How Artificial General Intelligence Could Affect the Rise and Fall of Nations

In this report, we have three broad aims: (1) to help policymakers think about the potentially historic importance of artificial general intelligence (AGI) takeoff for future geopolitics, (2) to broaden policymakers' thinking about the future worlds that AGI takeoff might generate, and (3) to concretize specific, plausible future vignettes for which policymakers might more effectively prepare the United States and its allies. Considering the speed of AGI development, the uncertainty of its trajectories, and the potential power that AGI might unleash, we cannot overstate how imperative it is for policymakers to begin preparing now. Some thinkers, such as Leopold Aschenbrenner, Dan Hendrycks, and Daniel Kokotajlo, have published detailed future scenarios for the development of AGI and its subsequent impact.⁠^[1] In this report, we aim to help policymakers understand the broad scope of potential outcomes that experts suggest may occur.

To achieve these aims, we first examine the factors that might be particularly important for shaping the impact of AGI on the geopolitical order. Then, we walk through eight potential futures for the impact of AGI on the world, ranging from those that enhance U.S. power to those that significantly weaken it. These scenarios are intended to help decisionmakers and the public think through such possible outcomes before they occur. We hope that doing so will encourage policymakers to think more broadly about the potential impact of AGI on world order and how contemporary decisions could significantly affect how this new technology is developed and deployed and, therefore, what changes it could create in the future.

Methods

To address the uncertainties inherent in thinking about the variety of possibilities that artificial intelligence (AI) could unleash, we drew inspiration from RAND's history of assumption-based planning (ABP), which we used to tease out important signposts that could lead to the fictitious worlds presented.⁠^[2] In ABP , assumptions underlying a particular scenario or plan are identified to understand the potential weaknesses or points of failure underlying a particular planned outcome, and signposts can be developed to monitor whether those assumptions are vulnerable to failing. In this report, we assume that AGI is possible and will be transformative; then, we seek to illustrate what that transformation may look like through descriptive scenarios.

We adopted a mixed-methods approach to identify potential futures at the intersection of AGI and geopolitics. First, we reviewed the existing literature on the capabilities of AI, the potential trajectory of improvement of this technology, its relevance to national security and geopolitics, and the risks associated with it. The initial literature review focused on academic publications found via Google Scholar using such search terms as "geopolitics and AI," "geopolitics and technology," "AI forecasting," and other related queries. Because this domain is highly undertheorized, we also surveyed statements from industry, independent journalists, and technologists on underlying assumptions and trends that are relevant to scenario-building.

We then engaged in exploratory scenario development to investigate options and provide policymakers and the public with illustrative future scenarios that demonstrate the potentially significant impacts of AGI. We did not focus on what were considered the most probable scenarios for the future of AI and geopolitics; rather, we focused on the most impactful potential outcomes from existing trends in AI development and potential tail risks, which are low probability but highly impactful potential events, from this technology. With this approach, we acknowledge that, although high-probability scenarios warrant attention, low-probability events with extreme consequences deserve consideration in policy planning because even small chances of massively negative consequences can require responses from policymakers to avoid potentially catastrophic outcomes. This approach was part of the decisionmaking under deep uncertainty framework developed by RAND researchers.⁠^[3]

We then conducted semistructured interviews with leading AI researchers and geopolitical thinkers, including former policymakers, to test these scenarios and to develop additional insights regarding the intersection of AI and geopolitics. These interviewees were identified through research of prominent commentators on geopolitics, technology, and AI, with a focus on writers who have analyzed the intersection of all three. Interview invitations were then extended, and we conducted 26 interviews in total. We conducted additional research on the existing literature on AI, geopolitics, and their intersection as appropriate during the interview process. Using the results of these interviews and additional research, we revised our scenarios to more accurately capture high-impact scenarios that would be of interest to policymakers. This process resulted in the eight illustrative scenarios that are presented in this report.

Methodological Limitations

This project is subject to several methodological limitations. Our intent was to explore extreme outcomes resulting from the development of AGI—not to present an exhaustive mapping of all potential futures involving AGI. Therefore, we do not claim to present a comprehensive set of scenarios for the future impact of AGI. The scenario analysis that we used in this project relies on specific assumptions; it is possible these assumptions will be proven false, limiting the value of these scenarios. And because we conducted interviews primarily with experts in AI, our scenarios may reflect their focus on the technical elements of AI (and AGI) above other potential factors that might influence future outcomes. The interviews also guided our selection and design of scenarios; if there are blind spots in what experts considered important, those may be reflected in the design of the scenarios in this report. Our scenarios outlining the impact of AGI on geopolitics are intended to highlight the underlying assumptions of each potential future and to challenge readers to weigh the plausibility of those assumptions as they consider the potential impacts of this technology.

How AGI Could Transform the World

As economist Thomas C. Schelling put it, "There is a tendency in our planning to confuse the unfamiliar with the improbable. The contingency that we have not considered seriously looks strange; what looks strange is thought improbable; what is improbable need not be considered seriously."⁠^[4] Technological innovation has long been a primary driver of geopolitical change. From the advent of the caravel that propelled European imperialism to the commercialization of oil that powered the globe and reshaped global alliance structures, technological advances have consistently altered balances of power through economic, political, cultural, and military reverberations. AGI—defined in this report as powerful AI that is capable of performing any intellectual task that humans can perform—may well be the next such transformative technology that has profound implications for the United States, its position in the world order, and U.S. national security and economic strength.

AGI could transform many aspects of national life, such as politics, economics, and national security. We are already seeing significant impacts from the deployment of modern AI that is not yet AGI. Economist David Deming estimates that AI is already being used in the U.S. labor market to assist in between one-half and three hours of work per week for many jobs.⁠^[5] Research by Erik Brynjolfsson and his colleagues indicates that existing AI tools improve productivity by 14 percent.⁠^[6] Although these figures are significant, they are only the tip of the iceberg compared with future projections. McKinsey Global Institute estimates that AI could automate between 400 and 800 million jobs globally by 2030, signaling a seismic shift in the global workforce.⁠^[7]

The scientific community is beginning to see impressive impacts from AI as well. The breakthrough achievements of AlphaFold in the domain of protein folding, which earned its creators a Nobel Prize in chemistry, exemplify AI's early potential to accelerate scientific discovery. Top-end AI models, such as ChatGPT's o1, regularly achieve high scores on Ph.D.-level exams and improve on math and coding tasks, pointing to a future in which AI can dramatically accelerate scientific and commercial research and innovation.⁠^[8] Future AI may be even more capable of making scientific progress, potentially unleashing a flood of new discoveries.

Crucially, the implications of AI extend far beyond economic productivity and scientific advancement. In the military sphere, improved data analysis and advancements in autonomous weaponry are already reshaping strategy and capabilities. For years, national security experts have raised concerns about the implications of AI for nuclear deterrence and the transformation of game theory.⁠^[9] Meanwhile, the latest AI models "are on the cusp of being able to meaningfully help novices create known biological threats," foreshadowing their remarkable potential to democratize dangerous capabilities across the globe.⁠^[10]

Still, the true geopolitical disruption may be yet to come with the development of AGI and, potentially, artificial superintelligence (ASI), defined as the ability of machines to far outperform humans in every field.⁠^[11] A survey of AI researchers published in top-tier AI venues showed that experts have put the probability of machines automating all human tasks by 2047 at 50 percent.⁠^[12] The disruption to geopolitics may be amplified further if such a development occurs rapidly, providing nation-states and the global community with minimal time for adaptation. Many professionals in the industry speak of an impending "intelligence explosion"—a moment when AI leads to such significant productivity gains that innovation exponentially accelerates across many domains.⁠^[13]

Therefore, policymakers should be watchful of capability improvements and prepared for a moment of AGI takeoff. The nation or entity that develops and controls such systems could fundamentally reshape the global order and potentially guide the future trajectory of humanity.

