An Evaluation Framework for National AI Regulation

• AI supports public administration, services, healthcare, hiring, and transport, but its value depends on data quality, performance, human review, and correction mechanisms.These systems can affect access to services, employment, mobility, and individual rights.Source: §1 · p. 1
• Statutes, strategies, voluntary guidance, standards, procurement rules, funding, sandboxes, and training create different duties, rights, incentives, and implementation demands.Institutional mandates and budgets determine whether commitments can be carried out, so national AI governance is better understood as a policy portfolio.Source: §1 · p. 2
• The paper develops a comparative method for eight jurisdictions that scores versioned national portfolios and documents the evidence supporting what policies provide on paper.It does not treat documented commitments as proof of compliance or social benefit.Source: §1 · p. 2

Metrics for Comparison.

• National capacity, infrastructure, political conditions, demand and talent explain implementation feasibility but remain context rather than inputs to the policy score.The comparison distinguishes public, private and cloud compute, records economic and external constraints, and avoids combining incomparable market, workforce or stability measures.Source: Metrics for Comparison · pp. 3–4

Analysis of Selected Jurisdictions.

• The eight-jurisdiction comparison is purposive, spanning different technical capacities and governance models rather than representing global prevalence.It includes China, India, Japan, Singapore, South Korea, the United Kingdom, the United States and the European Union, while keeping contextual advantages separate from policy scores.Source: Analysis of Selected Jurisdictions · p. 4

2.2 Cross-National Discussion of AI Policy Approaches

• Because jurisdictions combine laws, strategies, guidance, institutions and development programs differently, the framework evaluates portfolios and records legal force separately from substantive detail.The examples show why isolated instruments can mislead: Singapore relies heavily on voluntary guidance, Japan separates statutory status from coercive enforcement, Korea combines distinct AI classifications, and China distributes obligations across instruments.Source: §2.2 · pp. 4, 7, 8
• Emerging general-purpose, advanced and agentic systems require capability- and lifecycle-based coverage criteria rather than reliance on jurisdiction-specific labels.International coordination can produce shared practices without a stable national model, while agentic systems may be governed through general risk, human-oversight or sectoral provisions even without an agent-specific label.Source: §2.2 · p. 8

2.3 Related Work

• Prior comparative legal, principles-based, readiness and responsible-AI indices address related questions but generally use broader units or include implementation and capacity evidence.These approaches clarify institutional differences and enable structured comparison, yet they do not necessarily provide an operational protocol for scoring provisions across mixed national instruments.Source: §2.3 · p. 9
• This framework instead scores what a declared portfolio provides and how specifically it provides it, treating economic capacity as context and avoiding outcome inference.Corporate grading rubrics offer a methodological analogy through tiered, visible criteria, but national portfolios additionally encompass legislation, executive policy, strategies, mandates, procurement rules, guidance and budgets.Source: §2.3 · pp. 8–9
• A national evaluation must define its unit, cite the supporting provision for every score, separate commitment detail from legal status, and report category results.Without category results, an aggregate score can hide weaknesses in safety, rights, or institutional capacity.Source: §2.3 · p. 9
• Versioned portfolios and a cutoff are necessary because laws, guidance, budgets, and decisions can change which instruments coexist and how their commitments operate.A score without a cutoff may combine instruments that were never simultaneously in force.Source: §2.3 · p. 9
• Unsuccessful searches are not treated as proof that provisions are absent; the method records source searches, uses a separate not-scorable state, and reports rubric coverage.This prevents portfolios from being penalized solely because official material is difficult to access or differs in language and detail.Source: §2.3 · pp. 9–10
• The framework integrates principles, technical risk, rights, distributional effects, innovation, infrastructure, and enforcement questions while reporting their results separately.Domain-specific frameworks offer greater depth within individual areas, but the national comparison connects those areas to the full policy portfolio.Source: §2.3 · pp. 9–10

