Statecraft for the AI Transition

Today we are announcing the Center for Technology and Statecraft (CTS), a new non-partisan policy research initiative based in Washington, DC. Our mission is to solve unexplored AI policy challenges through technical, forward-looking research.

We are, first and foremost, a research organization. Our aim is to produce novel, foundational research, rather than just responding to the issues currently being debated in DC.

Our initial research focus will be on two primary policy pillars. First, we will investigate how policy can shape virtual and physical automation over the next decade as we reimagine the social contract in light of advanced AI. Second, we will study how to manage long-term competition between nations in AI while ensuring stability between great powers.

Both of these pillars will build on a foundation of analysis of the AI value chain, from chipmaking tools and tokens to AI-powered robots and software agents that affect the physical and virtual worlds. Our value-chain research will also assess the inputs needed to produce AI, such as compute, algorithms, and data.

CTS will operate as an independent initiative supported by the Institute for Progress (IFP), a non-partisan innovation policy think tank.

Where the world is headed

Building AI now rivals the largest capital projects ever undertaken. US AI capital investment already accounts for close to 1% of US gross domestic product (GDP) and the majority of US growth in private investment.

These enormous investments have bought commensurate returns in capabilities. Since 2012, when the current era of “deep learning” AI algorithms began, exponential scaling of computing power and training data has enabled rapid capability gains. With the advent of reasoning models in 2024, capability progress has accelerated further.

Today’s resulting AI capabilities would have seemed like science fiction as little as five years ago. Over the past year, agentic AI coding tools have shifted a large and growing share of coding tasks to AI, with the engineer’s role shifting from writing code by hand to orchestrating teams of AI coding agents. AI models can now prove noteworthy mathematical conjectures. They can autonomously execute end-to-end cyber vulnerability discovery, substantially increasing the rate of discovery in the last eight months. And they can now accelerate synthetic biology including the design of novel viruses. Even if progress plateaus and AI develops for a time as a normal technology, the widespread deployment of these capabilities could cause large shifts in economic production and national power.

But AI capability progress will almost certainly continue. Industry will keep scaling the production of AI computing power,^{Sidenote 1} enabling further capability scaling. Frontier AI companies are increasingly using their own AI models to automate AI R&D and are projecting full recursive self-improvement (RSI) in the late 2020s, which would enable an even faster exponential increase in capabilities. Unless we prepare now, this could overwhelm public institutions’ capacity to respond.

With continued capability progress, AI could transform both the virtual and physical world. In the limit, virtual AI remote workers could perform virtually any task involving a human behind a computer screen.^{Sidenote 2} Widespread deployment of AI-powered robots could enable the manipulation of the physical world at scale, potentially allowing AIs to broadly substitute for human labor. AI could reshape the balance of power within and between nations. In short, AI could be a transformative technology, precipitating a societal transition comparable to or greater in scale than the agricultural or industrial revolutions.^{Sidenote 3}

History provides mixed evidence about what this could mean for human welfare. Like the agricultural and industrial revolutions, transformative AI promises to dramatically improve productivity, plausibly accelerating overall economic growth. But while the transition from farming to industry was a boon to human welfare, the previous transition from foraging to farming arguably reduced it. Likewise, depending on economic factors and policy choices, transformative AI might cause human wages to increase exponentially, stagnate, or even fall to zero. Whether humanity retains agency in a world with more powerful AI is an open question. In sum, technological progress alone does not determine whether and how technology improves human welfare.

Policy will shape the outcome

AI’s future effects on the world are not fixed. Outcomes will depend in substantial part on choices about institutions, markets, and the development and deployment of AI systems — choices mediated by policy. We believe policy can act as a catalyst, spreading AI’s gains widely, mitigating avoidable harms, and preserving meaningful human agency. Or, absent sufficient attention, AI progress may leave most people no better off, erode their place in society, and strengthen authoritarian states rather than democratic ones.

Filling a gap in the research ecosystem

Many of the challenges arising from these transformations will be technically complex, will unfold at scale and over a period of years, and will not easily fit within established policy paradigms. This creates a challenge for policymakers, who must be responsive to the pressing issues of the day and have limited time in which to formulate new analytical frameworks. Meanwhile, technically grounded, policy-relevant investigations into technology remain relatively uncommon.

CTS will help fill this gap by producing research that is technically grounded, aimed at novel long-term challenges still in a pre-paradigmatic state, and accessible to policymakers.

