Below is Part IV of a five-part essay that I’m rolling out on NZN. (Part I is here, Part II is here, Part III is here, and all the parts are accumulating here.) The essay is a kind of call to action—an argument about what we need to do to successfully navigate the AI revolution and why we need to do it. If you’d like to help me get this message out, please share (and/or like) my tweet about it or my bluesky post about it—or use some other means of dissemination, like Facebook or email or actual in-person conversation with a flesh-and-blood human being.
Part IV: International and Intranational Destabilization
It’s common to break the AI safety issue down into distinct kinds of threats: bad actors using AI to build a bioweapon, bad actors using AI as a devastating cyberweapon, AI going rogue and wreaking havoc, and so on. But there’s a broader and in some ways deeper kind of AI danger—a danger that, depending on how exactly it took shape, might involve one or more of these threats but might not, and in any event would have other key ingredients as well. This is the danger of catastrophic destabilization—of multiple consequences of AI pushing a nation, or a whole international region, or even the whole world, beyond the edge of order.
Let’s start with international destabilization—with some of the ways AI could increase the chances of conflict among nations.
When people envision this technology being used to create a bioweapon or conduct a cyberattack, the destructive potential seems clear enough: a pandemic could kill millions if not hundreds of millions of people, and a cyberattack could take out whole pillars of the planet’s electronic infrastructure. But these kinds of weapons also have a property that, though not so obviously destructive, could lead to great carnage. Namely: Their point of origin can be hard to pin down. Unlike, for example, ballistic missiles, they can be delivered without a return address.
This cloak of anonymity could make it more tempting for both nation-states and non-state actors to use these weapons. And, perhaps more important, it could give these weapons an incendiary potential that, through a kind of chain reaction, could draw other kinds of weaponry into an expanding realm of conflict.
When nations face an attack of unclear origin, they can be quick to point fingers—and a consensus can quickly form, on the basis of scanty evidence and pre-existing animosities, that the most feared adversary of the day must be responsible. This in turn can lead to counterattack (perhaps to “establish deterrence”) that leads to escalation. So the age of AI makes it easier for a non-state actor, or a small-state actor, to launch a cyber-attack or bio-attack that leads to nuclear war between the US and China.
This generic prospect—a major war growing out of an AI-abetted attack of unclear origin—is rendered only more plausible by the variety of threat vectors that AI can intensify. In principle, for example, regime change could be orchestrated remotely, via the deployment of millions of bots that seem to be human (and that may sometimes use the commandeered social media accounts of real humans) but have superhuman persuasive powers and use them to ignite a rebellion. Faced with such an AI-astroturfed rebellion, will a nation’s political leaders point fingers at a foreign adversary even if they’re not sure whether the threat has a foreign or domestic origin? Well, political leaders have certainly been known to point fingers abroad when facing domestic discontent. What’s more, they’ve been known to consolidate shaky domestic support by fomenting conflict, even war, with an outside power.
It’s hard to say whether this last threat vector—AI-astroturfed rebellion—will turn out to be a big thing. And this kind of uncertainty is part of the problem: We are entering a period of ultra-rapid advance in a technology that could well pose various new threats to peace and stability at both the national and international level—including threats we don’t clearly anticipate. The good news is that, however diverse and unpredictable the emerging threat portfolio may be, one generalization would seem to hold: Structures for expanded and deepened transparency—structures that give nations a clearer idea of what’s going on with AI abroad—can be stabilizing if well designed. Transparency can not only discourage misdirected counterattacks; it can discourage the initial attacks by making it harder for countries to prepare and execute them without fear of detection.
A common response to the growing threat of cyber-attack and bio-attack is to say that America should increase its “resilience”—strengthen cybersecurity in the private and public sector, build an infrastructure for rapid-response vaccine production, and so on. Resilience is indeed worth building, and it could make a big difference. How big a difference is hard to say. In the realm of cyberattack for example, its efficacy depends partly on the question of whether, as AI supercharges both the offensive and defensive arsenals, the balance of increased power winds up favoring defense in the long run. But in any event there are powerful reasons to doubt that enhanced resilience can bring true security in the absence of robust and innovative international governance.
First, there are the aforementioned new threat vectors, like AI-astroturfed rebellion. For every threat vector there are things you can do to build resilience, but the more vectors there are, the harder it is to find security in a broad sense through resilience alone—especially if the emergence of some vectors is hard to predict.
And, even if we confine the analysis to the two best-known threat vectors—cyber-attack and bio-attack—reliance on resilience faces a big problem. Given how fast AI is advancing, and how unpredictable specific bursts of progress are, and how widely the use of “open source” AI is distributing the power to customize AI to nefarious ends, it may not always matter whether, “in the long run,” AI progress “on balance” strengthens defense more than offense. A single moment of asymmetry between the offensive capabilities of a bad actor and the defensive capabilities of a nation could have massively destructive consequences. So a trajectory of AI progress that repeatedly and abruptly creates such asymmetries could be a source of recurring large-scale lethality and ongoing instability.
This problem—that the speed of AI advance can create moments of incendiary asymmetry—isn’t confined to specific narrow threat vectors. There is a generic version of the problem.
A number of influential Americans—including Anthropic CEO Dario Amodei—have argued that winning the race to “artificial superintelligence” is critical to US national security. The premise is that superintelligence will entail such broad and diverse power that the first nation to achieve it will have unquestioned global hegemony. In this scenario, the US either gains this dominance by winning the race to superintelligence or gets dominated by the other likely winner, China.
