Promoting Stability on the Path to Superintelligence
Mutual restraint and resilience could be an alternative to threats of preventative war.
This essay is part of a project from the RAND Center for the Geopolitics of AGI that considers threats to stability in the race toward AGI. In this essay, expert Michael Mazarr proposes an approach that relies on mutual restraint, assurance, and resilience. We welcome your feedback.
As the world confronts the prospect of AI superintelligence—highly advanced AI capable of performing thousands of functions as well as or in some cases dramatically better than humans—the question of strategic stability during the transition period is becoming a real concern. If either the United States or China thinks that the other is on the verge of AI superintelligence, it may fear being subjugated in economic, technological, and military terms, or that superintelligence once unleashed will pose myriad threats to humanity. Great powers facing such threats have sometimes lashed out in preventive wars.
The obvious risk in such a power-rearranging transition is that the race to superintelligence could prompt instability and conflict. Probably the most widely-discussed diagnosis of and prescription for this challenge is the concept of “mutual assured AI malfunction,” or MAIM, proposed in the 2025 report Superintelligence Strategy by Dan Hendrycks, Eric Schmidt, and Alexander Wang. (Hendrycks and Adam Khoja later added a response to critics.) They argue that superintelligence could pose multiple dangers. “In the hands of state actors,” they argue, “it can lead to disruptive military capabilities and transform economic competition. At the same time, terrorists may exploit its dual-use nature to orchestrate attacks once within the exclusive domain of great powers. It could also slip free of human oversight.” As superintelligence looms on the horizon, great power relationships could become more volatile.
MAIM is compelling and sparked a necessary debate. But I’d like to propose an alternative to a world in which great powers threaten to attack each other to enforce controls on AI. My approach is grounded in a very different assessment of the dangers involved in the path to superintelligence. Rather than threats of preventive war, it relies on mutual restraint, assurance, and resilience.
The Difficulties with MAIM
To deal with the dangers of the superintelligence transition, Schmidt, Hendrycks and Wang offer a strategy with three components: Nonproliferation (preventing uncontrolled spread of AI capabilities such as advanced chips), competitiveness (ensuring that others don’t get to superintelligence first), and deterrence through the concept of MAIM. They term their overall approach the “Multipolar Strategy” because it embodies all three of these mutually supportive elements. But the linchpin of their approach, the core mechanism to avoid most risk, is MAIM.
With this strategy, they’re trying to control a complex and distinct set of risks. One is that “In the hands of state actors, it can lead to disruptive military capabilities and transform economic competition,” such as the creation of “superweapons” that could “enable military dominance.” A country “with sole possession of superintelligence,” they argue, “might be as overwhelming as the Conquistadors were to the Aztecs.” Such capabilities have the potential to destabilize balance of power, though the analysis doesn’t explain precisely how that would play out (except to claim that AI capabilities are, as a rule, dangerously offense-dominant).
A second risk is malign uses by non-state actors. Here they focus specifically on AI “as an amplifier of terrorist capabilities,” including bioterrorism and AI-empowered cyberattacks. The third risk is “losing control over the AI system itself”—classic loss of control or misalignment dangers of highly powerful AI that escapes human guardrails. They suggest MAIM as a way of controlling AI development and thus mitigating all three dangers.
MAIM reflects a sort of hyperactive Mutually Assured Destruction involving the relentless danger of preemptive action. They describe it this way:
If a rival state races toward a strategic monopoly, states will not sit by quietly. If the rival state loses control, survival is threatened; alternatively, if the rival state retains control and the AI is powerful, survival is threatened. A rival with a vastly more powerful AI would amount to a severe national security emergency, so superpowers will not accept a large disadvantage in AI capabilities. Rather than wait for a rival to weaponize a superintelligence against them, states will act to disable threatening AI projects.
