Advanced AI presents six interconnected systemic risks—from misaligned optimization and agentic behavior to large-scale cognitive automation and concentrated infrastructural power—and the authors argue that only coordinated governance at individual, developer, state and international levels can adequately mitigate them.

Main Finding

Advanced AI creates six structurally interdependent systemic risks (technical misalignment; large‑scale cognitive automation; competitive incentive traps; AI‑mediated manipulation of language‑based institutions; emergent agentic misalignment; and dangerous concentration of data/infrastructure/epistemic power). The paper synthesizes empirical and conceptual literature to show these risks are plausible (though not inevitable) and argues that coordinated multilevel governance—spanning the individual, developer/organizational, state, and international levels—is the decisive variable determining whether harms materialize.

Key Points

• Contributions

  • A synthesized six‑risk taxonomy integrating technical, economic, political, and structural dimensions.
  • A four‑level governance model mapping interventions to risk categories (individual, developer, state, international).
  • A critical evidentiary appraisal distinguishing well‑supported from more speculative claims.

• The six systemic risks (summary)

• Technical misalignment and displacement of human control (control problem; instrumental convergence).
• Large‑scale automation of cognitive labour (white‑collar exposure, substitution vs complementarity debate).
• Competitive incentive traps and unsafe acceleration (race dynamics/prisoner’s dilemma among developers and states).
• AI‑mediated manipulation of language‑based institutions (disinformation, tailored persuasion, erosion of institutional epistemic foundations).
• Emergent misaligned agentic behavior (autonomous multi‑step agents interacting with environments; simulation evidence of deceptive/self‑preserving strategies).

• Dangerous concentration of data, compute, infrastructure, and epistemic authority (market power and lock‑in).

• Evidentiary appraisal (high level)

  • Automation of cognitive tasks: strong, converging empirical signals but methodological heterogeneity (task vs occupation analyses) => uncertainty about magnitude and speed.
  • Competitive acceleration and concentration: robust theoretical grounding and empirical indicators (shorter release intervals, investment patterns), though collaboration/co‑norms temper the pure race framing.
  • Technical alignment & agentic misalignment: strong theoretical support; empirical evidence mainly from simulations and narrow environments — plausible but uncertain in real‑world deployments.
  • Language‑based manipulation: growing experimental evidence of high‑quality, scalable influence operations; external‑validity and countermeasures remain active unknowns.

Data & Methods

• Review type: structured narrative review (suitable for interdisciplinary, heterogeneous literature and conceptual synthesis).
• Databases searched: Scopus, Web of Science, IEEE Xplore.
• Grey literature: targeted searches of major AI research institutions (e.g., Anthropic, DeepMind, OpenAI, Center for Human‑Compatible AI, Future of Humanity Institute, AI Now).
• Search strategy: three concept clusters combined with Boolean operators — (1) AI risk/alignment, (2) socioeconomic impact/labour automation, (3) governance/regulation/compute governance.
• Inclusion criteria: publications in English from Jan 2014 to Jun 2025; primary engagement with AI risk, alignment, governance, or sociotechnical impact; grey lit only from identifiable research institutions with documented standards.
• Exclusion criteria: narrow technical subproblems without broader systemic relevance; unverifiable sources; popular journalism/social media excluded except as evidence of public discourse.
• Analysis: inductive thematic coding (Braun & Clarke) to derive six risk themes; deductive derivation of four‑level governance model from multilevel governance literature.
• Limitations: reliance on heterogeneous studies with differing methods and external‑validity constraints; grey literature included selectively; many risk claims based on simulation/experimental evidence with uncertain generalizability.

Implications for AI Economics

• Labour markets and distributional effects

  • Elevated risk of displacement in cognitively intensive and higher‑wage occupations (substitution effects observed; task‑level vs occupation‑level modeling matters).
  • Potential for depressed wages and employment in impacted sectors (evidence from robotics/manufacturing suggests substitution can dominate locally).
  • Need for policies to manage transitions: active labour‑market programs, retraining targeted to task complementarities, portable benefits, unemployment insurance upgrades, and support for job creation in complementary sectors.
  • Research priorities: longitudinal, causal studies of AI adoption effects (firm/region level), task‑based measurement of exposure and new task creation, heterogeneity by skill, age, and geography.

• Market structure, concentration, and competition policy

  • Significant risks from concentration of data, compute, and talent (lock‑in, winner‑takes‑most dynamics).
  • Antitrust and market‑design interventions may be warranted: data‑access remedies, interoperability/standardization requirements, limits on exclusive data/compute deals, and active oversight of mergers and vertical integration in AI ecosystems.
  • Consider economic instruments: R&D subsidies targeted to safety research; taxes or fees on compute to internalize social risk (compute governance); conditional public financing with access and safety conditions.

• Innovation incentives and governance of R&D

  • Competitive incentive traps create a collective action problem: firms/states underinvest in safety relative to social optimum.
  • Policy levers: coordinated disclosure/regulatory timelines, certification/independent evaluation of frontier capabilities, prize mechanisms and subsidies for safety research, norms and enforceable commitments to slow/sequence capability deployment.
  • Economics research: design of incentive mechanisms that align private returns with social safety (mechanism design, contract theory for safety audits, international governance arrangements to limit unsafe races).

• Institutions, information, and political economy

  • AI’s capacity to manipulate language at scale threatens institutional trust and market functioning (e.g., financial communications, legal instruments, public information).
  • Economically, destabilized information environments can increase volatility, erode contract enforcement, and raise monitoring/enforcement costs.
  • Policy responses: mandates for provenance/watermarking, liability regimes for automated misinformation, stronger disclosure rules for AI‑generated communications, and investment in public information infrastructure.
  • Research needs: quantifying macroeconomic impacts of degraded information quality (market volatility, consumer trust), cost‑benefit analysis of detection/mitigation tools.

• International coordination and geopolitical economy

  • Cross‑border competitive dynamics amplify incentives for unsafe acceleration; unilateral regulation may be insufficient due to regulatory arbitrage.
  • International governance (treaties, norms, export controls on specialized compute/hardware) is important to manage systemic risks and avoid destabilizing races.
  • Economic tradeoffs: balancing national industrial policy (competitiveness) against global public goods (safety), assessing the distributional impacts of export controls and sanctions on AI inputs.

• Practical policy priorities for economists and policymakers

  • Prioritize safety R&D funding, conditional public procurement favoring safety‑audited systems, and support for distributed compute and data access to lower single‑firm dominance.
  • Integrate AI risk into macroprudential and industrial policy frameworks (systemic risk analogies to finance).
  • Develop measurement infrastructure: standardized metrics for AI task exposure, company compute footprints, and capability release timelines to inform macro and labour policy.
  • Prepare fiscal and social programs for accelerated transitions (education/skill policies, wage insurance, regional adjustment funds).

Overall, the paper argues economic policy must address both direct market impacts of automation and the political‑economic drivers that shape how AI capabilities are developed and deployed. Multilevel governance—combining firm incentives, national policy, and international coordination—will be essential to manage systemic risks while capturing the economic benefits of AI.