Between Niti and Nyaya: A Society-Centered Framework for AI Governance
Organizations: Society-Centered AI Lab, Department of Computer Science, School of Data and Information Sciences, University of North Carolina at Chapel Hill, USA · Society-Centered AI Lab, School of Data and Information Sciences, University of North Carolina at Chapel Hill, USA
Abstract
Governments are making decisions about AI before they can measure its public effects, and those decisions determine whether a person can learn that AI influenced an outcome, challenge it, and see it corrected. Research has documented what AI governance produces: principles, strategies, registers, and audits. However, we lack shared language for understanding AI policy while it is still being made. We introduce the Society-Centered AI Governance Framework, which gives policymakers, researchers, and the public a way to assess instruments in any jurisdiction by whether a person can get notice, challenge a decision, and see it corrected. It reads a jurisdiction's instruments together by date and binding status, separates an agency duty from a right a person can invoke, and traces which later instrument must enforce each commitment. We evaluated it in North Carolina using interviews with nine AI Leadership Council members, conducted before and after the state's 2026 AI roadmap, and a provision-level reading, with human-reviewed model-assisted coding, of 49 AI governance documents from 19 states. For the state's own AI systems, we find visibility-first governance: visibility duties are in effect, notice exists as an agency duty with no right a person can invoke, and AI-specific challenge and correction remain advisory. The roadmap assigns owners and dates to the missing protections, so the gaps are documented but not yet closed. Recent federal preemption proposals focus on AI development, so recourse for a state's own systems is likely to remain a state decision. We identify "legitimacy before protection" as a testable risk of visibility-first governance and name the milestones that would test it. More broadly, judging AI policy by who can act on what it makes visible lets the field assess recourse while policy is still forming, before harm occurs.
Figures & tables
| Concept | Framework component | Key governance question |
|---|---|---|
| Niti (Sen) | Niti Formalization | What rules, inventories, and accountability mechanisms has the institution built? |
| Nyaya (Sen) | Nyaya Recourse | What can a person learn, challenge, and have corrected? |
| Polis (Stone) | Pluralistic Governance of AI | How do competing interests and actor power shape what gets built and who defines the problem? |
| Muddling through (Lindblom) | Adaptive Governance of AI with Incremental Revision | How does governance move forward as the technology and its consequences change? |
| Bounded rationality (Simon) | Bounded Rationality in AI Governance | What can institutions staff, interpret, enforce, and maintain, given limits of knowledge, capacity, and time? |
| Instrument | Visibility duty | Notice | Challenge | Correction | Later instrument | Status |
|---|---|---|---|---|---|---|
| Framework v1.1 (Aug 2025; ID-6) | Inventories; monitoring, audit, and documentation | Agency “shall provide notice”; disclaims enforceable rights | None | None | Roadmap; no right named | In effect as internal direction |
| Executive Order No. 24 (Sep 2025; ID-1) | Oversight teams; high-risk use cases collected for public review | None | None | None | Roadmap; no right named | In effect |
| Roadmap, agency AI generally (Jul 2026; ID-49) | Recommends public dashboards of high-risk use cases | As Framework | Recommends feedback mechanism (Dec 2026) | Recommends AI Liaisons for escalation (Dec 2027) | Owner and date named; no instrument named | Advisory |
| Roadmap, benefits eligibility (Jul 2026; ID-49) | As above | Recommends disclosure (Mar 2027) | None; general appeal only | Person keeps final determination | Owner named; no instrument named | Advisory |
| Harm | Deepest level specified | Instrument | Status | Responsible party |
|---|---|---|---|---|
| AI-generated sexual imagery of minors; altered private images | Enforcement: felony offenses; civil action and an order destroying the image | Session Law 2024-37 (NC, ID-21) | In effect | Individual offender |
| Deepfake harms to individuals | Enforcement: offense, civil cause of action, and image destruction | House Bill 934 (NC, ID-19) | Proposed | Individual creator or distributor |
| Errors when professionals use AI | Liability allocated: professional solely responsible; developer shielded | House Bill 934 (NC, ID-19) | Proposed | Learned professional |
| AI safety incidents | Tracked and routed: 72-hour report to the attorney general, with enforcement authority | Senate Bill 735 (NC, ID-25) | Proposed | Developer |
| AI in benefits decisions | Human-decision safeguard: no final determinations by AI; no complaint route | Roadmap (NC, ID-49) | Advisory | Deploying agency |
| Public concerns about agency AI use | Routed to staff: public feedback mechanism | Roadmap (NC, ID-49) | Advisory | Deploying agency |
Appendix figures & tables12 assets
Supplementary material from the paper’s appendix.