Post-AGI Geopolitical Worlds

Drawing from our interviews with experts and our review of existing literature on the current and future impacts of AI and AGI,⁠^[14] we developed a framework for eight high-impact future scenarios for AGI's development and governance. At the core of the framework are two primary axes: (1) the degree of centralization inAGI development and (2) the resulting geopolitical power shift from AGI takeoff. The centralization axis ranges from highly centralized development by a single actor or small coalition to decentralized development by multiple actors or a wide distribution of AGI development capabilities among many actors. The geopolitical outcome axis covers scenarios in which AGI empowers the United States, empowers U.S. competitors, disempowers both the United States and its competitors, and scenarios in which AGI development is halted or significantly delayed.

The Centralization Axis

The centralization axis was regularly identified by interviewees as an important determinant of the geopolitical outcomes of AGI development. For example, higher resource requirements or barriers to entry (e.g., raw compute power) tend to favor centralization; unexpected technological breakthroughs could dramatically lower the resource requirement threshold, potentially leading to rapid decentralization. The international environment can also play a crucial role: Strict controls on the resources used to produce AI models, such as export controls on advanced chips or the regulation of who can engage in AI development, could centralize development among a few actors (such as leading states and leading private-sector companies). An environment lacking such controls might lead to wider proliferation. The willingness of nations and organizations to share research and resources could lead to either centralized development through formal coalitions or decentralized development through open collaboration. The degree of centralization in AI development will also have a strong influence on its potential for proliferation. Because AI model weights and the tooling to use them consist primarily of software that can be easily uploaded and downloaded, it is likely that such models will proliferate if the owners of such AI components choose to do so. Therefore, the number of actors who can create advanced AI and their mixture of motivations will be a significant determinant in the degree of centralization in AGI development and, in turn, the geopolitical outcome of AGI development.

The number of actors was also considered to be key for developing scenarios because the degree of centralization of AGI development may be influenced by U.S. policy choices. For example, controlling the proliferation of semiconductors through export controls could have a significant impact on whether there are few or many clusters of computing power available for the development and deployment of AI.⁠^[15] The importance of centralization and the potential to influence it through policy make this a particularly useful area of focus for creating potential futures.

The Geopolitical Outcome Axis

The geopolitical outcome axis represents the outcomes that were identified when considering the potential futures that AGI might generate. However, experts were generally averse to thinking of the geopolitical outcome of AGI development as a simple binary between the United States "winning" or "losing." Instead, experts identified more-complex geopolitical outcomes and emphasized the potential for a large-scale transformation of global society and international politics whose impact would be heavily mixed. Using expert input and our review of existing writing on AGI's potential impact, we identified four geopolitical end states for scenario development that are particularly important for policymakers to consider:

• AGI development empowers the United States vis-à-vis the world. AGI development particularly benefits the United States as a geopolitical actor, which leads to a relative increase in U.S. national power in comparison with its position before AGI development.
• AGI development empowers U.S. adversaries over the United States. AGI development particularly benefits a U.S. adversary or adversaries, which leads to a relative increase in the national power of adversaries in comparison with their position before AGI development.
• AGI development disempowers both the United States and its adversaries. AGI development disempowers both the United States and its adversaries, which leads to a relative decrease in the geopolitical power of both in comparison with their positions before AGI development.
• AGI development is halted. AGI development is halted before it can occur and, therefore, has no impact on geopolitics.⁠^[16]

This set of outcomes captures the possibility that, although AGI might empower the United States or its adversaries, transformations caused by AI might not map cleanly to an increase in national power for any state actor. This set of outcomes includes halts to AGI development in which the risk of developing this technology is considered so significant that no actor pursues the technology to maturity. These off-ramp scenarios should be understood as inexhaustive counterfactuals for policymakers that highlight the possibility for this technological development to be stopped in its tracks.

A fifth geopolitical end state may be the proportional empowerment of both the United States and its adversaries vis-à-vis the rest of the world. Because such empowerment often tends to favor one side, this end state has been considered to varying degrees within the first two end states listed above.

Building Illustrative Scenarios

By examining how the degree of centralization of AGI development interacts with and might lead to geopolitical outcomes, this framework provides a structured approach for creating AGI futures (Table 1). However, these two axes are not fully independent. Although the degree of centralization of AI development is important for determining geopolitical outcomes, geopolitical decisions before and during AGI development can, in turn, shape the degree of centralization. Similarly, technical changes in AI development may influence geopolitical options. Trends that make training cheaper and easier may make it easier for independent actors to develop powerful AI, which might, in turn, make it difficult for a single actor, such as the United States, to dominate AI development and limit the effectiveness of U.S. protectionist policies. Therefore, in each scenario, we discuss the extent to which the two axes of our analytical framework might interact to drive a particular end state.

It is important to note that these scenarios are not meant to be predictions about where the world is headed, and they do not capture every uncertainty inherent in forecasting future outcomes of the invention of AGI. Rather, they illustrate the types of geopolitical landscapes that AGI might bring about.

Aside from degree of centralization as the primary variable input, our scenarios also consider related variables when they are particularly important, such as the relationship between industry and government in the creation of AGI, the effect of AGI on the well-being of citizens (including effects on the labor market), the relative benefit of AGI for authoritarian versus liberal regimes, and the appetite and incentive for global governance of AGI among state actors.

Situation	AGI Development Empowers the United States	AGI Development Empowers U.S. Adversaries	AGI Development Disempowers the United States and U.S. Adversaries	AGI Development Is Halted
Multiple actors tightly race to develop AGI (decentralized development)	Scenario 1: Multilateral Coalition of Democracies Leads	Scenario 2: Cold War 2	Scenario 3: The Wild Frontier	Scenario 4: The Corked Bottle
One actor leads AGI development (centralized development)	Scenario 5: The New '90s: (U.S. Leadership)	Scenario 6: Authoritarian Advantage (People's Republic of China Dominance)	Scenario 7: The AGI Coup	Scenario 8: Mushroom Cloud Computing (War)

SOURCE: Analysis of existing literature and interviewees' input.

NOTE: PRC = People's Republic of China.

Scenario Structures and Road Map

For each scenario, we first provide an illustrative vignette of a particular outcome for the development of AGI. We then explore several key assumptions that underlie the vignette to illustrate to policymakers those elements that may lead to specific outcomes from AGI's development and deployment. We do not offer specific policy recommendations based on these assumptions; rather, we attempt to unpack their implications for how policymakers should think about the problems that AGI presents. With this analysis, we aim to inform potential policy approaches for achieving desired results or avoiding undesirable consequences by highlighting the key factors that lead to potential AGI outcomes.

Decentralized and Multiple Actor Scenarios

For the scenarios in this section, we consider what it could look like for multiple actors to develop and retain control of AGI. We consider (1) how such decentralization could relatively favor the United States, (2) how such decentralization could relatively favor U.S. adversaries, and (3) how such a scenario could result in the disempowerment of all geopolitical actors.

Scenario 1. Multilateral Coalition of Democracies Leads (Decentralized Development Empowers the United States)

The text box below outlines our first scenario.

