Source-linked AI summary
Non-Great-Power Conflict and AI Risk
Kristina Kempkey, Seán Boddy, Catherine Ge-Wang
TL;DR
Research has focused on great-power conflict, leaving the relative importance of non-great-power conflict for AI-related catastrophic risk unclear. This paper examines three risk pathways and finds the case for treating non-great-power conflict as a minor contributor poorly supported, while loss-of-control effects remain unsettled.
Problem
Research has focused on great-power conflict, while the magnitude of non-great-power conflict’s risk relative to great-power conflict remains unclear.
Method
The paper uses causal models, literature reviews, and qualitative assessments to evaluate three pathways from non-great-power conflict to catastrophic risk.
Results
The null hypothesis that non-great-power conflict is a minor contributor is not well supported; loss-of-control effects remain the most speculative and unparameterized.
Takeaways & Limitations
Non-great-power conflict should not be treated as a negligible contributor to catastrophic risk relative to great-power conflict.
Takeaways & Limitations
The loss-of-control sub-hypothesis is the most speculative, and the paper does not attempt to parameterize its risk.
Abstract
from arXiv · showhide
Research on advanced AI and the risk of war has focused almost exclusively on great power conflict, on the grounds that confrontation between nuclear-armed adversaries poses the greatest risk of catastrophic or existential harm. Considerably less attention has been paid to non-great-power conflict (NGPC): wars between non-great powers, between non-great powers and great powers, civil wars, proxy wars, and conflicts involving nonstate actors. This paper evaluates the null hypothesis that NGPC is much less important than great power conflict (GPC) as a source of catastrophic risk in an era of increasingly capable AI, against the alternative that it is within an order of magnitude of GPC in importance. We assess three sub-hypotheses: that NGPC increases the likelihood of great power conflict; that it increases the expected harm from catastrophic terrorism; and that it increases the expected harm from loss of control over advanced AI systems. We find the null poorly supported for H1 and H2, and identify H3 as a priority for further work rather than a settled finding. We also identify five intermediate variables that recur across the pathways - information environment quality, decision-making timeline compression, great power threat perception, capability diffusion, and norm erosion - and argue that these shared nodes are the highest-priority targets for further investigation and intervention.
Introduction
The paper evaluates whether non-great-power conflict (NGPC) is within an order of magnitude of great-power conflict (GPC) as a source of catastrophic risk through three pathways. It finds the null poorly supported for H1 and H2, while treating H3 as a priority for further work rather than a settled finding.
- Research approach: The paper assesses whether NGPC increases GPC likelihood, catastrophic-terrorism harm, and loss-of-control harm through three sub-hypotheses.The analysis uses causal models, parameterized decompositions where evidence permits, and qualitative assessment of resulting parameters.
- Cross-cutting variables: The shared intermediate variables are information environment quality, decision-making timeline compression, great power threat perception, capability diffusion, and norm erosion.Because these variables recur across pathways, interventions shifting them could affect multiple risk pathways at once; interaction effects between pathways are not modeled.
- H1: NGPC and GPC: Across all three pathways, increasing NGPC is likely to increase the probability of GPC as AI capabilities advance.The pathways involve proxy escalation and alliance entrapment, strategic-environment deterioration, and AI capability testing and arms-race acceleration.
- H2: NGPC and catastrophic terrorism: Increasing NGPC is likely or probable to increase expected catastrophic-terrorism harm, driven primarily by capability transfer from conflict zones and capable AI or AI-assisted planning.Capability transfer is unlikely to increase attempt rate but likely or probably increases success rates near term, while capable agentic AI and AI-assisted planning likely or probably increase both attempt and success rates.
- H3: NGPC and AI loss of control: NGPC could potentially increase expected harm and the likelihood of loss-of-control events, but H3 remains the most speculative pathway.The proposed mechanism involves increasing the probability that misaligned or insufficiently controllable systems are deployed or given risky affordances when detection and mitigation are difficult or costly.
1 Methodology
The analysis builds causal models linking NGPC to each risk outcome, parameterizes some pathways where evidence permits, and qualitatively assesses others using literature, case studies, and abductive inference. It also identifies shared intermediate variables while treating probability judgments as directional, conditional assessments rather than precise forecasts.
- Methodology: The methodology constructs causal models, parameterized risk decompositions where evidence permits, and first-cut qualitative assessments of how advanced AI could alter relevant dynamics.The assessment draws on literature reviews, historical case studies, and abductive inference.
- Methodology: For great power conflict escalation and catastrophic terrorism, explicit risk decompositions multiply parameters such as attempt rate, success rate, and severity, with qualitative confidence assigned to each estimate.The models qualitatively assess how NGPC affects these parameters and are not intended to provide precise estimates.
