Policy Framing and Participation in Complex Technology Rulemaking: Evidence from AI Regulation

• In Event: Participation and Procedural Justice in AI Governance

Abstract

How can regulatory processes meaningfully incorporate diverse societal perspectives when the policy problems they address are highly technical? In the context of artificial intelligence (AI), governments increasingly rely on notice-and-comment rulemaking to gather input from affected stakeholders. Yet participation in these processes is uneven. Existing research has emphasized differences in stakeholder resources, expertise, and organizational capacity, or examined whether public comments influence final rules. Much less attention has been paid to an earlier stage of the process: how the framing of proposed rules may shape who participates, what issues they respond to, and how inclusive participation becomes.

This study examines that earlier stage by analyzing how policy framing structures participation in AI-related federal rulemaking. It addresses three questions: (1) how do different stakeholder groups vary in the issues they emphasize in AI rulemaking comments? (2) which features of policy framing in proposed rules are associated with variation in participation intensity, semantic breadth, and inclusiveness, and how do these relationships differ across stakeholder groups? and (3) how are these relationships between policy framing and participation manifested across different AI policy issues?

The study draws on research in AI governance, science and technology policy, and public participation in administrative rulemaking to examine how policy texts structure participation. It conceptualizes proposed rules as shaping the conditions under which participation occurs. In complex technological domains, policy framing influences how issues are interpreted, what counts as relevant knowledge, and who is able to engage. While prior research documents inequalities in rulemaking participation, it has not examined how these patterns may be associated with the design of policy texts themselves.

The analysis uses AI-related federal rulemaking dockets collected from Regulations.gov. From an set of 64 dockets, a final analytical sample of 37 documents is constructed. Using sentence-level semantic embeddings, policy texts are segmented and represented in a high-dimensional semantic space. A k-nearest-neighbor graph is constructed over sentence embeddings, and Louvain community detection is applied to identify clusters of semantically related sentences, resulting in 303 policy topics. Public comments (N = 13,137) are similarly processed and linked to these topics based on cosine similarity (threshold ≥ 0.5).

The analysis proceeds in three steps. First, the study constructs topic distributions for each stakeholder type and compares issue emphasis across industry, civil society, academia, government, international organizations, and individual commenters. Second, it measures narrative and semantic features of policy texts, including complexity, readability, specificity, and technical versus public-value orientation, and estimates their relationships with participation intensity, semantic breadth, and inclusiveness using multilevel regression models. Third, it visualizes policy topics in a reduced-dimensional embedding space and maps variation in framing and participation across topics.

The findings contribute to research on AI governance, policy framing, and administrative rulemaking by shifting attention from the influence of public comments on policy outcomes to the role of policy design in structuring participation. By linking framing features to participation patterns and situating these relationships across AI policy issues, the study provides new evidence on how policy design shapes engagement in complex technology rulemaking.

Author

• Jin Ai, Rutgers University - Newark
  Presenting Author