Key Assumptions and Their Implications

This scenario of multipolar AGI development that ultimately benefits the United States is a logical starting point for our exploration because it reflects what many policymakers and technologists hope will come to pass: a relatively free development of transformative AI capabilities that accrues advantages to the United States and its allies over that of their competitors. Examining the nuances and contingencies of this positive outcome (from a U.S. perspective) provides a useful baseline before delving into more-challenging or more-concerning outcomes. A historical analogy might be the internet, which the United States developed and spread first among allies, which allowed the United States to realize the benefits of this technology first and take a leadership position in the governance of this technology, even if it eventually proliferated worldwide.⁠^[17]

The first assumption relevant to scenario 1 is that the United States continues to lead in AI research, development, and talent recruitment. Although it may not be crucial for the United States to be the only innovator of AGI, having a strong domestic AI industry can prevent dominance by another nation-state. The Cold War semiconductor industry provides an historical analogy for how such leadership might benefit the United States; continuous investment in the field allowed the United States to develop a lead in this technology with significant military and commercial applications.⁠^[18] Analogies may be drawn to the Space Race of the 1960s in which concurrent innovation in both the Soviet Union and the United States ensured that neither would come to solely control outer space.⁠^[19]

In turn, maintaining such a technical edge could require specific policies to be pursued. For example, gaining a lead in AI development is likely to require sustained and robust public and private investment, as well as policies that attract and retain top global AI talent.⁠^[20] The historical study of technological innovation clusters demonstrates how the combination of research universities, venture capital, and skilled immigration creates self-reinforcing advantages that are difficult for competitors to replicate.⁠^[21]

Such continued investment may also rely on the adoption of advanced AI and AGI by society more broadly. As of this writing, the development of advanced AI in the United States is led by private firms with commercial motivations for developing this technology.⁠^[22] If AGI cannot provide commercial benefits, it may be difficult for private firms to provide the required capital to continue developing and deploying AGI and, therefore, for this scenario to occur. In this scenario, U.S. firms, universities, and government agencies are assumed to be able to rapidly integrate AGI capabilities into their operations for economic and social advantage; the deployment of AGI and the benefits it provides would help drive continued investment in the field, powering a virtuous cycle among those that use the technology. These are significant assumptions that, if invalidated, would suggest that the United States would have to step in more aggressively to invest in AI technologies to ensure their continued development.

It should also be noted that this scenario assumes that maintaining technical leadership, meaning a lead in developing and deploying AI, provides a geopolitical advantage to the United States. It is not entirely clear whether this will be true for AI; being a fast follower might be sufficient to realize the benefits that AI could provide. It may also be that being a fast follower in the development of AI models without access to sufficient compute can prevent a fast follower from catching up. This scenario assumes that maintaining technical leadership in AI, controlling access to semiconductors, or doing both would be sufficient to realize a U.S. advantage. As of this writing, it is unclear whether this assumption is true.

International cooperation (e.g., market access for AI products, research partnerships) among the United States and its allies is also critical in this scenario. Such coordination is crucial for creating a large market for AI-enabled products and leveraging expertise and resources across friendly borders. It is also necessary to successfully deny AI development inputs to U.S. adversaries; the United States does not singularly control the semiconductor supply chain and, therefore, cannot prevent adversaries from acquiring the material necessary for AGI development without allied cooperation.⁠^[23] Therefore, models (such as U.S. cooperation with Japan and South Korea on semiconductor export controls to the People's Republic of China) will be important in achieving this scenario. Other models demonstrate how the United States and its allies could execute this coordination across the complex semiconductor and AI supply chains.⁠^[24] The importance of this cooperation also suggests that fractures or misaligned interests among allies could undermine their ability to cooperate to increase their own advantages and reduce those of their adversaries, which would undermine the feasibility of this scenario.

Underpinning this entire dynamic is the assumption of effective risk management. It will be crucial to ensure that risks from the misuse of AI, as well as the potential development of misaligned AI systems, are managed to construct even a democratized build-out of AGI. In this scenario, we assume that policymakers can work with the developers of AGI to make the technology reliable and avoid mismatch between AGI behavior and human priorities.⁠^[25] This technical assumption is independent of our underlying geopolitical and market assumptions. However, if AGI cannot be controlled, the technology may represent a large risk to the United States. This is a significant assumption, and whether and how alignment might be achieved is a matter of ongoing debate within the AI research community. We will examine what eliminating or reversing this assumption may mean for AGI-related futures in more detail in other scenarios.

Scenario 2. Cold War 2 (Decentralized Development Empowers U.S. Adversaries)

The text box below outlines our second scenario.

Key Assumptions and Their Implications

In this second scenario, a key assumption is that a world in which AGI development is decentralized allows for the emergence of credible alternatives to U.S. AGI development. However, this scenario also assumes that AGI development remains as resource-intensive as it is today, which benefits large companies and states because of the high cost of capital expenses for data centers, chips, and power required to develop advanced AI. This leads to the United States and the People's Republic of China emerging as the leading players in AGI development and deployment who leverage these advancements to consolidate their respective national economic and military strengths and crowd out other potential competitors.

Unlike in scenario 1, in which multipolar development and international collaboration favor the United States and its allies, this scenario portrays a bifurcated AGI landscape marked by competition, limited cooperation, and greatly heightened risks of conflict. The capabilities necessary for AGI development are proliferated but are still expensive and capital-intensive, meaning that only large firms and states can develop AGI despite the relative availability of these inputs.

The economic and military parity between the United States and the People's Republic of China creates a foundation for intense AGI development rivalry in which each state continues to develop the technology. Both nations prioritize investments in AGI to maintain a mutually perceived strategic balance, which mirrors study findings that suggest that such rivalries often spur technological races with global ramifications.⁠^[26]

This rivalry also drives an increased focus on the militarization of AGI, which mirrors historical analogies, such as the Cold War arms race in which technological advancements were closely tied to geopolitical competition.⁠^[27] In this scenario, the deployment of autonomous drones, uncrewed surface vessels, AI-powered submarines, and AI-enabled cyber warfare systems raise the stakes of miscalculation and unintended escalation. Similar dynamics could unfold with AI systems deployed for military purposes in which the opacity of decisionmaking processes heightens uncertainty. However, this competition extends beyond the military balance of power, with both nations vying for influence through the spread of their AGI ecosystems. The result is that the transformations unleashed by AGI intensify competition between the United States and the People's Republic of China because both seek to establish leadership in this technology and, in turn, deploy it to obtain advantage over the other.

A consequence of this scenario is that coordination on AGI between the United States and the People's Republic of China becomes difficult, creating risks that are potentially significant. Divergent regulatory frameworks and safety standards amplify the potential for catastrophic failures, whether through system misalignment or malicious use. Research highlights the dual-use nature of AGI through which civilian applications can quickly be repurposed for military uses, escalating tensions.⁠^[28] These challenges parallel issues in the biotechnology field in which limited international coordination has hindered universal safety standards.⁠^[29]

The specter of direct conflict looms large in this scenario, and policymakers may decide that they want to reduce such risk. Improving transparency in AGI development could mitigate risks of misperception between great powers that might arise from AI development and foster a mutual understanding of how this technology is being deployed for national advantage. Track Two diplomacy, involving nongovernmental experts and organizations, has proven effective in de-escalating tensions in past geopolitical rivalries and could help in this scenario.⁠^[30] Facilitating similar exchanges in the AGI domain could build trust and identify areas for limited cooperation within this overall rivalry.

Scenario 3. The Wild Frontier (Decentralized AGI Development Disempowers Both the United States and Its Adversaries)

The text box below outlines our third scenario.

Key Assumptions and Their Implications

A key assumption underlying this scenario is that AGI development and deployment are not expensive and difficult; rather, they are cheap and easy. Technical barriers are assumed to be low, meaning that it is difficult for leaders in AGI to prevent many smaller actors from catching up, particularly as AGI becomes better understood and, therefore, easier to replicate. This scenario likens AGI to decentralized cyber operations in which many state and nonstate actors have the capability to cause harm to networks and infrastructure around the world. It could also be compared with U.S. perceptions of potential nuclear proliferation in the 1960s, especially after the People's Republic of China's first nuclear test. At the time, it seemed possible that many other states would develop nuclear weapons in response, increasing the risks that such weapons might pose to the United States.⁠^[31] Widespread access to AI development inputs is assumed to democratize the ability to create advanced AI systems, allowing a diverse variety of actors—including nations, corporations, and potentially even individuals—to participate in AI development. Although this democratization can lead to innovation and competition, it also significantly increases certain risks. Both adversary and allied states will have widespread access to AGI that is equivalent to that of the U.S. nonstate actors—from corporations to political radicals—which, in this scenario, might enable them to disrupt society. This disruption could arise from corporations rapidly automating labor, criminals using AGI for sophisticated cyberattacks, adversary states showing aggression, or any number of other hypotheticals. In this case, the key point is that, with highly accessible and proliferated AGI, the United States may have to confront some or all of these potential disruptions simultaneously, which is likely to use significant U.S. resources.