Appendix
| ID | State | Branch | Type | Document title |
|---|---|---|---|---|
| North Carolina: Executive Branch | ||||
| 1 | NC | Executive | Executive order | Executive Order No. 24: Advancing Trustworthy AI That Benefits All North Carolinians |
| 2 | NC | Executive | Press release | Governor Stein Announces Executive Order on AI |
| 3 | NC | Executive | Membership roster | AI Leadership Council Members |
| 4 | NC | Executive | Meeting minutes | 11/05/25 AI Leadership Council Meeting Minutes |
| 5 | NC | Executive | Meeting minutes | 01/21/26 AI Leadership Council Meeting Minutes |
| ID | Issuing body | Issue date | Version or edition | Retrieved |
|---|---|---|---|---|
| North Carolina: Executive Branch | ||||
| 1 | Office of the Governor | Sep 2, 2025 | — | May 16, 2026 |
| 2 | Office of the Governor | Sep 2, 2025 | — | May 16, 2026 |
| 3 | NCDIT | Undated | Living page | May 16, 2026 |
| 4 | NCDIT | Nov 5, 2025 | — | May 16, 2026 |
| 5 | NCDIT | Jan 21, 2026 | — | May 16, 2026 |
| Stage | Model | Model-assisted task and output | Human review |
|---|---|---|---|
| 1 | GPT-5.4 | Generated candidate inductive codes for each of the 49 documents from its full text. Each assignment included a high-level label, low-level label, definition, and confidence rating. This stage produced 8,589 candidate inductive assignments. | Human review checked semantic support against the source document. |
| 2 | GPT-5.4 | Generated candidate deductive assignments to the five framework components, reading each whole document as a governance artifact. | Human review checked that each quotation appeared in the document and supported the code. Retained assignments were linked to Niti Formalization, Nyaya Recourse, Pluralistic Governance of AI, Adaptive Governance of AI with Incremental Revision, and Bounded Rationality in AI Governance. |
| 3 | GPT-4o-mini | Generated candidate semantic mappings between inductive labels and framework components, with primary and secondary mappings and confidence scores. | Confidence scores were used only as triage aids. Human reviewers made final mappings. |
| 4 | Researcher review | Reconciled mismatches and ambiguous cases, including criminal statutes and technical-object codes. | Reviewers inspected flagged source passages and distinguished scope boundaries, mapping errors, and possible framework extensions. The final analytic dataset contains 8,589 mapped instances. |
| Framework component | Source theory | Operational definition |
|---|---|---|
| Niti Formalization | Sen’s Niti | Formal rules, procedures, institutions, legal duties, oversight bodies, inventories, audits, transparency requirements, and accountability structures that make AI governance administrable. |
| Nyaya Recourse | Sen’s Nyaya | Preconditions for public protection at three levels: recognition (harms, affected groups, benefit claims), tracking (outcomes), and recourse (complaint handling, recourse pathways, enforcement links, correction mechanisms). A passage at the first two levels offers no recourse. |
| Pluralistic Governance of AI | Stone’s polis | Membership, power, influence, public-interest claims, market logic, information, co-operation, and representation in governance. The component asks who participates, whose evidence counts, and whose interests set the agenda. |
| Adaptive Governance of AI with Incremental Revision | Lindblom’s muddling through | Pilots, phased implementation, living documents, staged reform, partial revision, and pragmatic adjustment under uncertainty. |
| Bounded Rationality in AI Governance | Simon’s bounded rationality | Constraints of information, time, expertise, authority, staffing, and administrative bandwidth. The component includes satisficing, attention allocation, problem decomposition, local search, and routine reuse. |
| Subcode | Definition |
|---|---|
| Pillar a: stakeholders | |
| Membership | Defining who belongs to the relevant community and who is excluded from participation, governance, benefits, or protections. |
| Groups | Recognizing groups and organizations as the primary actors through which political action and policymaking occur. |
| Power | Using influence, co-operation, loyalty, and collective action to subordinate individual self-interest to group or public interests. |
