Socially Grounded Agentic AI: Coordinating Plural P... | AI Research

Socially Grounded Agentic AI: Coordinating Plural Perspectives through Social Theory proposes a framework for AI alignment that treats the representation of multiple viewpoints as a problem of social coordination rather than simple output diversification. The authors, Matt Ratto, Abhishek Moturu, and Daniel Silver, argue that current pluralistic alignment methods—which often focus on listing diverse opinions—fail to account for the social structures that create and organize those perspectives. By applying sociological theories from George Herbert Mead, Jürgen Habermas, and Pierre Bourdieu, the paper outlines how agentic AI can be designed to represent roles, facilitate deliberation, and account for power dynamics within social fields.

Modeling Perspectives as Social Roles

The authors suggest that instead of asking models to generate generic "diverse" viewpoints, systems should be designed to represent specific social roles. Drawing on Mead’s concept of the "generalized other," the paper argues that perspectives are structured by institutional obligations, constraints, and expertise. In this framework, a role is not a persona, but a situated position within a field of action. Design implementations include role-indexed prompts and retrieval-augmented generation (RAG) systems that condition model outputs on the specific institutional logic of a role, such as a clinician or patient advocate, rather than treating all viewpoints as anonymous alternatives.

Deliberation Through Interaction

To address the limitations of steerable AI, the paper proposes moving from direct control to structured interaction. Using Habermasian theory, the authors frame alignment as a deliberative process where agents with differentiated roles exchange claims, provide evidence, and undergo critique. This approach utilizes multi-agent architectures where interaction protocols—such as claim-evidence-rebuttal formats—govern how the system reaches a conclusion. By making the reasoning process inspectable through interaction traces, the authors argue that systems can achieve more legitimate and transparent outcomes when handling conflicting values.

Field-Aware Population Representation

The paper addresses the challenges of distributional pluralism by applying Bourdieu’s theory of social fields. The authors note that simply reflecting the distribution of beliefs in a population can reinforce existing biases or structural inequalities. To mitigate this, they propose "field-aware" alignment, which models populations as configurations of positions with varying levels of authority and expertise. Implementation strategies include population-aware routing, where inputs are directed to agents representing specific social positions, and position-weighted aggregation, which adjusts contributions based on normative considerations like equity and expertise rather than just empirical prevalence.

Evaluating Trajectories Over Outputs

The authors argue that pluralistic alignment should be evaluated at the level of interaction trajectories rather than isolated model outputs. Because social legitimacy is an emergent feature of ongoing exchanges, the paper suggests that evaluation frameworks must account for how perspectives are sustained, revised, and coordinated over time. This requires auditing the full reasoning process, including intermediate steps, tool usage, and the conditions under which a system defers or escalates. The authors identify this as a necessary step for ensuring that agentic systems remain accountable and responsive to the social contexts in which they operate.