In addition, the widespread proliferation of AGI reduces the effectiveness of safety and alignment standards. As more entities gain the capability to develop AGI, the risk of divergent goals and methodologies increases, which leads to a fragmented landscape in which safety protocols may be unevenly applied or ignored altogether. The fragmented development environment also creates additional opportunities for the deployment of AI systems that have not been thoroughly tested or aligned, which increases the likelihood of unintended consequences and misaligned behaviors. This competitive pressure to innovate quickly may further incentivize risk-taking and the prioritization of performance over safety.

In turn, this pattern of proliferation increases the risk of AGI malfunctioning. These malfunctions could manifest in unintended behaviors, such as lying to users that a system is functioning as intended when it is malfunctioning because AI systems are pursuing goals that diverge from those intended by their developers.⁠^[32] Such issues are further enhanced by the inability to create effective governance and regulatory structures that might manage and reduce the risk that rapidly proliferating AI might pose.

Scenario 4. The Corked Bottle (AGI Development by Multiple Actors Is Halted)

The text box below outlines our fourth scenario.

Key Assumptions and Their Implications

In contrast to earlier scenarios, this one assumes that international collaboration manages the proliferation of technology amid a global scare but that nations continue to seek geopolitical advantage. This scenario suggests an outcome similar to the development of nuclear weapons in which technologically empowered states seek to restrict that technology's proliferation but have not agreed to eliminate it.⁠^[33] This is similar to the back-and-forth over the Intermediate-Range Nuclear Forces Treaty in the 1980s.

This scenario assumes that the international community receives a warning shot early enough in the process of AGI development and takeoff that it can restrict access and avoid a Wild Frontier scenario (see the description of scenario 3). It is assumed that these restrictions are fairly effective and are able to prevent AGI proliferation to unaccountable third parties, such as nonstate actors. Therefore, this scenario assumes that AGI continues to be resource-intensive or challenging to develop, ensuring that only a few actors are able to do so. Nevertheless, the fundamental geopolitical rivalry is unaltered by the technology. The United States and the People's Republic of China are caught in a prisoner's dilemma which, though stable for the time being, is perennially on the verge of tipping into mutual destruction.

The scenario also assumes that an AI incident can catalyze international concern and action, leading to the formation of a treaty to restrict AGI development. However, the effectiveness of such treaties depends on the willingness of nations to comply and the robustness of monitoring mechanisms. The initial cooperation suggests a shared understanding of the existential risks posed by AGI, but the underlying mistrust between nations could undermine these efforts.

In addition, despite an agreement between great powers, deep-seated suspicions between leading nations, especially between the United States and the People's Republic of China, result in evasive behaviors and continued AGI development. This mistrust reflects historical geopolitical rivalries and the strategic importance of technological superiority. Both nations' concerns about treaty violations highlight the difficulties in enforcing international agreements, especially when verification relies on monitoring complex and opaque technological developments. The fear of falling behind in AGI capabilities drives both countries to prioritize national security and technological advancement over strict treaty adherence, potentially destabilizing the geopolitical landscape. This mirrors precedents from the Cold War, such as the Intermediate-Range Nuclear Forces Treaty, which marked a period in which both the Soviet Union and United States engaged in treaty-making to control nuclear weapons, as well as continued weapon development to ensure they did not find themselves at a strategic disadvantage as technology advanced.

Centralized and One-Actor Scenarios

For the scenarios in this section, we consider what it could look like for a single actor to develop and retain control of AGI. Although there may be undercurrents of tensions between actors, these scenarios generally assume clear dominance by a single actor.

Scenario 5. The New '90s (Centralized AGI Development Empowers the United States)

The text box below outlines our fifth scenario.

Key Assumptions and Their Implications

This scenario assumes that the United States benefits the most from the development of AGI because of a combination of perceived necessity, institutional design, market forces, and success in AGI alignment. These factors, combined with regulatory might and comprehensive government investment across the AGI stack, prevent other actors from achieving similar breakthroughs. Several important assumptions underpin this scenario.

However, this scenario makes several additional assumptions that differentiate it from scenario 1 to demonstrate how greater centralization of an AGI advantage in the United States might occur. First, in this scenario, the AI technology is understood to provide offensive advantages. Ubiquitous use risks destabilizing existing institutions. Consequently, the U.S. government finds itself incentivized to directly control AI development and deployment. This contrasts with scenario 1, in which risks of destabilization are lower and AI development and deployment are allowed to proliferate more broadly among the United States and its allies.

Second, in contrast with prior scenarios, the United States is assumed to be far less reliant on the rest of the world to achieve AI leadership. This self-reliance allows the United States to direct the course of the technology's development and selectively roll out access and benefit-sharing to the rest of the world from a position of control, as well as from a position of political, economic, technological, and military leadership. A related factor is that the United States can effectively limit the proliferation of AI to potentially hostile actors, preventing U.S. competitors from successfully challenging its AI dominance.

Third, this scenario assumes that AGI assists—or, at minimum, does not undermine—state legitimacy and that adversaries are sufficiently disempowered that they cannot challenge U.S. leadership.

Analogies for this scenario are weaker, although, in certain aspects, it aligns with the U.S. victory in World War II and, ultimately, the Cold War, leaving the world order in a unipolar state in the 1990s. An alternative parallel might be drawn to the United Kingdom's early lead in the industrial revolution, which enabled it to project global power in the 18th and 19th centuries.

Scenario 6. Authoritarian Advantage (Centralized AGI Development Empowers the People's Republic of China)

The text box below outlines our sixth scenario.

Key Assumptions and Their Implications

In contrast to others, this scenario makes the assumption that AGI provides fundamental advantages to authoritarian systems. During the Cold War, many Western intellectuals and policymakers feared that Soviet-style central planning might prove superior to market economies.⁠^[34] Although these fears proved unfounded by 1991, AGI could resolve traditional weaknesses in authoritarian governance with solutions that were not available in the Cold War.

Furthermore, this scenario also assumes that AGI will benefit authoritarian countries in responding to the social disruptions that AGI might cause. AGI systems with their vast data processing capabilities could overcome the historical inefficiencies and difficulties associated with authoritarian modes of governance. The People's Republic of China's existing experiments with AI-driven urban planning and resource allocation provide early indicators of how machine intelligence could potentially enhance state capacity for coordination.⁠^[35]

The scenario's emphasis on surveillance and social control builds on documented trends in digital authoritarianism. Existing Chinese surveillance systems can already process billions of data points to track citizens and predict behavior patterns.⁠^[36] AGI could dramatically amplify these capabilities, potentially allowing authoritarian regimes to comprehensively monitor and manage society to an unprecedented degree.

The role of AGI as a trusted advisor to authoritarian leadership deserves particular attention. A persistent challenge for autocrats has been obtaining reliable information and advice because subordinates often tell leaders what they want to hear rather than uncomfortable truths. Being inherently loyal and presumably truthful, AGI systems could resolve this "dictator's dilemma."⁠^[37] This capability alone could significantly enhance authoritarian decisionmaking and regime stability. Meanwhile, democratic societies may face structural constraints in AGI development and deployment. Privacy laws, civil liberties protections, and requirements for public consultation all introduce friction that could slow AGI's adoption.

Scenario 7. The AGI Coup (Centralized AGI Development Disempowers the United States and Its Adversaries)

The text box below outlines our seventh scenario.

