AI creates systemic societal risks through feedback loops, market concentration and tightly integrated AI ecosystems that produce cascading harms and dependencies. Addressing these risks requires collective‑action and governance solutions rather than isolated technical fixes.

Main Finding

The paper proposes a conceptualisation of "systemic risks of AI" as emergent harms that arise from complexity, externalities, and collective‑action problems in sociotechnical systems. Rather than treating systemic risk as attributable mainly to single GPAI providers, systemic AI risk is best understood as macro‑level failures (or threats to emergent common goods such as democratic functioning, fundamental rights, environmental stability, or financial stability) produced by interactions, feedbacks, dependencies, and institutional incentives across actors, technologies, markets, and governance frameworks.

Key Points

• Definition & framing
  • Systemic AI risks are harms that emerge at societal or global scale from interactions among many elements of sociotechnical systems, not reducible to individual component failures.
  • The paper frames systemic risk through complexity science notions (emergence, nonlinearity, tipping points, feedbacks) and treats societal goals (e.g., rights, democracy, safety, environment) as emergent common goods.
• Three complementary actor constellations
  • Individual‑actor focus: risks traceable to a single actor or provider (e.g., misuse, malicious actors).
  • Interaction/interdependence focus: risks from complex interactions, cascading failures, and network structure.
  • Governance/institutional focus: risks rooted in incentive structures, regulation, or institutional design that shape interactions.
• Core mechanisms producing systemic AI risk
  • Collective action problems and externalities (public goods, free‑rider problems, common‑pool resources).
  • Feedback dynamics, rebound effects, and informational emergence.
  • Market concentration and network effects that create centralised nodes and dependencies.
  • Algorithmic monocultures and outcome homogenisation (amplify cascade/correlated failures).
  • Integration along AI value chains and “AI ecosystems” that increase structural dependencies and cascading risk.
  • Information asymmetries, proxy bias, and governance/institutional deficits that impede detection, coordination, and mitigation.
• Critique of prevailing regulatory framing
  • The EU AI Act’s emphasis on GPAI providers and high‑impact capabilities is useful but overly narrow; systemic risk also arises from many other models, supply‑chain integration, market structures, and governance failures.
• Assessment implications
  • Systemic risk assessment must reconstruct interacting processes and patterns likely to produce macro‑level harms (not just enumerate individual model risks).
  • Nonlinear accumulation, path‑dependence, and cross‑sector coupling complicate probabilistic and cost–benefit assessment.
• Open challenges
  • Measuring and modelling systemic externalities, tail risks, informational emergence.
  • Designing governance that addresses collective‑action failures, reduces incentives for risky centralisation, and enables coordination across jurisdictions and sectors.

Data & Methods

• Methodological approach: conceptual synthesis and theoretical framing.
  • Extensive literature review synthesising research on systemic risk, complexity and emergence, public‑goods theory, and nascent research on systemic harms from AI (including regulation such as the EU AI Act).
  • Sociotechnical perspective: analyses technical elements (models, datasets, architectures) together with human actors, business models, institutions, norms, and governance.
  • Abstraction into an analytical heuristic: treat systemic risks as complex externalities and collective‑action problems and identify recurring phenomena (feedbacks, network effects, monocultures, etc.) that drive emergence.
• Empirical scope and limits
  • Paper is conceptual/theoretical rather than empirical; builds on cross‑disciplinary evidence and analogies (finance, climate, discrimination, safety engineering) rather than new quantitative data.
  • Limitations include lack of calibrated empirical estimates of probabilities or magnitudes of specific systemic outcomes; intended as an analytical foundation for further empirical and policy work.

Implications for AI Economics

• Market structure and competition policy
  • Network effects, economies of scale, and integration in AI supply chains promote concentration and central nodes — increasing systemic fragility and correlated failure risk. Economic policy should treat such concentration as a systemic externality.
  • Antitrust and competition policy should consider systemic‑risk externalities (not only price/consumer welfare metrics): interoperability, data portability, and limits on vertical integration can reduce monoculture risk.
• Public goods and underprovision
  • AI safety, privacy protection, and protection of collective rights have public‑good character. Left to private incentives, these are likely underprovided (free‑riding on others’ investments in safety/compliance).
  • Economic instruments: public funding of safety R&D, standards development, and provision of shared safety infrastructure (testing, red‑teaming, incident reporting).
• Regulation as macroprudential policy
  • Analogous to macroprudential regulation in finance, regulators should design tools to internalise systemic externalities: mandatory resilience standards, capital/insurance requirements for systemically important AI providers, mandatory disclosure and stress‑testing of AI systems at scale.
  • Costs of precautionary rules should be weighed against nonlinear, tail, and systemic harms that standard microeconomic cost–benefit methods may mismeasure.
• Liability, insurance, and risk allocation
  • Insurance markets face correlated, fat‑tail risks from algorithmic monocultures and cascading failures; public backstops or compulsory pooling mechanisms may be needed.
  • Liability regimes should account for systemic externalities and incentives for safe design across supply chains (not only end‑user harms).
• Coordination, information, and institutions
  • Economic and regulatory solutions must address coordination failures: incentives for information sharing (incident data, model evaluations), platforms for collective action (industry safety consortia), and international cooperation to manage cross‑border spillovers.
  • Investing in measurement and monitoring institutions (real‑time metrics of concentration, dependency networks, model similarity) is an economic priority to reduce information asymmetries.
• Research and measurement agenda for economists
  • Quantify externalities: model how private decisions (model scaling, dataset reuse, vertical integrations) translate into systemic risk externalities.
  • Network and agent‑based models: simulate cascade dynamics, path‑dependence, and tipping points in markets and infrastructures with algorithmic agents.
  • Game‑theoretic analysis: cooperative vs free‑riding incentives for safety investments and standard adoption among competing firms.
  • Macro‑economic modelling of systemic AI shocks: estimate GDP, financial‑market, labor‑market, and welfare impacts of correlated AI failures or systemic misuse.
  • Empirical metrics: development of indicators for algorithmic monoculture (model similarity indexes), concentration (Herfindahl indexes adapted to model/data/control), and systemic importance of providers.
• Policy design highlights
  • Treat systemic risk as a market failure requiring public intervention beyond standard product regulation.
  • Combine competition policy, targeted regulation for systemically important AI actors, public investment in safety public goods, and internationally coordinated governance to manage cross‑border externalities.

If you want, I can: - Draft a short policy brief for economists or regulators summarising recommended economic instruments and concrete steps; - Outline an empirical research plan (data needs, modelling approaches) to quantify the systemic externalities discussed.