3.1 Framework Development and Design

• The framework uses purposive design synthesis, combining AI safety research, governance and readiness frameworks, and official instruments from eight jurisdictions.Technical research informed misuse, control, testing, intervention, and resilience, while governance sources addressed rights, institutions, enforcement, participation, and policy change (Hendrycks, Mazeika, and Woodside 2023; Slattery et al. 2025; Ji et al. 2025).Source: §3.1 · p. 10
• It converts broad policy concerns into separately scored questions about mechanisms, responsibilities, enforcement, accountability, oversight, liability, and redress.This prevents merely naming a risk from being treated as evidence of an operating governmental response.Source: §3.1 · p. 10
• Candidate concepts were grouped, split, or merged so each retained sub-criterion identified one distinct, source-justifiable policy feature without double counting.The design separates legal force from substantive detail because they can differ in practice.Source: §3.1 · p. 10
• The framework prioritizes serious AI risks while separately assessing implementation, rights, distributional effects, development policy, and responsible innovation.Responsible innovation cannot silently compensate for weak safety or rights provisions.Source: §3.1 · p. 10
• Ordered scoring anchors distinguish absence, general statements, partial mechanisms, clear operating provisions, and fully specified responses without treating the overall mean as the main diagnostic.Category and criterion results carry more information, while the score supports explicit judgment and repeat assessment.Source: §3.1 · p. 10
• The unit of analysis is a versioned national policy portfolio containing adopted or operative official instruments, institutional mandates, and budget commitments within a declared cutoff.The portfolio records instrument dates or versions, excludes subnational instruments unless specified, treats the European Union as a supranational comparator, and uses capability measures only as context.Source: §3.1 · p. 10
• Traceable evidence, preserved cutoff dates, category reporting, and rescoring make assessments auditable when portfolios change, but do not establish universal rubric reliability or validity.The protocol still requires testing by multiple evaluators, sensitivity analysis for weighting, and comparison with implementation evidence.Source: §3.1 · p. 10

3.2 Framework Structure

• Five categories and 25 criteria organize the framework, with sub-criteria serving as the scoring unit and indicators directing evidence searches without separate scores.The default scheme weights categories equally; Section 3.3 specifies remaining weights.Source: §3.2 · p. 11, Table 3 · p. 12

AI Risk Governance and Safety.

• AI risk governance evaluates controls for high-consequence systems, including testing, independent red teaming, auditing, human intervention, containment, misuse, concentration, and infrastructure resilience.Advanced or general-purpose systems qualify based on operative coverage rather than terminology, while long-term resilience and policy updating remain separate questions.Source: AI Risk Governance and Safety · p. 11

Effectiveness and Feasibility.

• Implementation readiness assesses institutional mandates, authority, expertise, funding, independence, clarity, enforceability, measurability, feasibility, and incentives before outcomes are observed.Effectiveness here means design and implementation readiness, not evidence that a policy caused an outcome.Source: Effectiveness and Feasibility · p. 11

Comprehensiveness and Scope.

• Comprehensiveness and scope examine responsibility across the AI lifecycle, coverage of technical, social, environmental, and security risks, stakeholder inclusion, redress, international coordination, and adaptability.A single instrument need not cover every lifecycle stage, but the portfolio should allocate responsibility across them and define an update process.Source: Comprehensiveness and Scope · p. 11

User Rights, Protection and Agency.

• Protections and agency assess fairness, transparency, privacy, data rights, accountability, independent oversight, complaint routes, liability, and meaningful redress for people affected by AI.The framework distinguishes public system descriptions from decision-specific explanations and scores overlapping protections separately.Source: User Rights, Protection and Agency · pp. 11–12

Socioeconomic Impact and Innovation.

• Socioeconomic impact and innovation assess public needs, innovation and competition programs, workforce support, labor protection, institutional expertise, social equity, vulnerable groups, and concrete trust measures.The category scores documented commitments and implementation provisions, not later economic growth, public confidence, or distributional outcomes without separate evidence.Source: Socioeconomic Impact and Innovation · p. 12

3.3 Scoring Methodology and Aggregation.

• Evaluators inventory a fixed, versioned portfolio and search protocol, then assign one evidence-backed score per sub-criterion rather than scoring indicators separately.Each score sheet records the cited provision, rationale, source-search notes, and the operative text supporting the selected anchor.Source: §3.3 · pp. 12–13

Sub-criterion Assessment on a 1–5 Scale

• The 1–5 scale uses ordered anchors selected as complete descriptions, while not-scorable items are excluded and renormalized instead of being treated as absent.A score of 1 requires evidence of absence; coverage rates and search notes disclose missing evidence, and an entirely unscorable category prevents reporting an overall score.Source: Sub-criterion Assessment on a 1–5 Scale · p. 13