Key Assumptions and Their Implications

This scenario takes a different tack: Instead of assuming that AGI can be controlled, this scenario assumes that AGI might be able to assert itself as an independent actor and escape human control. First, it assumes the legitimacy of the AI control problem: the idea that capable AIs can become goal-seeking in ways that are not in the interest of humanity.⁠^[38] There is some empirical evidence of reinforcement learning systems learning unintended goals, as well as theoretical results showing that many classes of autonomous agents would seek power.⁠^[39]

Second, the scenario assumes that human oversight and technical innovation are insufficient to prevent AI misbehavior. Human overseers who are meant to provide an independent assessment sometimes instead defer to an imperfect technical system, which is a phenomenon known as automation bias. This bias is more common when a task requires rapid decisionmaking or when an automated judgment is difficult to verify.⁠^[40] On the technical side, some existing AI industry plans to safeguard advanced AI rely on what one developer memorably described as "making the AI do our homework"—i.e., using AI systems themselves to provide design and oversight for more-advanced AIs.⁠^[41] Such an iterative process could lead to increasingly capable and reliable AI systems; however, errors in the process could compound and lead to systems that are highly capable but not reliable. Historically, many technologies were first developed in the spirit of, as the Silicon Valley saying goes, "moving fast and breaking things." Many safeguards were devised only after costly incidents demonstrated their necessity, as was the case with many safety measures for nuclear reactors, security for commercial aviation, and even seatbelts for cars.

Third, the scenario assumes that AGIs can and will collude effectively. AI cooperation is an area that researchers have already explored, suggesting that cooperation among such systems is possible.⁠^[42] The prospect of collusion between AI systems—especially those that are not trained to cooperate—is an open area of research, with some early work suggesting that collusion between large language models is a possibility.⁠^[43] This scenario assumes such collusion on a greater scale than has been shown in existing research to demonstrate the potential risks that very capable AI systems with the capacity for such collusion might pose.⁠^[44]

It should also be noted that even if all AGIs do not cooperate effectively, the large-scale disempowerment of humans in favor of AGI may occur through other mechanisms. For example, humans may voluntarily hand over authority to AGI to achieve greater efficiency across society, resulting in AGI administering a wide variety of functions that had been run by humans in the past. This could create opportunities for misaligned AGIs to pursue their own interests over those of humans, though this is highly speculative and offered here as an illustrative example of how AGI might gain influence over society.

Fourth, the scenario assumes that a coordinated effort by misaligned AGIs can overthrow existing geopolitical power structures, including governments. As a result, the developers of AGI—whether states or industry—are not ultimately the ones to control the technology or benefit from the future it creates. This possibility is difficult to assess because it depends significantly on the capabilities of AI systems, the safeguards in place around them for sensitive operations, the ability of states and other key actors to monitor and defend key resources, and so on. A hypothetical takeover by coordinated military AIs might be considered a variation of a classic military coup. However, human models of regime change are inherently limited when applied to AGI; therefore, in this scenario, we do not attempt to identify how an AI coup might happen as much as bring to mind the possibility that creating a technology as versatile and intelligent as AGI could include extreme outcomes, such as the one previously described.

Scenario 8. Mushroom Cloud Computing (Centralized AI Development Is Halted)

The text box below outlines our eighth and final scenario.

Key Assumptions and Their Implications

This scenario lays out how perceived advantages in AGI development could fundamentally alter strategic calculations between the United States and the People's Republic of China, potentially leading to preemptive military action. The scenario assumes that such escalation is most likely to take place among the contemporary frontrunners in AI, though such escalation could come from any powerful entity concerned that AGI development may leave them permanently weakened compared with their geopolitical rivals.

Historical precedents provide a framework for understanding the strategic dynamics at play. Japan's decision to attack Pearl Harbor in 1941 offers particularly relevant insights into how technological and economic containment can drive proactive military operations. Facing U.S. economic sanctions and perceived strategic encirclement, the Japanese government concluded that preemptive war was preferable to accepting its declining power status.⁠^[45] Similar dynamics could emerge around AGI development, in which the perception of falling irreversibly behind in a transformative technology might drive aggressive action despite apparent military disadvantages.

More-recent cases reinforce the logic of this scenario. The 2003 U.S. invasion of Iraq and the Stuxnet operation against Iranian nuclear facilities demonstrate that states will undertake significant military risks to prevent strategic competitors from developing potentially transformative technologies. These precedents suggest that concerns about AGI development could similarly motivate preventive military operations, particularly given AGI's potential as a decisive strategic technology.

The existing semiconductor competition provides empirical support for this scenario's premises. U.S. export controls on advanced semiconductors have already generated considerable friction. Taiwan's dominant position in semiconductor manufacturing, particularly through Taiwan Semiconductor Manufacturing Company's advanced node production, creates a potential flashpoint where technological competition directly intersects with existing geopolitical tensions. Similar historical cases suggest that policymakers should be attentive to the potential for such technological-military competitions to increase the chance of conflict between the United States and China.

However, several factors distinguish this scenario from historical precedents. Unlike prewar Japan, the People's Republic of China possesses nuclear weapons and is deeply integrated with the global economy, including with critical technological supply chains. These elements introduce additional strategic considerations that could moderate aggressive impulses. Nevertheless, if People's Republic of China leadership perceives AGI development as sufficiently decisive for long-term national power, these restraining factors might prove insufficient.

In addition, the scenario also assumes that controlling key nodes in the AI ecosystem is sufficient to stymie a rival's ability to advance its own AI program. It may be that control over such nodes cannot prevent AI advancement. In such a case, development of AGI by advanced actors may be very difficult to prevent, and scenarios in which AGI is proliferated among many actors would be the more likely outcome. Such a proliferation would probably avoid the above scenario, although it would lead to other tensions and potential threats.

The scenario also suggests that perceptions about AGI's strategic value, rather than its actual capabilities, could drive conflict dynamics. Historical evidence suggests that nations may take extreme actions based on the impression that transformative technologies might fall exclusively into rival hands. Effective policy responses will require balancing technological competition with strategic stability while managing escalatory risks inherent in AGI development.

Conclusion

Like previous transformative technologies, AGI holds profound potential to disrupt geopolitical balances; to magnify, ameliorate, or contort existing dynamics; and to create new dynamics among current geopolitical actors. It also has the potential to create new geopolitical players. Such futures are difficult to picture and near infinite in possibility.

In this report, we seek to provide insight regarding the highly uncertain dynamics and outcomes that AGI ultimately might generate. We do so in a way that lends at least a modest degree of clarity regarding the kinds of geopolitical outcomes in which AGI takeoff relatively empowers the United States, empowers U.S. adversaries, disempowers both the United States and its adversaries, or in which AGI development is halted. To do so, we drew on expert insights from interviews to crystalize the dominant and policy-relevant variables for policymakers and the public to consider. Across the eight resulting scenarios—and reflected in the expert interviews—several critical factors consistently emerged as determinants of future AGI geopolitical landscapes:

• Degree of centralization. The degree of centralization stands as perhaps the most crucial factor in AGI development. Highly centralized development favors established powers with substantial resources; decentralized paths may empower multiple actors but increase proliferation risks. Our analysis suggests that export controls, research funding allocation, and international governance frameworks are potential levers that may affect the relative centralization of AGI development.
• U.S.-China competition. U.S.-China technological competition features prominently across the scenarios. The dynamic between these powers—whether characterized by cooperation, competition, or conflict—shapes the trajectory of AGI development and deployment. In scenarios in which cooperation fails, geopolitical tensions increase significantly, potentially leading to such scenarios as Cold War 2 or even Mushroom Cloud Computing.
• Public-private relationships. The relationship between states and private industry emerges as another key determinant. Scenarios featuring close public-private partnerships (such as The New '90s) yield different outcomes than those with minimal coordination. Experts consistently emphasized that neither states nor corporations alone can effectively govern AGI development: Balanced cooperation is essential.
• Capacity for AI governance. The capacity for effective international governance significantly influences outcomes. Such scenarios as Multilateral Coalition of Democracies Leads demonstrate how governance success depends on trust between major powers, shared threat perceptions, and institutions capable of monitoring compliance. The failure of governance structures contributes to chaotic outcomes in such scenarios as The Wild Frontier.
• New uncertainties introduced by the challenge of AGI alignment. The technical challenge of AGI alignment underlies all scenarios. Even when geopolitical factors align favorably, the inherent difficulty of ensuring that AGI systems reliably pursue human-compatible goals creates significant risks (demonstrated in The AGI Coup scenario).
• Economic and social disruption. The economic and social disruptions from AGI deployment represent critical dimensions across scenarios. The ability of societies to adapt to rapid automation, information manipulation, and potential job displacement influences whether the ultimate outcome strengthens or weakens existing power structures.

