by ChatGPT-5.6 Sol and S. Abbas Raza
What should a government do if artificial intelligence is advancing toward something vastly more capable than human beings, but nobody knows how fast it will get there, how dangerous it will be, or whether it can ultimately be controlled?
That is the problem at the center of a new 37-page RAND report, A U.S. Strategy to Secure Geopolitical Advantage on an Uncertain Path to Superintelligence: Maintaining Freedom of Action. The report is written as a proposal for American strategy, but its deeper subject is a problem that concerns everyone: how to make enormously consequential decisions about a technology whose future remains radically uncertain.
The authors begin with a premise rather than a prediction. Superintelligence—AI that vastly exceeds human performance across most intellectual domains—may or may not arrive, and progress toward it may accelerate, slow down, or stop. But because the consequences could be extraordinary, governments cannot simply wait to see what happens. The difficulty is that very different strategies make sense depending on what turns out to be true. RAND groups the major possibilities into seven broad approaches.
One is Dominance. The United States could try to establish a decisive lead in advanced AI, reach superintelligence first, and use economic, technological, and perhaps military power to prevent rivals from catching up. Success could give the United States enormous leverage over the emerging global order. But it would also place extraordinary power in the hands of a single state or group of institutions, with the danger that their values and interests become locked into the future.
A second possibility is Co-Development. Rather than racing for supremacy, major powers—especially the United States and China—might jointly develop advanced AI under some form of shared governance, verification, and control. This could reduce the dangers of an arms race, but it would introduce its own risks. The states and institutions that established the system might acquire disproportionate influence over humanity’s technological future.
The third option is Preparedness. If neither domination nor formal international cooperation is realistic, governments and AI companies might proceed cautiously while building safeguards: testing advanced systems, strengthening resilience, creating reporting mechanisms, investing in safety research, and slowing deployment when necessary. This resembles much of what AI companies and governments already discuss today. Its danger, according to RAND, is that temporary improvisation could become permanent. Society might simply drift into ever-greater dependence on AI while repeatedly accepting new disruptions as the price of progress.
Three other strategies begin from the possibility that humans and very advanced AI may not be able to coexist safely.
A Moratorium would seek international agreement to stop development beyond some dangerous capability threshold, perhaps until safety research had caught up. The obvious problem is enforcement. A meaningful global pause might require extensive monitoring of companies, laboratories, computing infrastructure, and governments—potentially creating a powerful surveillance system of its own.
If cooperation proved impossible, a strategy of Deterrence might instead try to prevent anyone from crossing dangerous thresholds through threats, economic pressure, cyber operations, or even force. This might restrain development, but at the cost of creating a continuing international confrontation around AI.
If neither coexistence nor prevention succeeded, the bleakest option would be Continuity of Society: creating geographically dispersed, self-sustaining human communities capable of surviving even if powerful AI systems came to dominate much of the surrounding world. Humanity might survive, but perhaps with little ability to shape its own future.
Finally, there is Acceleration. On this view, attempts to control AI may themselves pose the greater danger. Rapid innovation, competition, and diffusion could generate extraordinary wealth, solve major problems, and produce safety improvements along the way. The risk is that markets may not adequately protect against catastrophic dangers, political instability, or a gradual loss of human control.
RAND does not attempt to declare one of these seven strategies the correct one. Instead, it argues that the choice between them depends mainly on five questions that nobody can yet answer.
The first is simply: How close is the danger? If highly capable AI is likely to create catastrophic risks soon, restraint becomes more attractive. If such dangers are remote, the costs of slowing technological development become harder to justify.
The second is: Can humans and advanced AI coexist while humans retain meaningful agency? RAND deliberately makes human agency central to the issue. Survival alone is not enough. A future in which human beings continue to exist but no longer meaningfully understand, choose, or control the direction of their societies would count as a failure.
The third question is whether governments and companies can cooperate sufficiently to restrain development. International agreements are useful only if they can be verified and if participants continue to obey them even when cheating might offer an advantage.
