications

The paper provides a bridge between technical AI development and AI Governance:

• For AI Development: It provides a principled way to measure how "agentic" a system is by looking at its world model and its ability to increase its own control through information gathering.
• For AI Governance: The authors argue that governance strategies must evolve alongside the agent's phenotype.
  • Low-agency agents can be controlled via external constraints (e.g., physical barriers or hard-coded rules).
  • High-agency agents cannot be controlled by external constraints alone; they require internalist governance (e.g., modulating their prior preferences or social/normative alignment), as they possess the reasoning capacity to navigate around simple external restrictions.