AI safety regimes emphasise keeping risks out but not organising responses when containment fails; adapting precommitment, shared protocols and a cross-actor 'note-exchange' of if-then response rules could close this structural coordination gap.

Main Finding

Frontier AI Safety Policies (FASPs) overwhelmingly emphasize prevention (capability testing, deployment gates, usage constraints) while under-investing in institutional coordination and robustness for when prevention fails. This "coordination gap" is structural: robustness yields diffuse social benefits but imposes concentrated private costs, producing systematic underinvestment. The paper proposes a practical institutional mechanism—the Scenario Response Registry (SRR)—to expose ex ante if–then response logic across actors, create common knowledge, and improve coordinated crisis response and learning.

Key Points

• Prevention-centric focus: Current FASPs (alignment testing, usage constraints, audits, regulatory norms) concentrate upstream control but leave mitigation, preparedness, and cross-actor coordination underdeveloped.
• Robustness gap defined: Policies assume control and accurate forecasts; in deep uncertainty they are brittle (gridlock or overconfidence) and cannot ensure acceptable system-level performance under shocks.
• Structural cause of underinvestment: Robustness is a public-good-like provision (diffuse social returns, concentrated private costs), compounded by strategic ambiguity and competitive pressures that favor flexibility over binding commitments.
• Lessons from other domains: Nuclear safety, pandemic preparedness, and cybersecurity use coordination primitives—precommitment, interoperable procedures, standing convening venues, standard vocabularies, and mandated escalation—to operationalize responses under uncertainty.
• Robustness thinking requirements: (1) make not only triggers but the decision logic transparent; (2) externalize triggers beyond internal technical metrics to include societal-impact signposts; (3) build on existing institutional capacity.
• Proposal — Scenario Response Registry (SRR):
  • A public authority curates a scenario library (technical, socio-economic, cross-domain).
  • Relevant actors file standardized, time-stamped if–then plans: "If [trigger], then [action], with [resources]."
  • The SRR harmonizes filings, identifies gaps/overlaps, stress-tests plans, and scores quality.
  • Incentives: regulatory forbearance, access to public compute, procurement eligibility linked to plan quality; bonds/insurance or forfeiture mechanisms for failing to act; dynamic scoring to influence audits/capital requirements.
  • Access can be tiered: public scenario library; sensitive filings restricted to relevant actors and authority.
• Complementary needs and constraints: incident databases, operational emergency frameworks, and drills must accompany SRR; geopolitical rivalry, competitive incentives, calibration risks, and maintenance costs complicate adoption.

Data & Methods

• Type of study: conceptual/policy analysis combining literature review, cross-domain institutional comparison, theoretical economic reasoning, and policy design.
• Sources and evidence:
  • Synthesizes existing FASPs (Anthropic ASL, OpenAI high/critical gates, DeepMind CCL) and recent transparency/coordination proposals.
  • Draws analogies from established risk regimes (IAEA, IHR, Sector Risk Management Agencies) and their coordination devices.
  • References empirical and experimental literature (e.g., climate threshold public-goods games) to illustrate cooperative failures under threshold uncertainty.
  • Leverages Robust Decision Making (RDM) as a methodological framing for stress-testing strategies across many plausible futures.
• Methods used:
  • Institutional comparison to extract coordination primitives and best practices.
  • Mechanism-design style reasoning to propose SRR architecture and incentive levers.
  • Conceptual modeling of incentive misalignment (public-good dynamics, free-riding, strategic ambiguity).
• Limitations:
  • No original empirical dataset or quantitative model is presented.
  • SRR remains a design proposal; claims about effectiveness are theoretical and rest on analogies and prior evidence from other domains.
  • Implementation feasibility (political, commercial, international) and calibration of incentives require empirical pilot testing and formal economic modeling.

Implications for AI Economics

• Market failure and investment allocation:
  • The paper identifies a clear market failure: private incentives favor speed and scale (first-mover, market share) while social welfare requires investments in robustness that firms will underprovide absent policy levers.
  • Economists should model how firms trade off deployment speed versus robustness investments and quantify social returns to robustness to inform optimal policy.
• Mechanism design opportunities:
  • SRR-style mechanisms create a space for mechanism design: linking disclosure quality to tangible regulatory or procurement benefits, designing bonds/insurance to create credible commitment, and dynamic scoring to internalize systemic risk.
  • Economic analysis can optimize penalty/reward structures (e.g., bond sizing, insurance pricing, procurement weightings) to achieve credible precommitment without unduly stifling innovation.
• Common knowledge and strategic interactions:
  • The SRR’s creation of common knowledge can transform assurance/coordination games that otherwise produce underinvestment; economists can formalize how publicized if–then commitments alter equilibrium behavior among competing labs and states.
  • Game-theoretic models (assurance games, threshold public-goods, dynamic games under uncertainty) can quantify tipping points where coordination becomes self-enforcing.
• New markets and risk pricing:
  • Anticipated demand for contingent-capacity commitments, bonds, and insurance products suggests market opportunities; economists can estimate required risk premia and capital costs and assess systemic liquidity implications.
  • Dynamic scoring and regulatory-linked capital requirements would affect firms’ cost of capital and valuation—modeling these effects is important for policy impact assessment.
• Competition, diffusion, and distributional effects:
  • Mandatory robustness disclosures or binding commitments may disadvantage small entrants or firms in adversarial jurisdictions; economists should study distributional effects and design proportional rules (tiering, exemptions, staged implementation).
  • International strategic dynamics: economists can analyze how SRR-like mechanisms interact with cross-border competition, leakage, and regulatory arbitrage.
• Empirical and modeling agenda suggested by the paper:
  • Estimate the social marginal benefit of additional robustness investment versus preventive measures across risk profiles.
  • Calibrate assurance-game experiments and large-scale simulations to test SRR designs and incentive schemes.
  • Cost-benefit analysis of SRR implementation (administrative costs, compliance burdens, effects on deployment timing).
  • Design and evaluate pilot programs (sector- or jurisdiction-limited SRRs) and measure behavioral responses, coordination outcomes, and market impacts.
  • Model insurance and bond markets for contingent response commitments, including moral hazard and verification issues.
• Policy evaluation metrics for economists:
  • Measures of improved coordination: reduction in mismatch/overlap of declared responses, time-to-harmonized action in drills, credibility measured by bond/insurance payouts.
  • Systemic metrics: reductions in systemic exposure to cascading harms, increased detection/response speed, and metricized learning rates from incidents.

Suggested next steps for applied economists working on this topic: - Build formal models (dynamic games and public-good frameworks) of SRR-like interventions. - Quantify costs to firms of credible readiness commitments and the social value of avoided cascading harms. - Design experimental pilots and field tests (e.g., limited-sector SRR) to observe signaling, compliance, and coordination outcomes.