Weighted Aggregation

• Default aggregation assigns equal weights hierarchically, while custom weights must be preregistered, justified, and accompanied by the default result for comparability.Sensitivity analysis is especially important when totals are close or missing evidence causes substantial renormalization.Source: Weighted Aggregation · p. 13

Final Score Interpretation (1-5 Scale)

• Interpretation reports overall, category, criterion, coverage, and instrument-level results, emphasizing evidence notes and reliability rather than treating totals as causal or independently decisive.Evaluator agreement should be measured before reconciliation, and portfolio updates should create new versions to preserve longitudinal comparability and the underlying evidence.Source: Final Score Interpretation (1-5 Scale) · pp. 13–14

Contribution

• The framework preserves policy differences by scoring technical risk governance, institutional capacity, lifecycle coverage, protections and agency, and public benefit separately.This prevents development programs from compensating silently for weak safety or rights provisions and connects broad governance comparison to traceable policy evidence.Source: §4 · p. 14

Legal force in context

• Legal force is one recorded dimension rather than a proxy for policy quality, substantive coverage, practical influence, enforcement, or overall portfolio performance.The EU AI Act and Republic of Korea’s AI Basic Act would score 5 for binding legal force, while Singapore’s nonbinding Model AI Governance Framework would score 2; these values address only the legal-status sub-criterion (European Union 2021b; Republic of Korea 2025; Personal Data Protection Commission Singapore 2020).Source: §4 · p. 14

Using and interpreting the framework

• Assessment requires a defined, versioned portfolio, documented search and translation procedures, cited evidence for each score, independent scoring, agreement statistics, and transparent handling of custom weights and missing evidence.Interpretation should begin with category and criterion results, while scores of 1, 2, and not-scorable distinguish absence, vague acknowledgment, and insufficient evidence, which imply different responses.Source: §4 · pp. 14–15

Limits of document review

• Document review improves auditability but can underrate effective informal practice, soft law, unpublished institutional capacity, and weakly enforced laws, while public-information gaps and cross-jurisdictional interpretation limit certainty.Independent scoring exposes disagreement, but legal and linguistic expertise remains important; the framework measures documented design and readiness, not outcomes or enforcement success.Source: §4 · p. 15

Future work

• Future empirical work should apply a frozen rubric to complete portfolios with multiple evaluators, publish inventories and evidence records, preserve initial and consensus scores, and test sensitivity to weights and missing-evidence rules.These materials would show which anchors are applied consistently and which require revision; a maintained public implementation could improve traceability and reuse without replacing expert judgment.Source: §4 · p. 15
• Specialized modules can address foundation models, agents, procurement, healthcare, and finance while preserving a stable core, whereas outcome validation requires longitudinal causal research linking design scores to enforcement, compliance, incidents, budgets, complaints, adoption, and broader outcomes.Modules should not alter the core score unless declared in advance, and outcome studies remain separate because the present content-based framework does not provide causal design.Source: §4 · pp. 15–16

• National AI governance combines binding rules, institutional authority, resources, and voluntary guidance in different ways, so no single prominent law or strategy is sufficient for comparison.Binding provisions may depend on later guidance and capable institutions, while development strategies and voluntary frameworks can shape practice without creating enforceable rights.Source: §5 · p. 16
• The eight-jurisdiction comparison shows why the portfolio is the appropriate unit of analysis, from comprehensive statutes in the European Union and Republic of Korea to distributed, sectoral, guidance-led, developmental, or platform-focused approaches.The United States distributes federal policy across executive action and agencies; the United Kingdom relies on sectoral regulators; Singapore and Japan emphasize guidance and coordination; India combines development policy and data protection; and China combines platform and content rules with cybersecurity and data law.Source: §5 · p. 16
• The framework freezes a national portfolio at a stated cutoff and scores 25 criteria covering serious risks, implementation capacity, scope, affected-person protections, and AI development.Each sub-criterion requires a cited provision and ordered judgment, while category results, coverage reports, instrument lists, and cutoff dates support traceability and later rescoring.Source: §5 · p. 16
• A strong score indicates documented design and implementation readiness, not enforcement, organizational compliance, improved safety, or stronger individual rights.Innovation and public welfare require separate outcome evidence, and the next step is applying one frozen rubric to complete portfolios with independent evaluators to test reliability and sensitivity.Source: §5 · p. 16