- Methodology: For loss of control, speculative mechanisms are assessed as candidate causal pathways using plausibility, novelty, and severity rather than rough parameterization.The paper avoids parameterization where it would produce false precision.
- Methodology: Shared intermediate variables across sub-hypotheses include AI capability diffusion to nonstate actors, governance-norm erosion, and institutional-capacity decay in conflict-affected states.Variables recurring across multiple pathways are treated as higher-priority targets for investigation, ceteris paribus.
- Methodology: Probability assessments indicate confidence in the direction, not size, of parameter changes, and every pathway is conditional on relevant non-great-power or nonstate actors gaining capable AI access.The timing of that access is uncertain and is not forecast.
2 Great Power Conflict Escalation
NGPC can raise great-power-conflict risk through direct escalation into opposing patron involvement or indirectly through accumulated instability, but the Cold War record shows that high escalation pressure can be contained. Advanced AI may intensify involvement, miscalculation, decision-time compression, and information degradation, although its net effect remains uncertain and may be small.
- Direct escalation: NGPC can escalate directly when great powers become involved on opposing sides and escalation pressures exceed successful de-escalation.The sequential model applies to interstate NGPCs and civil conflicts attracting external intervention.
- Indirect escalation: The indirect pathway treats NGPC as a multiplier on existing GPC risk by shifting background conditions that destabilize great-power relations.Model B risk can accumulate across many individually low-risk conflicts through ratcheting instability effects, unlike Model A risk from a specific conflict.
- Cross-cutting implications: Shared causal nodes affecting multiple pathways are higher-priority targets for investigation and intervention.The paper identifies these shared nodes as recurring across both direct and indirect models.
- Historical evidence: The Cold War produced dozens of proxy conflicts but zero proxy-to-GPC escalations, indicating successful containment rather than negligible escalation pressure.Proxy relationships can also serve as de-escalation tools, and Russia’s Donbas experience suggests abandoning proxies, rather than using them, produced escalation with NATO-supported forces.
- AI effects: AI may increase great-power involvement and escalation through autonomous misidentification, unauthorized engagement, deeper patron-proxy ties, disinformation, and compressed decision-making timelines.The assessed direction is upward and the pathway is somewhat likely to increase GPC probability, but the expected increase may be small because deniability, de-escalation, and AI-enabled crisis management could offset these pressures.
- AI effects: The indirect pathway is somewhat likely to increase political-strategic instability as patron realignment and disinformation degrade information environments, with future AI-generated synthetic media increasing attribution difficulties.Resilient norms, content-provenance standards, and clearer spheres of influence could dampen the effect, leaving its magnitude uncertain.
3 Catastrophic Terrorism
NGPC concentrates the pathways through which nonstate actors can acquire operational skills and capabilities, and conflict environments are associated with substantially more frequent and deadlier terrorism. However, implementation barriers still constrain catastrophic attacks, while AI and future agentic systems may progressively narrow the gap between planning and execution.
- Empirical basis: NGPC conflict environments contain roughly nine in ten terrorist attacks and 98 percent of terrorism deaths, with attacks five to six times deadlier than in peaceful countries.These environments also expose nonstate actors to foreign fighters, proxy capability transfers, and limited oversight.
- Constraining factors: Implementation barriers remain decisive because groups still need resources, organizational capacity, materials, and the ability to execute attacks without detection.Historically, freely available bomb-making information did not produce a surge in sophisticated attacks because execution barriers dominated.
- Capability transfer: As AI capabilities flow through conflict-zone transfer pathways, pairing them with battlefield-tested personnel could substantially raise potential damages, including through AI-coordinated drone swarms.The pathway is assessed with moderate confidence because transfer mechanisms are observed, but AI capability transfer itself remains inferential.
- AI-assisted planning: AI-assisted planning makes increases in attempt and success rates realistic possibilities in the near term, but severity remains unlikely because human execution bottlenecks persist.AI lowers the knowledge barrier through interactive guidance, while current evidence provides limited support that it has substantially improved outcomes beyond traditional methods.
- Agentic systems: Future agentic systems could make increases in attempt rate and success rate likely or probable, while increased severity remains a realistic possibility because deployment is least developed.Autonomous coordination of reconnaissance, procurement, and deployment could reduce human bottlenecks, but the gap between laboratory demonstrations and adversarial real-world use remains substantial.
4 Loss of Control
NGPC plausibly increases expected harm from loss of control by promoting risky AI deployment and affordances while making detection and mitigation more difficult or costly. This remains the most speculative sub-hypothesis, so the paper identifies causal pathways qualitatively rather than parameterizing the risk.
