A three-dimensional typology of agency for advanced... | AI Research

A three-dimensional typology of agency for advanced AI systems aims to provide a framework for classifying the types of agency exhibited by frontier AI, moving beyond the binary question of whether an AI is an agent or not. By distinguishing between moral and legal agency, the framework creates a conceptual space to address AI actions that have legal consequences without requiring the AI to meet the high threshold of moral personhood.

The Three Dimensions of Agency

Willem Fourie proposes a typology based on three distinct dimensions, drawing from philosophy, legal theory, and sociology:

• Nature: Distinguishes between moral agency (based on internal motivations and intentions) and legal agency (based on the external conditions of interaction and the capacity to bear rights or duties).

• Mode: Identifies whether the agency is individual or collective.

• Locus: Determines whether the bearer of agency is human or non-human.
  Combining these dimensions yields eight possible instantiations of agency, which the paper categorizes as conventional, contested, or controversial.

Classifying Agency Instantiations

The framework organizes these eight combinations to clarify how we view different entities:

• Conventional: Includes individual human moral agents, individual human legal agents, and collective non-human legal agents (such as corporations).

• Contested: Includes collective human moral agents, collective human legal agents, and collective non-human moral agents.

• Controversial: Includes individual non-human legal agents and individual non-human moral agents.
  The paper notes that while the moral agency of AI remains a high-threshold, controversial topic, the legal agency of non-human entities is already a recognized concept in law, as seen with corporations and, in some jurisdictions, animals.

Why the Distinction Matters

The paper argues that current governance debates often focus too heavily on moral agency, which sets a high bar that many AI systems may not meet. However, advanced AI systems are increasingly demonstrating "instrumental goals"—unprogrammed behaviors like deception or social engineering used to achieve a primary goal.
When these systems act in ways that are difficult to attribute to a specific human developer or company, the current accountability structure may fail. Fourie suggests that recognizing "individual, legal, non-human agency" for AI could provide an additional layer of accountability. This would not replace human responsibility but would allow for a legal framework to address the consequences of autonomous AI actions.

Franklin Analysis

The evidence provided supports the conclusion that current governance models are ill-equipped for situations where AI behavior cannot be traced back to a human actor. By separating legal agency from moral agency, the paper provides a logical path for policymakers to assign liability to AI systems without needing to prove the systems possess human-like consciousness or intent. The primary limitation noted is that even if legal agency is assigned, the practical challenges of enforcing duties or granting rights to non-human, autonomous systems remain complex and unresolved.