| Pillar b: interests | |
| Motivation | Accounting for both self-interest and altruism as motivations for political action and policymaking. |
| Subcode | Definition |
|---|---|
| Satisficing ( Simon, 1955 ; Simon, 2000 ) | Accepting a workable option without searching for an optimal one. |
| Attention allocation ( Jones and Baumgartner, 2005 ; Ocasio, 1997 ) | Prioritizing some problems over others because attention, expertise, and institutional bandwidth are limited. |
| Problem decomposition ( March and Simon, 1993 ) | Splitting a complex governance problem into smaller parts addressed by different actors or at different times. |
| Local search ( Cyert and March, 1963 ) | Looking to nearby examples, existing standards, peer states, templates, or available models for solutions. |
| Routine reuse ( Nelson and Winter, 1985 ) | Reusing existing procedures, offices, definitions, or templates to reduce the burden of new governance work. |
| Code | Definition |
|---|---|
| Multilevel government authority | Different levels of government hold different forms of enforcement authority and act on the same AI problems at once. |
| Trust as precondition | Participants treat trust and a sense of protection as conditions for policy uptake and describe the council as a trust-building institution under uncertainty. |
| AI heterogeneity and policy granularity | Grouping traditional machine learning, generative AI, and agentic AI under one framework creates tension between broad policy and field-specific guidance. |
| Inevitability of AI | Participants treat AI adoption as already underway and difficult to stop, regardless of state action. |
| Experiential learning | Participants treat hands-on use of AI tools as a way to build practitioner intuition and governance knowledge. |
| Growth optimism | Participants frame AI mainly through innovation, scaling, industry attraction, or economic development, and environmental or ethical concerns sometimes receive less attention. |
| Passage function | Example content | Final component | Rationale |
|---|---|---|---|
| Harm recognition | Names sexual exploitation, CSAM, deepfakes, sextortion, intimate-image abuse, privacy loss, cybersecurity harm, unequal access, or workforce displacement. | Nyaya Recourse | The passage makes a public consequence visible, even if it does not specify measurement, recourse, or correction. |
| Formal legal response | Creates offenses, penalties, civil remedies, liability standards, court procedures, reporting duties, campaign rules, or whistleblower protections. | Niti Formalization | The passage establishes a rule, duty, remedy, procedure, or accountability structure. |
| Object of governance | Defines AI capabilities, legal categories, tool types, covered systems, synthetic media, or regulated uses. | Niti Formalization | The passage sets the scope of what the policy governs. |
| Implementation constraint | Describes enforcement difficulty, legal uncertainty, staffing limits, expertise gaps, agency burden, or interpretive uncertainty. | Bounded Rationality in AI Governance | The passage identifies what institutions can staff, interpret, enforce, or maintain. |
| Technical boundary case | Describes model architecture, compute, technical foundations, or infrastructure without connecting them to governance action. | Retained because governance relevant | Purely technical description was treated as outside the framework unless tied to authority, risk, deployment, oversight, or public consequence. |
| Question | Model-assisted pass | Planning reading and interviews |
|---|---|---|
| What do the documents commit to? | Applied one codebook to all 49 documents | Separated obligations from intentions by reading operative language and instrument type |
| Where does recourse appear? | Retrieved passages on harms and response mechanisms from all 49 documents | Recorded the date and binding status of each instrument (Section 5.1 ) |
| Who influences decisions? | Retrieved passages naming publics across the corpus | Compared named publics with rosters, testimony lists, and interview accounts (Section 5.4.2 ) |