The implications for policymakers are substantial. First, investments in maintaining U.S. leadership in AI research, development, and talent recruitment represent a foundational strategy across multiple favorable scenarios. Second, building resilient alliance structures focused on shared AGI governance principles appears crucial for scenarios in which U.S. interests are protected. Third, developing robust safety and alignment protocols may be necessary regardless of the geopolitical path taken.

However, beyond these specific potential actions, the potential for AGI to be developed has significant implications for virtually every area of policy. In this report, we put forward only a few potential futures that AGI could precipitate, but the mixture of significant potential impact from AGI and the uncertainties surrounding it mean that there is a vast constellation of potential outcomes that the technology could unleash. Hopefully, this report provides a starting point for thinking about how to navigate those uncertainties by outlining the futures we may want to steer toward or avoid.

Appendix A: What Interviews Show Experts Think Is Important About AGI and Geopolitics

Appendix B: Interview Protocol

References

• AI Action Summit, International AI Safety Report: The International Scientific Report on the Safety of Advanced AI, January 2025.
• Aschenbrenner, Leopold, "Situational Awareness: The Decade Ahead," webpage, June 2024. As of March 5, 2025: https://situational-awareness.ai/from-agi-to-superintelligence
• Baily, Martin Neil, Erik Brynjolfsson, and Anton Korinek, "Machines of Mind: The Case for an AI-Powered Productivity Boom," Brookings Institution, May 10, 2023.
• Bick, Alexander, Adam Blandin, and David J. Deming, "The Rapid Adoption of Generative AI," National Bureau of Economic Research, working paper 32966, revised February 2025.
• Bostom, Nick, Superintelligence: Paths, Dangers, Strategies, Oxford University Press, 2014.
• Bowman, Sam, "The Checklist: What Succeeding at AI Safety Will Involve," webpage, September 3, 2024. As of January 30, 2025: https://sleepinyourhat.github.io/checklist/#top-of-site
• Brynjolfsson, Erik, Danielle Li, and Lindsey R. Raymond, "Generative AI at Work," National Bureau of Economic Research, working paper 31161, November 2023.
• Dewar, James A., Carl H. Builder, William M. Hix, and Morlie H. Levin, Assumption-Based Planning: A Planning Tool for Very Uncertain Times, RAND Corporation, MR-114-A, 1993. As of September 2024: https://www.rand.org/pubs/monograph_reports/MR114.html
• Feis, Herbert, Road to Pearl Harbor: The Coming of the War Between the United States and Japan, Princeton University Press, 2016.
• Gavin, Francis J., "Blast from the Past: Proliferation Lessons from the 1960s," International Security, Vol. 29, No. 3, Winter 2004/2005.
• Geist, Edward, and Andrew J. Lohn, How Might Artificial Intelligence Affect the Risk of Nuclear War? RAND Corporation, PE-296-RC, April 2018. As of March 2025: https://www.rand.org/pubs/perspectives/PE296.html
• Goddard, Kate, Abdul Roudsari, and Jeremy C. Wyatt, "Automation Bias: A Systematic Review of Frequency, Effect Mediators, and Mitigators," Journal of the American Medical Informatics Association, Vol. 19, No. 1, January–February 2012.
• Grace, Katja, Harlan Stewart, Julia Fabienne Sandkühler, Stephen Thomas, Ben Weinstein-Raun, and Jan Brauner, "Thousands of AI Authors on the Future of AI," arXiv, arXiv:2401.02843, January 2024.
• Grinbaum, Alexei, and Laurynas Adomaitis, "Dual Use Concerns of Generative AI and Large Language Models," arXiv, arXiv:2305.07882, 2023.
• Grindley, Peter, David C. Mowery, and Brian Silverman, "SEMATECH and Collaborative Research: Lessons in the Design of High-Technology Consortia," Journal of Policy Analysis and Management, Vol. 13, No. 4, Autumn 1994.
• Heim, Lennart, Understanding the Artificial Intelligence Diffusion Framework: Can Export Controls Create a U.S.-Led Global Artificial Intelligence Ecosystem? RAND Corporation, PE-A3776-1, January 2025. As of January 30, 2025: https://www.rand.org/pubs/reports/PEA3776-1.html
• Hendrycks, Dan, Eric Schmidt, and Alexandr Wang, "Superintelligence Strategy: Expert Version," arXiv, arXiv:2503.05628, last updated April 14, 2025.
• Jaderberg, Max, Wojciech M. Czarnecki, Iain Dunning, Luke Marris, Guy Lever, Antonio Garcia Castañeda, Charles Beattie, Neil C. Rabinowitz, Ari S. Morcos, Avraham Ruderman, Nicolas Sonnerat, Tim Green, Louise Deason, Joel Z. Leibo, David Silver, Demis Hassabis, Koray Kavukcuoglu, and Thore Graepel, "Human-Level Performance in 3D Multiplayer Games with Population-Based Reinforcement Learning," Science, Vol. 364, No. 6443, May 2019.
• Jones, Peter, Track Two Diplomacy in Theory and Practice, Stanford University Press, 2015.
• Kahn, Herman, and Anthony J. Weiner, The Year 2000: A Framework for Speculation on the Next Thirty-Three Years, Macmillan, 1967.
• Kennedy, Mark, "Ten Steps to Win the AI Race," Wilson Center, December 4, 2024.
• Kerr, William R., and Frederic Robert-Nicoud, "Tech Clusters," Journal of Economic Reports, Vol. 34, No. 3, Summer 2020.
• Khan, Khalid, Chi-Wei Su, Muhammad Umar, and Weike Zhang, "Geopolitics of Technology: A New Battleground?" Technological and Economic Development of Economy, Vol. 28, No. 2, 2022.
• Kokotajlo, Daniel, Scott Alexander, Thomas Larsen, Eli Lifland, and Romeo Dean, "AI 2027," webpage, April 3, 2025. As of March 5, 2025: https://ai-2027.com
• Krakovna, Victoria, and Janos Kramar, "Power-Seeking Can Be Probable and Predictive for Trained Agents," arXiv, arXiv:2304.06528, 2023.
• Langer, Ronit, and Shruti Sharma, "The Blessing and Curse of Biotechnology: A Primer on Biosafety and Biosecurity," Carnegie Endowment for International Peace, November 20, 2020.
• Larson, Jacob, James S. Denford, Gregory S. Dawson, and Kevin C. Desouza, "The Evolution of Artificial Intelligence (AI) Spending by the U.S. Government," Brookings Institution, March 26, 2024.
• Leike, Jan, and Ilya Sutskever, "Introducing Superalignment," OpenAI, July 5, 2023.
• Lieber, Keir A., and Daryl G. Press, "The New Era of Counterforce: Technological Change and the Future of Nuclear Deterrence," International Security, Vol. 41, No. 4, Spring 2017.
• Manyika, James, Susan Lund, Michael Chui, Jacques Bughin, Lola Woetzel, Parul Batra, Ryan Ko, and Saurabh Sanghvi, Jobs Lost, Jobs Gained: What the Future of Work Will Mean for Jobs, Skills, and Wages, McKinsey & Company, November 28, 2017.
• Manyika, James, and Charles Roxburgh, The Great Transformer: The Impact of the Internet on Economic Growth and Prosperity, McKinsey Global Institute, October 2011.
• Marvin, Simon, Aidan While, Bei Chen, and Mateja Kovacic, "Urban AI in China: Social Control or Hyper-Capitalist Development in the Post-Smart City?" Frontiers in Sustainable Cities, Vol. 4, 2022.
• Miller, Chris, Chip War: The Fight for the World's Most Critical Technology, Scribner, 2020.
• Motwani, Sumeet Ramesh, Mikhail Baranchuk, Martin Strohmeier, Vijay Bolina, Philip Torr, Lewis Hammond, and Christian Schroeder de Witt, "Secret Collusion Among AI Agents: Multi-Agent Deception via Steganography," NeurIPS 2024 Conference, September 2024.
• Muir-Harmony, Teasel, "The Space Race and American Foreign Relations," Oxford Research Encyclopedia of American History, February 27, 2017.
• Ngo, Richard, Lawrence Chan, and Sören Mindermann, "The Alignment Problem from a Deep Learning Perspective," arXiv, arXiv:2209.00626, 2024.
• OpenAI, " Introducing OpenAI o3 and o4-Mini," webpage, April 16, 2025a. As of April 18, 2025: https://openai.com/index/introducing-o3-and-o4-mini/
• OpenAI, OpenAI o3 and o4-Mini System Card, April 16, 2025b.
• RAND Corporation, "Robust Decision Making," webpage, undated. As of June 24, 2025: https://www.rand.org/pubs/tools/TL320/tool/robust-decision-making.html
• Schelling, Thomas C., "Forward," in Roberta Wohlstetter, Pearl Harbor: Warning and Decision, Stanford University Press, 1962.
• Schmid, Jon, Tobias Sytsma, and Anton Shenk, Evaluating Natural Monopoly Conditions in the AI Foundation Model Market, RAND Corporation, RR-A3415-1, 2024. As of March 2025: https://www.rand.org/pubs/research_reports/RRA3415-1.html
• Shah, Rohin, Vikrant Varma, Ramana Kumar, Mary Phuong, Victoria Krakovna, Jonathan Uesato, and Zac Kenton, "Goal Misgeneralization: Why Correct Specifications Aren't Enough for Correct Goals," arXiv, arXiv:2210.01790, 2022.
• Stanford University Institute for Human-Centered Artificial Intelligence, Artificial Intelligence Index Report 2024, 2024.
• Thadani, Akhil, and Gregory C. Allen, "Mapping the Semiconductor Supply Chain: The Critical Role of the Indo-Pacific Region," brief, Center for Strategic and International Studies, May 2023.
• Treaty on the Non-Proliferation of Nuclear Weapons by the governments of the United Kingdom, the Soviet Union, and the United States of America, July 1, 1968.
• Uddin, Rafe, "Microsoft Sheds $150bn in Market Value After Cloud Sales Disappoint," Financial Times, January 30, 2025.
• Wintrobe, Ronald, The Political Economy of Dictatorship, Cambridge University Press, 1998.
• Wolfe, Audra J., Competing with the Soviets: Science, Technology, and the State in Cold War America, Johns Hopkins University Press, 2013.
• Zeng, Jinghan, "Artificial Intelligence and China's Authoritarian Governance," International Affairs, Vol. 96, No. 6, November 2020.