Under the concept of MAIM, “any state’s aggressive bid for unilateral AI dominance is met with preventive sabotage by its rivals.” States could try to kneecap an opponent’s AI progress in various ways, “ranging from covert operations that degrade training runs to physical damage that disables AI infrastructure.” The result, they argue, would be “a dynamic similar to nuclear Mutual Assured Destruction (MAD), in which no power dares attempt an outright grab for strategic monopoly, as any such effort would invite a debilitating response.”
This is a useful and thought-provoking idea, and it has sparked critical dialogue on the issue of stability on the road to superintelligence. But in an earlier essay, two RAND colleagues and I pointed out several major shortcomings. The basic threat involved in MAIM just isn’t credible: Neither country has an ability short of the use of nuclear weapons to hit hundreds of AI-related targets in the other’s homeland. Then, too, it’s hard to imagine that intelligence on the other side’s AI progress would be reliable enough to justify starting wars, in part because there’s no obvious threshold at which to launch an attack.
The threat of such attacks likely won’t deter AI progress, especially because any side which can race to superintelligence before it gets hit will be able to prevent the other side from launching a preemptive attack, basically creating a veto power on the other’s MAIM strategy. Worst of all, the doctrine would be profoundly destabilizing, locking in a mutual threat to start a war over ambiguous conditions. As we argued in our response, because “it will be so difficult to know when the critical moment of danger has arrived in a rival’s development of advanced AI, a MAIM-like balance could be highly unstable. … Each side would be scrutinizing the others’ AI development with an intensity verging on paranoia, and the potential for misperception would become very great. … The result could be a hair-trigger balance of AI terror.”
The Risks We’re Trying to Control
Yet the strategic risks and uncertainties on the path to superintelligence could be very real, and so we need some form of strategy to deal with them. An alternative can be grounded in a less binary conception of those risks from that proposed in the MAIM concept, which sees dangers of instability from fears of a sudden, rapidly destabilizing transition to superintelligence. If instead we assume that the transition is not a singular event but many accumulating steps, that it will be drawn out rather than a single event or threshold, and that it will be jagged rather than uniform in the power it conveys, we can envision a very different strategy for dealing with it.
My proposal makes a key distinction: between rapidly emerging, binary risks and gradually accumulating ones. Contrary to some arguments about strategic stability, I’d argue that the dangers of superweapons that effectively disarm other nations are not so great. MAIM envisions a “grab for strategic monopoly,” a sort of potent leap for superintelligence on a massive scale. But superintelligence will emerge from a thousand incremental steps rather than a few massive and easily recognizable ones, probably over a longer time frame than some advocates have suggested. In its focus on a sudden, rapid transition, MAIM is describing a world that’s not likely to exist. If superintelligence ends up being a gradually emergent phenomenon rather than a stark threshold, no single moment will offer a potential to impose absolute hegemony.
Nor will superintelligence arrive in all domains at the same time. It is more likely to emerge very unevenly, across different sectors of the economy and social domains. Neither side will confront a comprehensive advance that gives the other an absolute advantage. Some areas, like biology, broader scientific discovery and strategic decision making, are too causally complex, interactive, and stochastic to allow linear effects from hugely intelligent AI. Even in domains where it can offer real gains, many advantages conveyed by advanced AI will probably take years to work themselves into tangible advantage—for example through scientific discoveries that need to be validated or manufactured goods that need to be produced. The stability risks of the path to superintelligence will be eased by the gradualism of its emergence.
Finally, there’s growing evidence that the road to superintelligence is likely to be increasingly diverse rather than monolithic. New approaches to model training beyond transformer-based LLMs are likely to be needed to get to true superintelligence. Compute power will come as much from complex architectures of linked semiconductors, ecosystems of computing power, as from the power of the chips themselves. A more heterogenous pathway to superintelligence will make a jagged, uneven emergence of capabilities even more likely.
Of course, in any strategic balance, perceptions will be at least as important as reality. If one or both sides believe that advantages will arrive quickly and be more binary than I have argued, they might be tempted to act in destabilizing ways, trying to undermine the other’s emerging AI ecosystem as MAIM would predict.