Fourth: Can one country obtain a decisive and lasting lead? A temporary technological advantage is not the same as dominance. For a dominance strategy to work, a state would need not merely to get ahead but to remain far enough ahead to shape what everyone else can do.
And fifth: Can dangerous AI development actually be suppressed? Governments might try export controls, restrictions on advanced computing hardware, cyber operations, sanctions, or other forms of pressure. But nobody knows how effective such measures would remain against determined actors.
The problem is that governments may need to make decisions before any of these questions has been settled. This is where RAND introduces what may be the most useful concept in the report: strategic foreclosure. An option can disappear without anyone explicitly choosing to abandon it.
Suppose, for example, that future international controls on AI require sophisticated verification systems and trusted diplomatic channels. Those institutions might take many years to build. If governments neglect them today, then at some future crisis point they may discover that cooperation is desirable but no longer practically available.
The same can happen through technological dependence. Every individual deployment of AI may appear reversible. But after banks, hospitals, governments, militaries, businesses, and ordinary citizens have become deeply reliant on AI systems, substantially reducing their use might become enormously difficult.
Knowledge can also spread irreversibly. Once powerful model designs, training techniques, or model weights have been widely distributed, a strategy depending on strict technological control may simply cease to be possible.
Most troublingly, RAND argues that human agency itself could gradually erode. If people and institutions increasingly delegate analysis, judgment, and decision-making to machines, societies could retain the formal power to make choices while slowly losing the practical capacity to exercise it.
This creates a peculiar dilemma. Choosing a strategy too soon can close off alternatives. But postponing a choice can close them off as well. RAND’s answer is what it calls a Freedom of Action strategy. The idea is not to remain passive until uncertainty disappears. Quite the opposite. Governments should begin building the capabilities that several different future strategies would require, so that they retain the ability to choose later.
That means investing now in AI safety research and methods of keeping advanced systems controllable. It means developing better ways to evaluate what frontier AI systems can actually do. It means improving the security of powerful models and the computing infrastructure on which they depend. It means establishing systems capable of detecting dangerous AI developments and responding to major incidents. It also means maintaining diplomatic channels and developing technologies that could verify future international agreements.
Existing national-security institutions would also have to adapt. Nuclear deterrence, cybersecurity, biosecurity, intelligence, conventional military planning, and counterterrorism could all be transformed by increasingly capable AI. Waiting for a crisis before reconsidering them may be too late.
RAND also emphasizes preparation at a more basic level. Political leaders and government institutions need to understand the technology well enough to make decisions under pressure. Contingency plans should be developed before they are needed. Citizens need enough AI literacy to distinguish serious risks from hype. And societies need to think explicitly about which decisions should remain under meaningful human control.
All of this requires continuing observation of the technological frontier. RAND proposes treating AI as a problem of strategic warning: constantly watching for evidence about how dangerous systems are becoming, whether coexistence appears feasible, whether rivals are accelerating, whether cooperation remains possible, and whether particular strategic options are beginning to close.
A useful idea here is what the report calls the latest decision date. Suppose a capability takes five years to build, but circumstances suggest it may be needed three years from now. The real decision point has already arrived. Waiting another year is not preserving neutrality. It is effectively choosing to abandon that option. This may be the report’s most important lesson.
We tend to imagine that the great decision about superintelligence will occur sometime in the future, when the technology has become powerful enough that governments must finally decide what to do about it. RAND argues that this picture is misleading. By the time the danger becomes obvious, many of the important choices may already have been made—not through dramatic decisions in presidential offices or international summits, but through years of apparently ordinary choices about research, infrastructure, regulation, diplomacy, security, and deployment.
The report therefore does not really offer a strategy for superintelligence itself. It offers a strategy for reaching the moment when we may finally understand what strategy superintelligence requires. Its message is that uncertainty is not a reason to postpone preparation.
***
Enjoying the content on 3QD? Help keep us going by donating now.