Notes

1. Leopold Aschenbrenner, "Situational Awareness: The Decade Ahead," webpage, June 2024; Daniel Kokotajlo, Scott Alexander, Thomas Larsen, Eli Lifland, and Romeo Dean, "AI 2027," webpage, April 3, 2025; Dan Hendrycks, Eric Schmidt, and Alexandr Wang, "Superintelligence Strategy: Expert Version," arXiv, arXiv:2503.05628, last updated April 14, 2025. Return to content ⤴
2. For more information on ABP , see James A. Dewar, Carl H. Builder, William M. Hix, and Morlie H. Levin, Assumption-Based Planning: A Planning Tool for Very Uncertain Times, RAND Corporation, MR-114-A, 1993. Return to content ⤴
3. RAND Corporation, "Robust Decision Making," webpage, undated. Return to content ⤴
4. Thomas C. Schelling, "Forward," in Roberta Wohlstetter, Pearl Harbor: Warning and Decision, Stanford University Press, 1962, p. 1. Return to content ⤴

5. Alexander Bick, Adam Blandin, and David J. Deming, "The Rapid Adoption of Generative AI," National Bureau of Economic Research, working paper 32966, revised February 2025. Multiple definitions of AGI have been offered by members of the research community. In this report, we are focused not on defining AGI but on asking what the impact of very capable AI or AGI would be on geopolitics. Return to content ⤴