- Overview: The loss-of-control sub-hypothesis is the most speculative because its prospective mechanisms have little direct historical or empirical support.The paper assesses these causal pathways qualitatively and does not attempt to parameterize the risk.
- Overview: NGPC increases expected loss-of-control harm by making misaligned AI systems more likely to be deployed or given risky affordances when detection and mitigation are difficult or costly.The paper consolidates this claim into three pathways: deployment, risky affordances, and difficult or costly detection and mitigation.
- H3.1: Deployment: Prolonged, unstable NGPC can increase time pressure, uncertainty, delay costs, and incentives to deploy imperfectly aligned AI in critical infrastructure.Deployment is defined as integrating AI into military command, logistics, and intelligence; the pathway is described as plausible and likely.
- H3.2: Risky affordances: Risky affordances can increase loss-of-control likelihood by giving misaligned systems more information, power, and autonomy, especially in fragile states experiencing prolonged conflict.Conflict diverts resources, personnel, political attention, and coordination capacity, making AI attractive for administrative, intelligence, decision-making, and coordination roles.
- H3.3: Detection and mitigation: NGPC may make detection and mitigation harder or costlier by accelerating path-dependent decisions and encouraging risky, irreversible action under perceived first-move advantages.These perceptions are expected in fragmented, uncertain NGPC environments that serve as testing grounds for new military AI technology.
- H3.3: Detection and mitigation: Multi-agent failure risks are strongest in proxy, regional multi-actor, and multipolar conflicts, but may not arise in siloed or limited-context deployments.The paper’s key uncertainties include alignment-capability tradeoffs, conflict decision priorities, institutional effects, and the prominence of perceived early AI military advantage.
5 Conclusion
The paper finds that NGPC materially affects mechanisms across all three catastrophic-risk vectors, but does not establish that its overall risk is within an order of magnitude of GPC. H1 has the strongest evidence, H2 has clear but magnitude-uncertain pathways, and H3 remains speculative and merits further investigation.
- Overall conclusion: NGPC materially affects mechanisms behind escalation to GPC, catastrophic terrorism, and loss of control over advanced AI, without yielding a precise NGPC-to-GPC risk ratio.The analysis uses causal models, literature reviews, and qualitative assessments, while leaving the order-of-magnitude comparison unresolved.
- Great-power conflict escalation: H1 has the strongest evidence: NGPC may escalate GPC through opposing patronage relationships, degraded stability, proxy entanglement, patron realignment, and military-AI testing grounds.De-escalation mechanisms can manage escalation pressure but do not eliminate the identified pathways.
- Catastrophic terrorism: H2 has clear causal pathways but uncertain magnitude, as conflict zones may provide tacit operational knowledge, lower planning barriers, improve coordination, and enable capability transfer.These mechanisms are expected to increase some combination of terrorist-attack attempt rates and success rates.
- Loss of control: H3 is the most speculative hypothesis: NGPC may increase pressure to deploy misaligned AI, incentives to grant risky affordances, and difficulty or cost of detecting and mitigating misaligned behavior.Each causal link remains uncertain, though the pathways may be empirically testable in the future.
- Limitations and contribution: The analysis is preliminary and identifies no negligible pathway, but it neither demonstrates that NGPC rivals GPC nor rigorously estimates interaction effects or overall risk.The paper concludes that the question remains open, the mechanisms are identifiable and sometimes observable, and further empirical and tractability work is warranted.
6 Appendix
The appendix maps five shared intermediate variables across the paper’s three pathways: information environment quality, decision-making timeline compression, great power threat perception, capability diffusion, and norm erosion. These shared nodes are identified as priority targets because interventions could affect multiple risk pathways simultaneously.
- Cross-cutting intermediate variables: Information environment quality affects de-escalation and baseline trust, while disinformation and synthetic media degrade both crisis management and great-power coordination.Degraded information environments reduce the ability to negotiate off-ramps and build trust during crises.
- Cross-cutting intermediate variables: Decision-making timeline compression increases escalation risk because autonomous systems and AI-assisted targeting reduce the time available for human judgment.AI-driven speed compresses tactical timelines within conflicts and strategic timelines during great-power crises.
- Cross-cutting intermediate variables: Great-power threat perception, capability diffusion, and norm erosion respectively drive intervention and instability, shift military balances, and lower thresholds for proxy involvement and escalation.Capability transfer can increase NGPC rates when it reaches nonstate actors and increase instability when it reaches rival great powers.
- Cross-cutting intermediate variables: The five shared variables appear across all three pathways and are the highest-priority targets for further investigation because interventions could shift multiple risk pathways at once.The appendix explicitly identifies information environment quality, decision-making timeline compression, great-power threat perception, capability diffusion, and norm erosion as shared nodes.