This paper introduces a new way to look at the rapidly evolving landscape of AI policy, moving beyond just listing principles to asking a fundamental question: What can a person actually do when an AI system affects their life?
TL;DR
The authors propose the Society-Centered AI Governance Framework, which evaluates AI policy based on whether it provides citizens with notice, the ability to challenge decisions, and a path to correction. Through a deep dive into North Carolina’s AI roadmap, they reveal a trend of "visibility-first governance"—where the state tracks its own AI use but leaves affected individuals with no enforceable rights.
What Problem Does This Paper Address?
AI governance research typically focuses on the outputs of policy: ethics principles, strategies, and audit reports. However, there is a lack of shared language to understand AI policy while it is still being written.
Policymakers face an "evidence dilemma": acting too early might entrench ineffective rules, but waiting for harm to occur leaves society vulnerable. The authors argue that current policies often create "procedural visibility" (the state knows what AI it is using) without "public contestability" (the public can do something about it). This gap creates a risk of "legitimacy before protection," where transparency tools like public inventories build trust before any actual legal protections are in place.
Methods
The authors developed a framework anchored in five theoretical components, drawing on the work of Amartya Sen, Deborah Stone, Charles Lindblom, and Herbert Simon.
The framework components are summarized in Table 1:
To evaluate this framework, the team used a mixed-method approach:
- Provision-Level Reading: They analyzed the legal force and binding status of policy instruments.
- Model-Assisted Coding: They used GPT-5.4 and GPT-4o-mini to code a large corpus of documents, followed by rigorous human validation (achieving a 97% match).
- Semi-Structured Interviews: They interviewed nine members of North Carolina’s AI Leadership Council to understand the reasoning and constraints (like "bounded rationality") behind policy decisions.
Datasets
The study utilized an AI Governance Corpus consisting of 49 documents published between 2020 and 2026:
- North Carolina (31 documents): A mix of executive orders, legislative bills, judicial orders, and the state's 2026 AI Roadmap.
- Comparative State Corpus (18 documents): Reports from task forces and committees across 18 other U.S. states (e.g., California, Texas, Washington) to identify recurring patterns in tool vocabularies and oversight structures.
Key Insights and Ideas
- Niti vs. Nyaya: Borrowing from Indian legal philosophy, the authors distinguish between Niti (formal rules and institutions) and Nyaya (justice as it is actually realized in people's lives). Most current AI governance is heavy on Niti but light on Nyaya.
- Agency Duty vs. Affected-Person Right: A critical distinction. A policy might mandate that an agency should provide notice (an internal duty), but if that doesn't grant the citizen a legal right to receive notice or sue if they don't, the protection is hollow.
- Visibility-First Governance: In North Carolina, the state has successfully made its AI systems visible to itself through inventories and oversight teams. However, AI-specific rights to challenge or correct decisions remain advisory or scheduled for years in the future.
- The "Standing Gap": The authors found that while policies often mention "the public," the actual drafting bodies and testimony lists are dominated by officials and industry experts, often excluding the very communities most likely to be harmed.
Key Results
- Recourse is uneven: In North Carolina, enforcement is strongest for harms with an "intentional offender" (like AI-generated deepfake imagery of minors, which is a felony). It is weakest for "diffuse harms" caused by the state's own systems, such as AI used in benefits eligibility determinations.
- Documented Gaps: The North Carolina Roadmap documents missing protections but assigns them future dates (2026-2027). Until then, notice remains an agency duty without an enforceable right for the person affected.
- Model Efficiency: The model-assisted coding workflow successfully mapped 8,589 code instances, demonstrating that LLMs can be effective "research assistants" for large-scale qualitative policy analysis when paired with human oversight.
Limitations and Future Work
The authors acknowledge that their findings are primarily focused on North Carolina. The "visibility-first" pattern might be weaker in jurisdictions where the legislature acts before the executive branch (e.g., the European Union or Colorado).
Furthermore, the study did not directly interview residents affected by these AI systems. Future work should focus on measuring "lived outcomes"—asking if these inventories and feedback channels actually make AI more contestable and correctable for the average citizen. The authors also suggest that researchers should use their "four public questions" to test new AI bills and international agreements at the drafting stage.
Summary updated