So far, despite the breakneck pace of AI investment in both the United States and China, neither side seems to be thinking this way. U.S. officials and AI leaders assume they are ahead and can remain there. Most reports on Chinese official perceptions agree that for now, Chinese experts do not perceive an approaching moment of profound strategic disadvantage and view the competition as a more gradual affair and one in which application and diffusion are as important as outright model capability and open source models, while less powerful, have important long-term advantages in capturing global AI network power.
This could all change if a single, profound threshold emerged on the road to superintelligence, something that forced one or both sides to confront the possibility of strategic defeat. But the objective reality of a jagged, gradual, diverse frontier of superintelligence is likely to be apparent enough to leaders on all sides of the race that it will calm the temptation to take violent action—especially because MAIM-like attacks are so impractical and profoundly risky.
Even on a gradual path, however, we are likely to begin seeing—possibly within the next several months—the first of a long chain of AI-empowered threat moments, actions that highlight the growing dangers of advanced AI. Some of these will be hostile use of AI by dangerous individuals or groups: Inventing a lethal pathogen and trying to manufacture it; using AI-magnified cyber capabilities to attack critical infrastructure; manipulating chatbots to threaten vulnerable populations. Others could be loss of control events in which AI models spontaneously cause harm in various ways. Some versions of either of these categories of threat could be profoundly dangerous—far more than mere nuisances, threats that pose serious risk to the American homeland. (In fact, some see the recent events surrounding the AI collaboration realm Moltbook as a harbinger of serious risks to come.)
It could well be, then, that the real danger as the path to superintelligence becomes clearer is not a sudden, comprehensive power shift that feeds the risk of preemptive strikes, but an accumulating set of serious but discrete perils generated by increasingly powerful AI. Few of the intentional malign uses of AI might originate from other great powers; none of them would create the possibility of a comprehensive strategic hegemony by any one country.
Elements of a Strategy to Deal with the Transition
That analysis suggests that a strategy for dealing with the transition to superintelligence can be less about avoiding Conquistador moments or mitigating temptations for preventive war and more about building resistance against gradually emergent and uneven threats—threats that do endanger America’s and China’s security, but not in ways that can be remedied with mutual AI decapitation strikes. A standoff characterized by brinksmanship, intimidation and enforced limits is less appropriate to this challenge than resilience, reassurance, and staying even in AI progress.
At this point I’m not trying to offer a comprehensive strategy to deal with the risks of the superintelligence transition. A wide range of work is underway at RAND and in other places looking at this issue. My goal here is to contribute to that thinking by offering one tentative idea for an alternative strategy and defining its constituent parts. A transitional risk mitigation strategy could be based on four elements:
Restraint and reassurance. In a mirror image of the MAIM idea, the United States and China could make formal commitments not to preempt the other’s AI progress. The United States could enshrine this as a formal doctrine and offer transparency and public assurances about its posture. Such promises alone would hardly sweep away all concerns, partly because there is too little trust between the United States and China to make promises strong enough to solve the problem on their own. But they would create a baseline set of pledges on which to build further confidence building measures. This step could include coordination on specific risks both would fear, such as terrorist use of AI to generate biological weapons. It could also involve demonstrated restraint in the use of new AI-fueled capabilities to avoid the impression that the United States will employ them to subjugate others.
Intelligence and awareness. A second component of the strategy would seek to guarantee maximum possible knowledge of AI programs and progress in China and beyond. It would also build a basis of reporting and sensing to identify potential loss of control and misuse risks, be able to attribute attacks and understand which elements of the jagged frontier of superintelligence are moving more quickly. As AI capabilities advance, intelligence on their movement and risks associated with them will become more important than reporting on potential adversary militaries.