| Does the framework fit? | Generated inductive labels independent of the framework and flagged 77 mismatches | Added Bounded Rationality in AI Governance from interviews and resolved mismatches by function (Section 3 ) |
| Requirement | Source theory and the condition AI governance adds | Concept or indicator | Observable indicator | Component |
|---|---|---|---|---|
| 1. Separating visibility to the state from recourse for the public | Sen separates institutions from lives, and one AI instrument can inform the state without serving the public | Procedural visibility and public contestability | Whether the state, the public, or both can use what an instrument records; whether a provision is an agency duty or an affected-person right | Niti Formalization; Nyaya Recourse |
| 2. Evaluating policy across instruments and over time | No single theory. Implementation research often follows one statute or program, and AI rules are split across instruments with different legal force | Binding sequence | Date written, binding status (in effect, enacted but not yet applicable, advisory, proposed, or absent), and who can enforce each duty | All five |
| Handoff specification | Whether a recommendation names who acts (implementation handoff) and the instrument that will enforce it (enforcement handoff) | All five | ||
| 3. Judging protection before outcomes exist | Sen judges justice by realized lives, and officials write AI rules before anyone can observe them | Preconditions of protection | Three levels kept separate: recognition (harms, groups, benefit claims), tracking (outcomes), recourse (complaints, pathways, enforcement, correction) | Nyaya Recourse |
| Depth of recourse specification | Whether a harm is named, tracked, routed to an office, or tied to enforcement and correction | Nyaya Recourse | ||
| Obligation rule | Binding commitments separated from statements of intent | Niti Formalization |
| Instrument | Who is affected? | What harm is tracked? | Where can a person complain? | Who can correct it? | Status |
|---|---|---|---|---|---|
| Responsible Use Framework, version 1.1 (NC, ID-6) | Those who may be impacted | AI use | No complaint route specified | Agency; no right enforceable by a person | In effect as internal direction |
| Executive Order No. 24 (NC, ID-1) | Implied by listed high-risk use cases | High-risk AI use | No individual complaint procedure specified | Agency AI oversight teams; no binding correction for individuals | In effect |
| Roadmap (NC, ID-49) | Broad publics | Broad risks | Public feedback mechanism, read by staff | No binding correction; a person makes final benefits determinations; proposed AI Liaisons for accountability escalation | Advisory |
| Senate Bill 735, edition 0 (NC, ID-25) | Public and users of covered models | AI safety incidents | No individual complaint procedure; developers report incidents to the attorney general | Attorney general enforcement | Proposed |
| House Bill 934, edition 2 (NC, ID-19) | Persons aggrieved by deepfakes | Deepfake harms | Civil action | Courts | Proposed |
| Session Law 2024-37, effective Dec 1, 2024 (NC, ID-21) | Identifiable minors; identifiable people in altered private images | AI-generated sexual imagery; altered private images | Criminal reporting; civil action | Courts and prosecutors; court may order the image destroyed | In effect |
| Recommendation | Primary owners | Timeline |
|---|---|---|
| Minimum statewide trustworthy AI standards | Department of Information Technology | Sep 2026 |
| External reporting with public feedback | Department of Information Technology; Agency AI Oversight Committees | Dec 2026 |
| Benefits-program transparency and human determination | Department of Information Technology | Mar 2027 |
| Policy package on AI harms | Governor’s Office; General Assembly | Mar 2027 |
| Policy package on worker displacement | Department of Commerce; General Assembly | Mar 2027 |
| Infrastructure-impact monitoring | Department of Environmental Quality; General Assembly; local governments | Jun 2027 |
AI Governance
Artificial intelligence governance studies how laws, institutions, and technical controls shape AI development, deployment, accountability, and social effects.
See all 412 papers