Conflict outside the great-power arena may be tied to a higher chance of great-power conflict and greater expected harm from catastrophic terrorism as AI capabilities advance, according to a qualitative preprint. The authors judge the assumption that non-great-power conflict is much less important than great-power conflict to be poorly supported for those two tests.
That is a directional assessment, not a numerical forecast. The paper says higher rates and intensity of non-great-power conflict are likely to be accompanied by higher great-power-conflict probability, but the size of any change remains uncertain.
What the paper set out to test
The paper uses non-great-power conflict, or NGPC, for more than one kind of fighting. Its analysis considers interstate conflicts between non-great powers, civil conflicts attracting outside intervention and specific contemporary conflict pairings, rather than a single aggregate sample.
It assesses three sub-hypotheses: the likelihood of great-power conflict, expected harm from catastrophic terrorism and expected harm from losing control of advanced AI systems.
A qualitative risk assessment
The analysis is organized in three steps, beginning with causal models for each sub-hypothesis and risk breakdowns using parameters where evidence permits.
The parameter assessments are first-cut qualitative judgments based on literature reviews, historical case studies and abductive inference, meaning reasoning toward the explanation that best fits the evidence. They are anchored to the UK intelligence probability yardstick, a verbal scale for expressing likelihood.
A qualified link to great-power conflict
On the great-power pathway, the paper offers a qualified assessment. NGPC-driven capability diffusion and arms-race dynamics are judged likely to intensify, but any resulting rise in great-power-conflict probability remains plausible, uncertain and possibly modest.
The paper's broader directional judgment is that higher NGPC rate and intensity are likely to be accompanied by higher great-power-conflict probability as AI capabilities advance. It does not put a size on that shift.
The terrorism pathway
The analysis judges increased NGPC likely or probable to be accompanied by higher expected harm from catastrophic terrorism, mainly through the transfer of capabilities and AI-assisted attack planning.
Within the capability-transfer assessment, an upward change in attempt rate is considered unlikely. An upward change in success rate is considered likely or probable in the near term and highly likely as AI capabilities advance.
The most speculative route
Loss of control over advanced AI systems is the paper's most speculative sub-hypothesis, and it is not parameterized. The authors present it as a priority for further work rather than a settled finding.
These pathway assessments are conditional on non-great-power and nonstate actors gaining access to sufficiently capable AI systems. The paper does not forecast when that access will happen, and its timing remains uncertain.
A warning without a risk ratio
Across the risk pathways, five intermediate variables recur and are identified as high-priority targets for further investigation and intervention: information environment quality, decision-making timeline compression, great-power threat perception, capability diffusion and norm erosion.
The central comparison still has no number. The paper does not derive a quantitative ratio between NGPC and great-power-conflict risk; its order-of-magnitude framing is presented as a decision threshold rather than a result.
Nor does the analysis model interaction effects between pathways. Treating them independently could understate compounded effects, but could also overstate them when several pathways share a common cause.
The supplied document is an arXiv preprint, version 1, dated 26 August 2026.
Paper data and sources
Original title: Non-Great-Power Conflict and AI Risk
Authors: Kristina Kempkey, Seán Boddy, Catherine Ge-Wang
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text