Scope and role

• The thirteen-jurisdiction infrastructure inventory informed purposive selection but does not enter the policy score.The documented resource base includes public and private facilities used in the comparison.Source: A Compute Infrastructure Review · pp. 17, 20

Public and private compute

• Singapore combines S$270 million for National Supercomputing Centre expansion and training with public, commercial, and major-cloud-provider compute capacity.AWS, Microsoft Azure, and Google Cloud operate Singapore cloud regions, alongside commercial private cloud and accelerator capacity.Source: A Compute Infrastructure Review · p. 17
• Japan’s ABCI 3.0 became publicly available in January 2025 with 6,128 NVIDIA H200 GPUs and a stated 6.22 EFLOPS FP16 peak.The comparison does not equate commercial cloud regions with scientific supercomputers.Source: A Compute Infrastructure Review · p. 17

Contextual conditions

• The contextual tables compile economic, political, demand, research, and talent signals from heterogeneous sources rather than a common stability, demand, or talent measure.The entries do not establish causal effects or provide a comparable cross-jurisdiction ranking.Source: A Compute Infrastructure Review · pp. 18, 19, 21

• Universities and digital-skills programs train AI workers, but differing occupational definitions prevent combining available sources into one workforce count; Chinese firms and public entities procure or provide domestic AI services.The Chinese providers include Baidu, Alibaba, Tencent, government agencies, and state-owned enterprises.Source: A Compute Infrastructure Review · p. 22
• Universities, national laboratories, the ELLIS network, and member-state research programs support AI research across Europe.Source: A Compute Infrastructure Review · p. 22
• EU institutions and national programs support AI adoption in manufacturing, finance, healthcare, and public services.Source: A Compute Infrastructure Review · p. 22
• China produced 23.2% of AI publications in 2023, while 11% of China-educated elite AI researchers in the 2024–2025 conference sample worked in China, down from 16% in 2019.These figures come from the 2025 AI Index and MacroPolo’s conference sample, respectively.Source: A Compute Infrastructure Review · p. 22
• Supercomputing centers, universities, research institutes, and company laboratories contribute to AI research.Source: A Compute Infrastructure Review · p. 22

• As of August 15, 2026, the tables report what cited sources state and do not infer causal effects.Source: B Cross-National Policy Overview · p. 23

United States

• The United States combines executive orders with voluntary AI risk-management guidance, while its operative federal policy portfolio excludes state AI laws and emphasizes leadership, infrastructure, standards, security, and diplomacy.A prior executive order was revoked on January 20, 2025, and the AI Diffusion Rule was rescinded in May 2025 before taking effect.Source: B Cross-National Policy Overview · p. 24

• Across the portfolios, instrument scope and legal force vary: some measures create binding duties, while others provide coordination, standards, governance processes, or sectoral guidance without approval powers or penalties.The U.S. AI Risk Management Framework is voluntary and creates no independent approval requirement, the National AI Initiative Act authorizes programs rather than model-specific approval, and Singapore’s healthcare guidance is not a standalone approval statute.Source: B Cross-National Policy Overview · pp. 24–25

Singapore

• Singapore’s portfolio combines binding, technology-neutral personal-data duties with nonbinding organizational, healthcare, and financial-sector AI guidance.The PDPA imposes applicable consent, purpose, notification, access, correction, protection, retention, and transfer duties, while sectoral guidance addresses validation, monitoring, fairness, explainability, accountability, cybersecurity, and patient-centered design (Personal Data Protection Commission Singapore 2020; Ministry of Health Singapore, Health Sciences Authority, and Integrated Health Information Systems 2023; Monetary Authority of Singapore 2021).Source: B Cross-National Policy Overview · p. 25

India

• India combines a nonbinding national AI strategy with binding data-protection law and platform-focused advisory duties rather than a general AI approval or audit regime.The Digital Personal Data Protection Act 2023 and 2025 Rules provide rights and duties with staged implementation, while the revised MeitY advisory addresses unlawful content, discrimination, electoral integrity, unreliable outputs, and synthetic misinformation without requiring prior permission for under-tested models (NITI Aayog 2018; Government of India 2023b; Press Information Bureau, Government of India 2025; AZB and Partners 2024).Source: B Cross-National Policy Overview · p. 28