Continued investment in AI development and application to avoid falling decisively behind. As much as progress even at the threshold of superintelligence is likely to remain gradual, the United States has to ensure that it does not surrender a significant lead to other AI actors. That seems highly unlikely given current and prospective AI research and development but ensuring the outcome must remain a U.S. priority. It can help the United States avoid the kinds of fears that would prompt desperate actions to protect its security. This component of the strategy is not only about AI model capabilities but also broader investments in the societal requirements for success in the AI era.
Broad-based societal resilience against AI-generated risk. Finally and most importantly, as MAIM is the fulcrum of the Schmidt, Hendrycks and Wang strategy, resilience is the centerpiece of this one. Part of the resilience agenda could be focused on ensuring that no one can launch a preemptive strike against the U.S. AI ecosystem. Steps to achieve this can include hiding, concealing and dispersing elements of an AI architecture; providing active defense against virtual or physical attack; hardening targets; and increasing redundancy and resilience. More broadly, the United States could pursue a general society-wide program of resilience in many forms to guard against the risks posed by superintelligence: Critical infrastructure protection, a radically higher degree of cybersecurity, societal and population training, disaster recovery capabilities, preparations for operating in degraded modes, and a broad-based approach to AI alignment and control.
Part of the goal with such resilience is to create an analogue to the true source of Cold War-era nuclear peace: Guaranteed retaliation. MAIM inverts the true logic of Cold War MAD. That strategic reality was based on the inability of either side to launch a first strike, not the institutionalization of such strikes. The “mutual destruction” element came from the fact that both sides maintained a secure second strike, not a constant threat of preemption. To be sure, both sides sought to escape this reality to some degree, toying with first strike-capable nuclear systems and warfighting doctrines. But senior political leaders always recognized the futility of large-scale nuclear strikes. The goal is to preserve an ability to retaliate across a wide range of tools so that the attacker could not envision getting away with a superintelligence-spurred attack.
This strategy addresses the risk of perception-driven instability and fears of sudden power shifts but does so in different ways than MAIM. It uses commitments for restraint, broader assurances and confidence building measures, and strengthened intelligence and awareness to reduce the perceived danger of preventive war, and investments in AI equivalence and resilience to increase mutual confidence that each side would not be vulnerable to a sudden, superintelligence-driven effort to impose hegemony. Even if moments do arise that feel more like a brink than a gradual path, the strategy has shock absorbers to avoid instability.
We could call the resulting strategy Resilient Equilibrium (or as a shorthand, R-EQ). I asked ChatGPT for a summary of its basic elements. This approach, it noted, rests on the judgment that transitional risks don’t emerge so much from a mutual temptation to destroy the other’s AI ecosystem but from a more gradually accumulating series of discrete dangers. Given this,
Strategies centered on brinkmanship, enforced limits, or threats of disruption risk amplifying those dangers by encouraging secrecy, arms racing, and crisis instability. By contrast, Resilient Equilibrium seeks to dampen escalation pressures by combining restraint and reassurance with superior intelligence and sustained investment in AI research and application sufficient to avoid decisive technological disadvantage. Its central premise is that confidence—confidence in awareness, in relative technological position, and in societal robustness—reduces the incentives for preventive war more effectively than coercion. Rather than attempting to freeze progress or deter through mutual vulnerability, Resilient Equilibrium prioritizes resilience: hardening critical systems, dispersing and defending AI infrastructure, and preparing society to operate through disruption. Stability in a superintelligence transition, under this approach, is achieved not by threatening catastrophe, but by making catastrophe strategically unnecessary.Nicely said. That sort of a strategy—ensuring world-class AI capabilities and awareness of others’ progress while promoting restraint and reassurance alongside a comprehensive new effort at national resilience—is the way to deal with the risks of the transition to superintelligence. It’s appropriate to the nature of the threat and a lot more feasible, and less provocative, than MAIM. Broadly speaking, it’s the direction we should head as we begin to contemplate those final years and months on the road to extremely powerful AI.