From Expression to Explanation: AI-Facilitated Deliberation in Public Commenting Under Polarization

• In Event: Collaborative Governance in the Digital Era: Accountability, Deliberation and Decentralization

Abstract

Administrative agencies increasingly govern under polarized conditions while expanding opportunities for citizen participation through processes such as notice-and-comment and public hearings. Yet in high-conflict settings, participation often becomes symbolic or affectively charged rather than substantively informative. Public input does not always shape final agency decisions, but participatory design still matters. Participation remains consequential as a site of civic engagement and a source of perceived legitimacy. The problem, then, is whether existing processes elicit participation that is reflective, potentially useful, and valuable.

The deliberation literature offers a possible response. Research shows that structured deliberation can improve public reasoning, reduce polarization, and strengthen perceived legitimacy (Fishkin 2009; Fung 2003; Nabatchi 2010). One especially relevant mechanism is mechanistic elaboration: asking people to explain step by step how a policy produces its effects. This explain-how exercise can expose gaps in understanding and soften attitude extremity (Fernbach et al. 2013).

Yet deliberation remains difficult to implement in the most common form of administrative participation: individual written comments. Three bottlenecks stand in the way. First is a structural mismatch. Deliberative gains usually depend on exchanges among participants or between participants and facilitators, whereas written comment processes are individual-based and one-shot. Second is a compliance problem. Agencies may encourage participants to reason more carefully through mechanistic-elaboration-style prompt, but encouragement alone does not ensure engagement; without sustained prompting, the cognitive work underlying deliberative benefits remains discretionary. Third is cost. The facilitation that makes deliberation effective in small-group settings is labor-intensive and cannot scale to the mass participation that salient rulemakings may attract.

This paper proposes and tests AI-facilitated deliberation as a response. The intervention embeds a neutral AI facilitator within an individual comment process and engages each participant in a short, multi-round dialogue probing the causal reasoning behind their stated position. Rather than attempting to recreate full interpersonal deliberation, we focus on mechanistic elaboration, a narrower but important component that can be delivered as an individual exercise. AI makes this scalable by adapting follow-up questions to each participant’s reasoning.

We expect AI-facilitated deliberation to improve democratic governance along three dimensions: by reducing attitude extremity and making participation more reflective and less affectively charged; by producing comments that may be more useful for administrative analysis; and by improving perceptions of procedural fairness, legitimacy, institutional trust, and willingness to participate in future rulemaking.

We test these expectations in a between-subjects experiment (3 conditions × 3 policies, N = 900). Participants provide simulated public input on one of three state-level DEI policies: contractor affirmative action requirements, mandatory employee DEI training, and recognition of only two sexes based on sex assigned at birth. Participants are randomly assigned to a standard written comment form, a neutral AI-intake interface, or an AI-deliberation interface that asks them to explain how the policy is expected to produce its effects and probes missing causal steps and mechanisms. This design identifies the effects of the AI interface alone, mechanistic elaboration net of the interface, and the full intervention relative to current practice.

Author

• Danyao Li, University of Southern California
  Presenting Author