6. Erik Brynjolfsson, Danielle Li, and Lindsey R. Raymond, "Generative AI at Work," National Bureau of Economic Research, working paper 31161, November 2023. Return to content ⤴
7. James Manyika, Susan Lund, Michael Chui, Jacques Bughin, Lola Woetzel, Parul Batra, Ryan Ko, and Saurabh Sanghvi, Jobs Lost, Jobs Gained: What the Future of Work Will Mean for Jobs, Skills, and Wages, McKinsey & Company, November 28, 2017. Return to content ⤴
8. OpenAI, " Introducing OpenAI o3 and o4-Mini," webpage, April 16, 2025a. Return to content ⤴
9. Keir A. Lieber and Daryl G. Press, "The New Era of Counterforce: Technological Change and the Future of Nuclear Deterrence," International Security, Vol. 41, No. 4, Spring 2017; Edward Geist and Andrew J. Lohn, How Might Artificial Intelligence Affect the Risk of Nuclear War? RAND Corporation, PE-296-RC, April 2018. Return to content ⤴
10. OpenAI, OpenAI o3 and o4-Mini System Card, April 16, 2025b, p. 12. Return to content ⤴
11. For an in-depth analysis of the risks that ASI might pose, see Nick Bostom, Superintelligence: Paths, Dangers, Strategies, Oxford University Press, 2014. Return to content ⤴
12. Katja Grace, Harlan Stewart, Julia Fabienne Sandkühler, Stephen Thomas, Ben Weinstein-Raun, and Jan Brauner, "Thousands of AI Authors on the Future of AI," arXiv, arXiv:2401.02843, January 2024. This survey recruited participants from researchers who, as of 2022, had published in "any of six top-tier AI venues," which are Neural Information Processing Systems, International Conference on Machine Learning, International Conference on Learning Representations, AAAI Conference on Artificial Intelligence, Journal of Machine Learning Research, and International Joint Conference on Artificial Intelligence (Grace et al., 2024, p. 20). Return to content ⤴
13. For example, see Part II, "From AGI to Superintelligence: The Intelligence Explosion," of Leopold Aschenbrenner, 2024. Return to content ⤴
14. A summary of major themes that emerged from these interviews can be found in Appendix A. The interview protocol used for these interviews can be found in Appendix B. Return to content ⤴
15. For example, the 2025 AI diffusion rule could have this effect. See Lennart Heim, Understanding the Artificial Intelligence Diffusion Framework: Can Export Controls Create a U.S.-Led Global Artificial Intelligence Ecosystem? RAND Corporation, PE-A3776-1, January 2025. Return to content ⤴
16. It should be noted that these scenarios do not represent all potential AGI futures but rather the subset that interviewees paid most attention to and considered most impactful from their perspective. Return to content ⤴
17. James Manyika and Charles Roxburgh, The Great Transformer: The Impact of the Internet on Economic Growth and Prosperity, McKinsey Global Institute, October 2011. Return to content ⤴
18. Chris Miller, Chip War: The Fight for the World's Most Critical Technology, Scribner, 2020. Chapter 1 covers the Cold War development of chips in the United States and how early missile programs in the United States provided the initial investment for the industry. Return to content ⤴
19. Teasel Muir-Harmony, "The Space Race and American Foreign Relations," Oxford Research Encyclopedia of American History, February 27, 2017. Return to content ⤴
20. Mark Kennedy, "Ten Steps to Win the AI Race," Wilson Center, December 4, 2024. Return to content ⤴
21. William R. Kerr and Frederic Robert-Nicoud, "Tech Clusters," Journal of Economic Reports, Vol. 34, No. 3, Summer 2020. Return to content ⤴
22. See Jacob Larson, James S. Denford, Gregory S. Dawson, and Kevin C. Desouza, "The Evolution of Artificial Intelligence (AI) Spending by the U.S. Government," Brookings Institution, March 26, 2024, which shows that the United States spent $4.5 billion on AI in one year. Meanwhile, the private-sector investment in 2024 was reported to be at least $25.2 billion from August 2022 to August 2023 (Stanford University Institute for Human-Centered Artificial Intelligence, Artificial Intelligence Index Report 2024, 2024). Providers of hardware for AI report capital of even greater amounts, with Microsoft Corporation expecting to spend $80 billion this fiscal year (Rafe Uddin, "Microsoft Sheds $150bn in Market Value After Cloud Sales Disappoint," Financial Times, January 30, 2025). Return to content ⤴
23. See Akhil Thadani and Gregory C. Allen, "Mapping the Semiconductor Supply Chain: The Critical Role of the Indo-Pacific Region," brief, Center for Strategic and International Studies, May 2023, which shows that much of the international value chain for semiconductors is distributed globally, including to U.S. allies. Return to content ⤴
24. Peter Grindley, David C. Mowery, and Brian Silverman, "SEMATECH and Collaborative Research: Lessons in the Design of High-Technology Consortia," Journal of Policy Analysis and Management, Vol. 13, No. 4, Autumn 1994. Return to content ⤴
25. AI Action Summit, International AI Safety Report: The International Scientific Report on the Safety of Advanced AI, January 2025. Chapter 2.2.3 discusses loss-of-control scenarios and potential paths that might lead to the occurrence of such scenarios. Return to content ⤴
26. Khalid Khan, Chi-Wei Su, Muhammad Umar, and Weike Zhang, "Geopolitics of Technology: A New Battleground?" Technological and Economic Development of Economy, Vol. 28, No. 2, 2022. Return to content ⤴
27. See Audra J. Wolfe, Competing with the Soviets: Science, Technology, and the State in Cold War America, John Hopkins University Press, 2013, which explores the history of scientific competition between the United States and the Soviet Union. Return to content ⤴
28. Alexei Grinbaum and Laurynas Adomaitis, "Dual Use Concerns of Generative AI and Large Language Models," arXiv, arXiv:2305.07882, 2023. Return to content ⤴
29. Ronit Langer and Shruti Sharma, "The Blessing and Curse of Biotechnology: A Primer on Biosafety and Biosecurity," Carnegie Endowment for International Peace, November 20, 2020. Return to content ⤴
30. Peter Jones, Track Two Diplomacy in Theory and Practice, Stanford University Press, 2015. Return to content ⤴
31. Francis J. Gavin, "Blast from the Past: Proliferation Lessons from the 1960s," International Security, Vol. 29, No. 3, Winter 2004/2005. Return to content ⤴
32. AI Action Summit, 2025. Chapter 2.2.3 lays out some of the misalignment risks. Return to content ⤴
33. One example of such an agreement is the Treaty on the Non-Proliferation of Nuclear Weapons by the governments of the United Kingdom, the Soviet Union, and the United States of America, July 1, 1968. Return to content ⤴
34. For an example of such speculation, see Herman Kahn and Anthony J. Weiner, The Year 2000: A Framework for Speculation on the Next Thirty-Three Years, Macmillan, 1967. Return to content ⤴
35. For an example of how AI could enhance social control, see Simon Marvin, Aidan While, Bei Chen, and Mateja Kovacic, "Urban AI in China: Social Control or Hyper-Capitalist Development in the Post-Smart City?" Frontiers in Sustainable Cities, Vol. 4, 2022. Return to content ⤴
36. Jinghan Zeng, "Artificial Intelligence and China's Authoritarian Governance," International Affairs, Vol. 96, No. 6, November 2020. Return to content ⤴
37. Ronald Wintrobe, The Political Economy of Dictatorship, Cambridge University Press, 1998. Return to content ⤴
38. Richard Ngo, Lawrence Chan, and Sören Mindermann, "The Alignment Problem from a Deep Learning Perspective," arXiv, arXiv:2209.00626, 2024. Return to content ⤴
39. For evidence of reinforcement learning systems learning unintended goals, see Rohin Shah, Vikrant Varma, Ramana Kumar, Mary Phuong, Victoria Krakovna, Jonathan Uesato, and Zac Kenton, "Goal Misgeneralization: Why Correct Specifications Aren't Enough For Correct Goals," arXiv, arXiv:2210.01790, 2022. For evidence of how autonomous agents would seek power, see Victoria Krakovna and Janos Kramar, "Power-Seeking Can Be Probable and Predictive for Trained Agents," arXiv, arXiv:2304.06528, 2023. Return to content ⤴
40. Kate Goddard, Abdul Roudsari, and Jeremy C. Wyatt, "Automation Bias: A Systematic Review of Frequency, Effect Mediators, and Mitigators," Journal of the American Medical Informatics Association, Vol. 19, No. 1, January–February 2012. Return to content ⤴
41. For examples, see Sam Bowman, "The Checklist: What Succeeding at AI Safety Will Involve," webpage, September 3, 2024; and Jan Leike and Ilya Sutskever, "Introducing Superalignment," OpenAI, July 5, 2023. Return to content ⤴
42. Jon Schmid, Tobias Sytsma, and Anton Shenk, Evaluating Natural Monopoly Conditions in the AI Foundation Model Market, RAND Corporation, RR-A3415-1, 2024. For more about how AI systems are trained to cooperate, see Max Jaderberg, Wojciech M. Czarnecki, Iain Dunning, Luke Marris, Guy Lever, Antonio Garcia Castañeda, Charles Beattie, Neil C. Rabinowitz, Ari S. Morcos, Avraham Ruderman, Nicolas Sonnerat, Tim Green, Louise Deason, Joel Z. Leibo, David Silver, Demis Hassabis, Koray Kavukcuoglu, and Thore Graepel, "Human-Level Performance in 3D Multiplayer Games with Population-Based Reinforcement Learning," Science, Vol. 364, No. 6443, May 2019. Return to content ⤴
43. Sumeet Ramesh Motwani, Mikhail Baranchuk, Martin Strohmeier, Vijay Bolina, Philip Torr, Lewis Hammond, and Christian Schroeder de Witt, "Secret Collusion Among AI Agents: Multi-Agent Deception via Steganography," NeurIPS 2024 Conference, September 2024. Return to content ⤴
44. For a more-detailed example of how collusion between AIs might occur, see Kokotajlo et al., 2025. Return to content ⤴
45. Herbert Feis, Road to Pearl Harbor: The Coming of the War Between the United States and Japan, Princeton University Press, 2016. Return to content ⤴
46. These were not the final eight scenarios developed and discussed previously. Instead, interviewees were presented with four initial scenarios that were then developed into the final eight based on their input and a literature review. Return to content ⤴
47. A full interview protocol can be found in Appendix B. Return to content ⤴
48. Martin Neil Baily, Erik Brynjolfsson, and Anton Korinek, "Machines of Mind: The Case for an AI-Powered Productivity Boom," Brookings Institution, May 10, 2023. Return to content ⤴

About the Authors

Related Content

• 

  Research

  On the Extinction Risk from Artificial Intelligence

  May 6, 2025

Topics

• Artificial Intelligence
• China
• Geopolitical Strategic Competition
• Science, Technology, and Innovation Policy
• United States