In doing so, CTS will follow the lead of other organizations — both inside and outside Washington. The research these organizations produced was unusually useful because it was technical, forward-looking, or both, and therefore opened up new fields of policy activity.

Starting in 2019, CSET contributed early work on AI chips and their supply chains, China’s chip industry, and their implications for national security policy well before these were salient issues in DC (and in greater detail than typical policy analyses). This work developed an in-depth picture of how semiconductors and the scaling of computing power would shape the future of AI and national security,^{Sidenote 4} adding structure to what before then was a pre-paradigmatic field that the national security policymaking community was only beginning to appreciate.^{Sidenote 5} Since around 2022, Epoch AI, MIT FutureTech, and SemiAnalysis have developed AI forecasting and trends analysis, which have become central to studying and managing AI’s effects. METR pioneered frontier AI safety frameworks and pre-deployment evaluations in 2022–2023, which are now widely adopted by leading US AI companies and national AI institutes and form the basis of several regulatory frameworks. A range of organizations seeded the research and policy agenda in 2023 for the dual-use convergence of AI and biotechnology. In 2024, RAND defined the playbook for securing AI models, which now shapes industry and government practice. Across 2023–2024, several leading technologists conceptualized defensive acceleration (def/acc), an approach to societal resilience in the face of AI progress, which IFP’s Launch Sequence has been operationalizing since 2025.

Ultimately, however, these types of work are rare in government-facing AI policy, which generally focuses on issues under current debate. A core motivation for CTS is scaling this kind of technical and forward-looking research, through a dedicated research team and agenda. Prioritizing research means we will have less time for commentary and may go longer than is typical for think tanks between publishing reports.

We are also thinking carefully about how to build a durably useful research organization as AI capabilities advance. In addition to hiring full-time researchers, we have dedicated budgets for compute and datasets to take advantage of frontier AI systems at scale.

Grounding research in the AI value chain

A core design principle of our process at CTS will be grounding our policy research in rigorous empirical work and deep technical analysis of the entire AI value chain. Our research will span empirical and technical questions across development, deployment, diffusion, and comparative national assessments, both for today’s AI systems and their supply chains. We will also explore emerging technologies and approaches, such as alternative AI hardware architectures, robotics and embodied AI, emerging data center and system architectures, and AI verification technologies. The work will also analyze the factors of AI production, including both direct factors (compute, algorithms, and data), as well as supporting factors (investment, talent, and energy).

We will focus on the value chain for three reasons:

First, it is important: much of how the future of AI unfolds depends on how the value chain evolves. The final output — such as an AI system — is a function of many inputs. Policy that responds only to the final output — such as regulation of frontier AI systems — neglects the upstream levers, like export controls, where much of the leverage lies. Analysis that ignores the physical and economic stack misses these levers.

Second, it is predictable: The structures involved are large-scale, physical, and slow to change, with predictable underlying regularities. These structures have an empirical depth amenable to large-scale data gathering, integration, and modeling, enabling long-range predictions about the future and how policy actions can reshape those futures. As an example, Moore’s Law and semiconductor industry roadmaps provide the structure that enables long-range chip supply-chain analysis. Many other aspects of the broader AI value chain have a similar kind of structure.

Third, it is productive: research on the value chain reliably generates insights. These systems often have fractal-like complexity, where increasing the depth of investigation reveals new policy-relevant structure. There are detailed empirical facts and trends to discover at each level, which means time spent on this research is rarely wasted.

Policy pillars

Building on our value-chain analysis, every major policy research investigation at CTS will be organized around a strategy, rather than around policy tools (e.g., export controls or tax mechanisms) or specific technical questions. One way to solve policy challenges relating to AI is to develop and execute on an all-things-considered grand strategy, a concept traditionally used to describe US foreign policy. However, defining and operationalizing a grand strategy is complex. A more tractable approach is to begin with smaller component strategies — our policy pillars. Each pillar’s analysis begins with a strategic picture: what the current world-state is, how it will evolve by default, and how policy can steer it. That picture will inform which empirical questions to ask, especially as they relate to the AI value chain.^{Sidenote 6}

Initially, we are organizing our work around two policy pillars:

First, managing the transition to a world of advanced AI will likely require shaping the trajectory of the virtual and physical automation of society and reimagining the social contract to account for the effects of this automation. With continued capability advancements, AI may automate a growing share of tasks, with AI agents automating virtual work and embodied AI automating the physical economy, including key industrial supply chains. This process could occur quickly if post-RSI AI systems generalize well to new domains, but AI capabilities may remain jagged, with automation proceeding through a longer, iterative process that takes years.^{Sidenote 7} Even still, accelerating virtual and physical automation could rapidly expand society’s productive capacity and overall economic growth,^{Sidenote 8} potentially enabling an industrial explosion in the limit in which AIs, operating as agents or embodied in robots, become gross substitutes for human labor, allowing capital to accumulate with no humans in the loop. It could also have large and destabilizing effects on core aspects of the modern social contract, including employment, distribution of resources, and market freedom and competition.