South Korea

• South Korea combines a 2025 binding AI framework with data-protection rights for automated decisions and nonbinding ethics guidance.The framework establishes planning institutions, transparency duties, safety measures for high-performance AI, and safeguards for high-impact AI; the data-protection law provides explanation rights for qualifying automated decisions, while ethics guidance sets ten requirements and creates no direct penalty or approval mechanism (Republic of Korea 2025; Jon 2026; Personal Information Protection Commission, Republic of Korea 2023; Ministry of Science and ICT, Republic of Korea 2020).Source: B Cross-National Policy Overview · p. 30

Scope and methodological distinction

• The reviewed frameworks differ from the present approach because they assess country conditions, governance capacity, companies, systems, or policy formation rather than versioned national policy portfolios.The present framework isolates the documented design of a declared policy portfolio, whereas several alternatives combine policy content with implementation evidence or broader readiness conditions.Source: C Existing Policy Evaluation Framework… · p. 33

National readiness frameworks

• UNESCO’s readiness methodology supports national, multi-stakeholder diagnosis and reform recommendations, but it is not a fixed comparative rubric for policy-portfolio content.It combines document review and stakeholder participation with legal, social, economic, scientific, educational, and technical conditions and institutional capacity.Source: C Existing Policy Evaluation Framework… · p. 33
• The AI Governance International Evaluation Index, Global Index on Responsible AI, and Government AI Readiness Index measure broad national governance or readiness across countries and jurisdictions.Their coverage includes policy and implementation, civil-society engagement, enabling conditions, government uses, governance practices, research outputs, risk exposure, institutions, infrastructure, skills, and private-sector conditions.Source: C Existing Policy Evaluation Framework… · p. 33

Company safety frameworks

• Company-focused frameworks demonstrate detailed scoring of voluntary or frontier safety commitments, but their units are companies rather than national portfolios of laws, strategies, budgets, and guidance.One study applies 65 weighted criteria across four risk-management dimensions to 12 company frameworks and shows how rubrics expose vague or missing commitments.Source: C Existing Policy Evaluation Framework… · p. 33

Supporting evaluation frameworks

• Other reviewed frameworks inform terminology, lifecycle mapping, stakeholder roles, privacy and security criteria, implementation, or policy formation without providing the present cross-national portfolio-scoring structure.The meta-framework lacks an operational scoring protocol, while system and clinical frameworks use narrower units than national policy.Source: C Existing Policy Evaluation Framework… · p. 33

D Complete National AI Policy Evaluation Framework

• Each of the five categories receives 20%, with criteria equally weighted within categories and sub-criteria equally weighted within criteria.This implements the default weighting rule described for the framework.Source: D Complete National AI Policy… · p. 35

Category 1. AI Risk Governance and Safety. Weight 20%

• AI risk governance and safety covers red teaming, auditing, human oversight, system interfacing, long-term governance, and infrastructure resilience.The supplied framework materials identify these as sub-criteria within the first 20%-weighted category.Source: Category 1. AI Risk Governance and… · pp. 36–38, Table 20 · p. 38, Table 21 · p. 39

Category 2. Effectiveness and Feasibility. Weight 20%

• Effectiveness and feasibility assesses institutional capacity, authority, resources, compliance monitoring, enforcement, legal force, measurability, and incentive alignment.Its indicators distinguish oversight mandates, independence, expertise, monitoring, penalties, binding status, reporting, performance metrics, and resource requirements.Source: Category 2. Effectiveness and… · pp. 39, 40, 42

Category 3. Comprehensiveness and Scope. Weight 20%

• Comprehensiveness and scope evaluates lifecycle coverage, risk range, stakeholder inclusion, adaptability to technological advances, and international harmonization.International cooperation is assessed through information sharing, joint research, and coordinated policy development.Source: Category 3. Comprehensiveness and… · pp. 44, 45, 50

Category 4. User Rights, Protection and Agency. Weight 20%

• User rights, protection and agency includes fairness and non-discrimination, transparency and explainability, and accountability criteria.The category is represented through dedicated sub-criteria and indicators for these protections.Source: Category 4. User Rights, Protection… · p. 50, Table 36 · p. 51, Table 37 · p. 51

Category 5. Socioeconomic Impact and Innovation. Weight 20%

• Socioeconomic impact and innovation evaluates societal benefit, innovation, workforce development, market competitiveness, economic growth, and public trust.Its indicators address retraining, education, job placement, antitrust measures, fair competition, and support for smaller enterprises.Source: Table 40 · p. 53, Table 41 · p. 53, Table 42 · p. 54