But, as Dan Hendrycks, Eric Schmidt, and Alexandr Wang noted in a 2025 paper called Superintelligence Strategy, a race between two superpowers toward superintelligence poses a risk: If both countries believe that reaching that threshold confers enduring dominance on the winner, there is a strong incentive for the would-be loser—the country that lags in the race as the finish line approaches—to pre-emptively sabotage the rival AI program. (The idea of a distinct “finish line” is of course too simple, but for purposes of this analysis, you don’t need to imagine the future being binary in that way; sufficiently rapid acceleration—a singularity-style curve of AI progress—could be enough to inspire existential fear in the lagging nation as the curve starts to “go vertical.”) The sabotage might come in the form of cyber-attack but might also, given the stakes, be kinetic—like, for example, bombing AI data centers or factories that make advanced AI chips.
In either event, and especially in the kinetic scenario, escalation toward nuclear war is imaginable. And the kinetic scenario is rendered all the more plausible by the fact that the only factories capable of producing the world’s most advanced AI chips are within easy striking range of China in territory that China considers its own—and China is currently denied access to those chips.
Hendrycks, Schmidt, and Wang expressed the hope that war between rival AI superpowers could be prevented by a kind of naturally developing deterrence (“mutually assured AI malfunction,” or MAIM). This hope was dismissed by some as unrealistic, and it indeed presupposes, among other things, not obviously plausible demonstrations of unilateral restraint by national leaders. Still, one premise of this hopeful scenario—that a nation leading the race to superintelligence would perceive the growing incentive of its rival to take drastic action—is reasonable. And this could in principle be the basis for a bi-national agreement that brought verifiable commitments of mutual restraint, accompanied by the formalized transparency required for such verification.
A headlong international AI race can bring destabilization not just among countries but within them. So long as the US sees winning a race with China as vital to national security, the race will tend to take priority over domestic regulation that might slow the advance of American AI. (As MIT physicist and AI researcher Max Tegmark has put it, American AI companies “have found that by far the easiest way to avoid being regulated is just to say this two-word sentence: ‘But China!’”) And, given the many realms of life where AI will have a dislocating or disorienting effect—jobs, education, politics, child-rearing, romance, and on and on—rapid AI advance could push American social and political life even further from equilibrium than it now is.
This prospect of domestic destabilization becomes all the more plausible in light of one goal AI companies share with social media companies: to maximize engagement. One way to keep people engaged with an information product is to give them information that feels good—including information that bolsters their belief in their rightness and the wrongness of their rivals and adversaries. So there is good reason to believe that AI, if shaped by market forces alone, will be tribalizing, intensifying divisions among American factions.
There’s an irony in the prospect of domestic destabilization via AI—via its collective impact along various economic, social, psychological, and political dimensions. Prominent among the premises underlying the US foreign policy establishment’s consensus that the US must “win” the AI race with China is the belief that China aims to impose its system of governance on other nations and would use increased international leverage to that end. It’s not clear that this premise is correct (it is so widely accepted as to be rarely examined), but it’s clear how it could backfire: If fear of China-imposed authoritarianism fuels an American sprint toward superintelligence that is essentially untrammeled by regulation, the result could be sharply increased domestic discontent, disorientation, and disorder. And history shows that this is fertile ground for authoritarianism. So authoritarian rule in America—the threat that a sprint toward superintelligence is intended to head off—could wind up being a product of that sprint.
In the scenario I’ve just painted, the destabilization of American society and politics would in some sense be the product of a racing dynamic between the US and China. But it’s worth underscoring how that dynamic works: It’s not that the race with China accelerates the pace of American AI development, pushing it beyond the pace that intense competition among American companies would by itself sustain. Rather, it’s that the race with China precludes the kind of regulation that might otherwise slow the advance of AI. So, by the lights of this analysis, the “natural” rate of AI progress—the rate that American capitalism produces even in the absence of international adversaries—poses a threat to the stability of the United States.
This threat—that the nation could dissolve into violent disarray—may be the most underappreciated of the threats AI poses to the United States (and to other countries). To be sure, many of the individual elements that together constitute the threat get a fair amount of attention. There is broad and growing awareness of the possibility of massive job displacement. And the challenges AI poses to education and to child rearing and to the mental health of vulnerable adults are also much discussed. There is even some discussion, if less, of the way ambitious actors could use armies of AI agents to shape our opinions and beliefs and allegiances, creating new religious cults or ideological cults or other tribes with malign potential. And there is some awareness that radically new kinds of scams and cybercrimes are taking shape. Meanwhile, the Hugging Face incident has turned the dangers of rogue AIs into a common discussion topic. But there is still little discussion of the fact that the emergence of these and other challenges within a short time frame—like, the next two or three years—might overwhelm our capacity for social and political adaptation, carrying a whole nation off the rails.
Yet history tells us that, though societies can adapt to technological change, adaptation takes time, and in the meanwhile there is dislocation. And in all of history there has never been as much dislocation, in as short a period of time, as seems to be in the offing.
The prospect of a nation collapsing into chaos is, on the face of it, a national problem, especially when the problem emanates from the nation’s own economy—from technological progress that is driven by companies residing within the nation and is intensified by competition among those companies. Yet even when the problem is in this sense national, national policy alone can’t solve it. No significant slowdown of AI progress is politically feasible unless it is internationally coordinated—involving, at a minimum, both the US and China. Whether the problem is international destabilization or national destabilization, a big part of the solution must lie in international governance.
Coming next week: Part V: The Profound Challenge of International Governance. All parts of the essay accumulate here.
Meanwhile: If you want to check out my book on AI, The God Test, you can read the introductory chapter and excerpts from all other chapters at thegodtest.net.
Banners and graphics by Clark McGillis.