AI can strengthen state capacity in delivering public services but risks entrenching bias and eroding democratic accountability if deployed without institutional safeguards. The paper sets out a practical 'before–during–after' governance sequence—algorithmic impact assessment, human-in-the-loop controls, periodic external audits, traceability and appeal mechanisms—to balance efficiency gains against equity, transparency and legitimacy.

Main Finding

The authors argue that AI can strengthen state capacity and improve administrative efficiency and service personalization, but only if deployed within an institutional architecture that preserves equity, transparency and democratic legitimacy. They propose an operational, sequential governance framework (before, during, after deployment) built around four evaluative pillars — efficiency, equity, transparency, legitimacy — and specify verifiable indicators, minimum thresholds, and veto conditions to ensure that AI benefits do not come at the expense of social justice or democratic control.

Key Points

• Four analytic pillars:
  • Efficiency: operationalized via processing time, cost reductions, prediction capacity; rule: automation is acceptable only with verifiable performance metrics and active error-correction mechanisms.
  • Equity: focuses on distributive impacts; risks from biased or unrepresentative data can exacerbate structural inequalities; rule: require prior algorithmic impact assessments and ongoing mitigation mechanisms.
  • Transparency: stresses explainability, documentation, traceability and auditability (not necessarily full source-code disclosure); indicators include documentation, external audit access, and decision trace logs.
  • Legitimacy: concerns participation, oversight, and the preservation of democratic deliberation; rule: AI must not replace core deliberative functions and decisions affecting rights require human oversight and appeal channels.
• Distinction between personalization types:
  • Political personalization (communication, representation) vs administrative personalization (tailoring services and resource allocation). The paper focuses on administrative personalization: it can improve targeting and service concentration but risks amplifying inequality when data or oversight are weak.
• Governance sequence (operational):
  • Before: algorithmic impact assessment (distributive and rights risks), consider non-algorithmic alternatives, set minimum thresholds (data representativeness, auditability), contractual clauses for data access/transfer to avoid vendor lock-in.
  • During: human oversight at key decision points, documented objectives/assumptions, traceability of decisions.
  • After: periodic external audits, public disclosure of rationale and criteria, practical appeals mechanisms, continuous monitoring & corrective procedures.
• Practical constraints and veto assumptions: systems of high-impact on rights should not be automated absent representative data, auditability, or enforceable human supervision.
• Evidence base: a structured narrative review (2020–2025) and comparative cases in health, education, social protection inform the framework; the paper is conceptual, not an empirical evaluation.

Data & Methods

• Methodological design: mixed, combining (1) conceptual-analytical work to operationalize the four pillars and (2) a structured narrative review of interdisciplinary literature (2020–2025).
• Sources: academic articles, technical reports from multilateral organizations, policy documents, and comparative studies on AI in public services (health, education, social protection). Databases used include Google Scholar and Scopus; repositories on digital governance were consulted.
• Inclusion/exclusion: prioritized public-sector AI literature and governance/ethics studies; excluded purely technical, commercial-profit–oriented works and publications before 2019. No original empirical data collection or case-study fieldwork — limitations acknowledged.
• Outcome: a prescriptive, operational governance sequence with verifiable indicators and normative thresholds rather than statistical estimation or causal inference.

Implications for AI Economics

• Trade-offs and welfare:
  • Efficiency gains (lower administrative costs, faster processing, better prediction) can improve social welfare but may create regressive distributional effects if biases concentrate benefits or burdens; welfare analysis must internalize both efficiency gains and distributional harms.
• Measurement and metrics economists should adopt:
  • Efficiency: processing time, unit administrative cost, predictive accuracy (AUC, RMSE) on policy-relevant outcomes.
  • Distributional impacts: subgroup false positive/negative rates, changes in benefit coverage by income/ethnic/vulnerable status, Gini or other inequality measures on access and outcomes.
  • Transparency/legitimacy: rates of appeals, overturns of automated decisions, public trust indices, administrative compliance costs (audits, oversight).
• Empirical research directions:
  • Estimate net welfare effects of AI adoption in public programs via RCTs, stepped-wedge designs, difference-in-differences, or synthetic controls where feasible.
  • Audit and fairness studies to quantify bias-related welfare losses across groups (audit studies, targeted counterfactual simulations).
  • Cost-effectiveness of safeguards: compare marginal benefits of algorithmic accuracy versus costs of audits, human review, and data-improvement interventions.
  • Market and contract studies: measure vendor lock-in risks, pricing of data-access/transfer clauses, and how procurement design shapes incentives for transparency.
• Policy and regulatory economics:
  • Design of incentive-compatible regulations: calibrate thresholds for allowable automation depending on impact level, require minimum data representativeness and auditability, and mandate contractual provisions to reduce switching costs.
  • Consider compliance costs and administrative capacity constraints: smaller governments may face higher relative costs to implement recommended safeguards, implying potential unequal adoption and externalities across jurisdictions.
• Labor and organizational effects:
  • Reallocation of public sector labor from routine processing to oversight, appeals, and user-facing roles; need to model transitional unemployment/ retraining costs and productivity gains.
• Limitations and empirical gaps:
  • The paper is conceptual; economists should empirically validate assumed links between governance safeguards and outcomes (e.g., do mandatory audits materially reduce discriminatory outcomes?).
  • Data needs: representative administrative datasets, recorded decision logs, and access to procurement/contracts to estimate economic consequences.
• Practical recommendation for economic researchers and policy designers:
  • Incorporate distributional tests and transparent performance metrics into impact evaluations of public-sector AI.
  • Build the cost side of analyses to include ongoing governance (audits, traceability, appeals) when comparing AI-based interventions with non-algorithmic alternatives.