The Grammar of Diffusion: How Policy Formality Shapes the Spread of AI Governance Across U.S. States

• In Event: Governance of Science, Technology, and AI

Abstract

Policy diffusion, the process through which innovations spread across jurisdictions, has long been a central focus of policy studies. While extensive research has explored the economic, political, and social determinants of diffusion, comparatively little attention has been given to how the internal structure of policies themselves influences their adoption and spread. The rapid emergence of artificial intelligence (AI) technologies provides a timely and important context for revisiting this question. In the United States, where federal guidance on AI governance remains limited and fragmented, individual states have begun formulating their own AI policies, strategies, and oversight mechanisms. This decentralized policy landscape offers a unique opportunity to study how the design of policy content shapes diffusion dynamics.

This study integrates institutional grammar theory into the analysis of AI policy diffusion. Using the ADICO framework—attributes, deontics, aims, conditions, and or else—institutional grammar identifies three levels of policy formality: AIC (shared strategies with non-prescriptive advice for action), ADIC (norms with prescriptive statements), and ADICO (full rules backed by enforcement mechanisms). Building on this structure, the study examines three central questions: (1) whether higher levels of policy formality are associated with greater effectiveness and efficiency in diffusion; (2) whether the nature of prescriptions, encouraged versus obligatory, affects diffusion outcomes; and (3) whether specifying sanctions or consequences strengthens diffusion effectiveness.

To test these questions, the study constructs a comprehensive dataset of state-level AI policies across all 50 U.S. states. Policies are identified through systematic web searches of government websites, legal databases, and news sources, supplemented by ChatGPT-assisted information extraction to ensure coverage of both legislative actions and executive initiatives. The dataset includes legislative bills, executive orders, strategic plans, and official task force reports related to AI governance, spanning the period from 2010 to 2024. Each policy document is processed using computational text analysis tools. Customized dictionaries and machine-learning classifiers are developed to automate the coding of institutional grammar components, with manual validation performed on a subset of the data to ensure reliability and accuracy. Event history analysis is then used to assess how internal policy structure relates to patterns of diffusion, controlling geographic proximity, political ideology, and economic capacity.

By combining institutional grammar theory with policy diffusion analysis, this research advances understanding of how the micro-foundations of policy design influence broader diffusion processes. The findings offer both theoretical contributions to diffusion scholarship and practical insights for crafting more effective and transferable AI governance frameworks.

Authors

• Yiying Chen, Syracuse University
  Presenting Author

• Tong Chen, Rutgers University
  Non-Presenting Co-Author