Governments should govern frontier AI through adaptive, scenario-aware institutions rather than fixed compliance regimes; effective oversight requires capability monitoring, conditional controls, and institutional redesign to remain robust across divergent technological futures.

Main Finding

Public-sector governance of frontier general‑purpose AI should be organized as adaptive risk management under deep uncertainty through 2030: combine capability monitoring, differentiated risk tiers, conditional (triggered) controls, layered technical and organizational safeguards, and continuous institutional learning rather than relying on static compliance rules or a single forecast of technological progress.

Key Points

• Evidence dilemma: AI capabilities are advancing rapidly but unevenly, while empirical knowledge about harms, safeguards, and intervention effectiveness lags — so governments cannot safely wait for full evidence nor can they adopt blanket restrictions based on speculation.
• Multiple plausible trajectories: OECD foresight treats AI progress through 2030 as several plausible, non‑probabilistic scenarios (slowdown, steady progress, acceleration). Governance needs robustness across these paths.
• Differentiated risk taxonomy: Risks grouped into malicious use, malfunctions (operational unreliability), and systemic risks (labor disruption, concentration, institutional dependence). Different risks demand different governance responses.
• AI in government is sociotechnical: Effective adoption requires organizational redesign (routines, accountability, training), data stewardship, and governance capacity — not just procurement of better models.
• Static regulation insufficient: One‑time rules can quickly become mismatched to capabilities or deployment modalities. Regulation should be adaptive, threshold‑based, auditable, and interoperable.
• Proposed 6‑layer operational framework for public institutions:
• Capability intelligence — active monitoring of autonomy horizons, multimodal and inference‑time capabilities.
• Risk tiering — classify use cases by risk type, sector criticality, and reversibility.
• Conditional controls — pre‑defined if‑then triggers to escalate controls when indicators appear.
• Defense‑in‑depth — layered technical, process and human safeguards.
• Sociotechnical implementation — documented organizational redesign and accountability for each high‑impact deployment.
• Learning & revision cycle — regular reviews of incidents, updates, procurement dependencies, and autonomy changes.
• Leadership & accountability: senior officials must define risk appetite, allocate monitoring resources, and ensure responsibility mapping across developers, integrators, deployers and oversight bodies.
• Data governance is foundational: reliable, auditable public‑sector AI depends on interoperable metadata, clear sharing rules, and federated data stewardship.
• Limitations: the article is a synthesis/framework proposal based on reports and literature; it does not provide original empirical testing of governance effectiveness.

Data & Methods

• Approach: conceptual synthesis and framework building based on institutional reports, OECD foresight, and recent digital‑government scholarship. The paper integrates policy documents and peer‑reviewed literature rather than presenting new empirical data.
• Main sources cited:
  • International AI Safety Report 2026 (Bengio et al.) — capability updates, incident typology, safety frameworks.
  • OECD papers (2026): Exploring Possible AI Trajectories through 2030; Trends in AI Incidents Reported by the Media; The OECD.AI Index — scenario methods, illustrative task‑horizon analyses, incident mapping, cross‑country readiness metrics.
  • Government Information Quarterly literature (2024–2025) on AI adoption, sociotechnical transformation, and data collaboration in public administration.
  • Earlier reviews on e‑government research limitations (Bannister, 2007) for theoretical grounding.
• Methods used in sources: foresight methods (trend analysis, horizon scanning, driver mapping), scenario building (non‑probabilistic trajectories), literature review, index construction. The paper synthesizes these findings into an operational governance framework and derives implications for institutional design.

Implications for AI Economics

• Uncertainty and policy design:
  • Macroeconomic and microeconomic policy must be robust to multiple AI progress scenarios; traditional cost‑benefit analyses that assume point forecasts are unreliable for high‑impact decisions.
  • Option‑value thinking and adaptive policy instruments (conditional triggers, staged rollouts) reduce the risk of locking in poor long‑run outcomes.
• Labor markets and productivity:
  • Heterogeneous capability progress implies uneven sectoral impacts. Economists should model distributions of task automation risk (task autonomy horizons) rather than uniform job‑level displacement.
  • Public‑sector adoption affects labor demand in government services, potentially accelerating complementarities (augmentation) or substitution depending on implementation and organizational redesign.
• Market structure and concentration:
  • Systemic risks include concentration of capability and data. Governance that depends on vendor transparency and standards can affect competition economics — well‑designed interoperability rules and data‑sharing incentives can mitigate lock‑in and market power.
• Externalities, measurement, and incident data:
  • Lack of reliable incident and prevalence data creates underspecified externalities. Improved public incident reporting and standardized metrics (beyond headline media counts) are priorities to enable meaningful economic valuation of harms and benefits.
  • OECD.AI Index and similar tools are useful but incomplete; economists need richer, time‑series, and comparable indicators (capability measures, deployment intensity, incident severity) to quantify welfare effects.
• Public goods and policy capacity:
  • Investments in public‑sector policy capacity (capability intelligence, data stewardship, monitoring institutions) are public goods that reduce coordination failures and negative externalities from misaligned deployments.
  • The cost of governance (monitoring, audits, conditional controls) should be modeled as part of the social cost of AI adoption; funding these functions may be justified by avoided harms and improved trust/efficiency.
• Regulatory economics and incentives:
  • Adaptive, tiered regulation alters firms’ compliance costs and innovation incentives. Well‑calibrated thresholds and auditability can encourage safer development while avoiding excessive barriers that shift activity to less regulated jurisdictions.
  • Conditional controls and reporting obligations change the strategic calculation for malicious actors and market entrants; enforcement design (liability, audits, transparency) will shape equilibrium behavior.
• Policy experiments and evaluation:
  • The proposed learning/revision cycles imply the need for randomized or quasi‑experimental evaluations of deployment models, procurement rules, and governance instruments to estimate causal effects on outcomes (service quality, fraud rates, labor outcomes).
• Research agenda for economists:
  • Develop models of governance‑technology co‑evolution (how regulatory regimes and organizational redesign influence diffusion and economic outcomes).
  • Build scenario‑based macro and sectoral models incorporating capability intelligence signals (task‑horizon indicators) to stress‑test policy options.
  • Quantify trade‑offs between rapid adoption (efficiency gains) and exposure to emergent systemic risks, including dynamic general equilibrium effects.

If you want, I can (a) convert the six operational layers into a short checklist for procurement and budgeting decisions, or (b) outline specific economic models and empirical strategies to measure the public‑sector impacts discussed above.