From AI Scenarios, December 3:
Abstract
We model national strategies and geopolitical outcomes under differing assumptions about AI development. We put particular focus on scenarios with rapid progress that enables highly automated AI R&D and provides substantial military capabilities.
Under non-cooperative assumptions—concretely, if international coordination mechanisms capable of preventing the development of dangerous AI capabilities are not established—superpowers are likely to engage in a race for AI systems offering an overwhelming strategic advantage over all other actors.
If such systems prove feasible, this dynamic leads to one of three outcomes:
- One superpower achieves an unchallengeable global dominance;
- Trailing superpowers facing imminent defeat launch a preventive or preemptive attack, sparking conflict among major powers;
- Loss-of-control of powerful AI systems leads to catastrophic outcomes such as human extinction.
Middle powers, lacking both the muscle to compete in an AI race and to deter AI development through unilateral pressure, find their security entirely dependent on factors outside their control: a superpower must prevail in the race without triggering devastating conflict, successfully navigate loss-of-control risks, and subsequently respect the middle power's sovereignty despite possessing overwhelming power to do otherwise.
Executive summary
We model how AI development will shape national strategies and geopolitical outcomes, assuming that dangerous AI development is not prevented through international coordination mechanisms. We put particular focus on scenarios with rapid progress that enables highly automated AI R&D and provides substantial military capabilities.
Race to artificial superintelligence
If the key bottlenecks of AI R&D are automated, a single factor will be driving the advancement of all strategically relevant capabilities: the proficiency of an actor's strongest AI at AI R&D. This can be translated into overwhelming military capabilities.
As a result, if international coordination mechanisms capable of preventing the development of dangerous AI capabilities are not established, superpowers are likely to engage in a race to artificial superintelligence (ASI), attempting to be the first to develop AI sufficiently advanced to offer them a decisive strategic advantage over all other actors.
This naturally leads to one of two outcomes: either the "winner" of the AI race achieves permanent global dominance, or it loses control of its AI systems leading to humanity's extinction or its permanent disempowerment.
In this race, lagging actors are unlikely to stand by and watch as the leader gains a rapidly widening advantage. If AI progress turns out to be easily predictable, or if the leader in the race fails to thoroughly obfuscate the state of their AI program, at some point it will become clear to laggards that they are going to lose and they have one last chance to prevent the leader from achieving permanent global dominance.
This produces one more likely outcome: one of the laggards in the AI race launches a preventive or preemptive attack aimed at disrupting the leader's AI program, sparking a highly destructive major power war.
Middle power strategies
Middle powers generally lack the muscle to compete in an AI race and to deter AI development through unilateral pressure.
While there are some exceptions, none can robustly deter superpowers from participating in an AI race. Some actors, like Taiwan, the Netherlands, and South Korea, possess critical roles in the AI supply chain; they could delay AI programs by denying them access to the resources required to perform AI R&D. However, superpowers are likely to develop domestic supply chains in a handful of years.
Some middle powers hold significant nuclear arsenals, and could use them to deter dangerous AI development if they were sufficiently concerned. However, any nuclear redlines that can be imposed on uncooperative actors would necessarily be both hazy and terminal (as opposed to incremental), rendering the resulting deterrence exceedingly shaky.
Middle powers in this predicament may resort to a strategy we call Vassal's Wager: allying with one superpower in the hopes that they "win" the ASI race. However, with this strategy, a middle power would surrender most of their agency and wager their national security on factors beyond their control. In order for this to work out in a middle power's favor, the superpower "patron" must simultaneously be the first to achieve overwhelming AI capabilities, avert loss-of-control risks, and avoid war with their rivals.
Even if all of this were to go right, there would be no guarantee that the winning superpower would respect the middle power's sovereignty. In this scenario, the "vassals" would have absolutely no recourse against any actions taken by an ASI-wielding superpower.
Risks from weaker AI
We consider the cases in which AI progress plateaus before reaching capability levels that could determine the course of a conflict between superpowers or escape human control. While we are unable to offer detailed forecasts for this scenario, we point out several risks:
- Weaker AI may enable new disruptive military capabilities (including capabilities that break mutual assured destruction);
- Widespread automation may lead to extreme concentration of power as unemployment reaches unprecedented levels;
- Persuasive AI systems may produce micro-targeted manipulative media at a massive scale.
Being a democracy or a middle power puts an actor at increased risk from these factors. Democracies are particularly vulnerable to large scale manipulation by AI systems, as this could undermine public discourse. Additionally, extreme concentration of power is antithetical to their values.
Middle powers are also especially vulnerable to automation. The companies currently driving the frontier of AI progress are based in superpower jurisdictions. If this trend continues, and large parts of the economy of middle powers are automated by these companies, middle powers will lose significant diplomatic leverage....
....MUCH MORE, that's just the summary.
Here's the version at the Social Science Research Network:
Modeling the Geopolitics of AI Development23 Pages
Posted: 3 Dec 2025