Managing the trajectory of AI-enabled automation, therefore, will soon represent a defining policy challenge. Our work in this area will assess the volumes and characteristics of AI labor and the potential and realized consequences of AI-driven automation. Drawing on this analysis, we will assess options for reimagining aspects of the social contract to address the consequences of automation. The social contract here refers to the terms of cooperation, standing, and sharing of benefits in society, including whether and how AIs figure as agents within it, and how those terms can be set to maintain human agency and welfare. We may assess policy options to keep humans in the societal loop, including through tax, corporate, and IP law, and explore the degree to which policy should recognize AIs as economic agents over time. We will examine mechanisms for preserving competition, dynamism, and pluralism, including through taxation, innovation and regulatory policy, and ownership structures. And we will study how policy can distribute benefits, domestically and internationally, while managing rapid disruptions. As part of this work, we will also explore how policy should shape robotics and industrial supply chains, their development trajectory, and how AI-powered robots could accelerate AI development and enable self-sustaining AI supply chains.

Second, we will study the long-term strategy for strategic AI competition with China while maintaining stability. US strategy has coalesced around a bipartisan consensus on industrial policy, export controls, and other competition measures to maintain America’s technology advantage, but also on the need to engage with China on AI risks that cross borders. However, many basic questions underpinning the US-China AI balance are not fully understood. For one, we lack a deep quantitative picture of how the inputs to AI progress have contributed to relative US-China model capabilities and deployment scale and have even less understanding of the key determinants of diffusion in the economy and the national security enterprise, including through robotics and embodied AI. For another, China represents a long-term pacing challenge to US security and standing on the global stage. The policy apparatus must respond to projected changes in AI development and its technical paradigms, effects of export controls, both countries’ integration of AI into the physical world, and the broader US-China relationship, all over a long time horizon extending well into the 2030s.

We aim to develop long-term US-China capability assessments and generate new policy ideas for ensuring democracies maintain and extend their lead in AI. In doing so, we will explore how the US can prudently use the lead to reduce catastrophic risks, including by reducing the risk of great-power conflict and ensuring space for coordination and cooperation on AI safety, stability, resilience, and verification. This includes creating option value for coordinated crisis response if we see rapid, transformative jumps in capability resulting from RSI.^{Sidenote 9}

Table 1 summarizes our initial research priorities.

Table 1: Research priorities

The need for flexible mindsets

Our research problems span different policy domains, but society needs coordinated solutions to all of them to respond to a fully general-purpose technology like AI. We believe that realizing robust solutions will require moving between two mindsets that are common in AI policy — we call them the “security mindset” and the “diffusion mindset.” (Table 2 describes how these mindsets approach problems in different ways.)

Put simply, the security mindset orients towards the possibility that the most important AI effects may be sudden, discontinuous, and irreversible, thereby overwhelming governing institutions’ capacity to respond. It concludes that the most important AI policy questions therefore involve shaping the behavior of a few strategic actors (e.g., leading governments and AI companies) as they develop and deploy increasingly capable frontier AI systems, which could pose discrete, sudden risks or confer decisive advantages.

Conversely, the diffusion mindset orients primarily towards the possibility that the most important AI effects may arise through a coevolutionary process with existing institutions, particularly economic markets. It focuses on policy questions related to shaping the behavior of many decentralized actors (e.g., firms and consumers) as AI diffuses throughout society, often with a focus on ensuring widely shared prosperity and preserving human agency and welfare.

In reality, many thinkers combine dimensions of these different mindsets, breaking the binary. Nevertheless, the dimensions associated with these mindsets frequently correlate.

Table 2: Mindsets

The security mindset is common among frontier AI companies, national security policymaking circles, and the AI safety community. It is a useful frame for our competition and stability focus area, as it relates to China policy, national security, and discrete, sudden catastrophic risks. Meanwhile, the diffusion mindset is common among the business community, antitrust and labor policy circles, and academia. It is a useful frame for our automation and social contract focus area and is relevant to labor and social policy, economic growth, innovation policy, assessing the speed of and bottlenecks to AI-enabled automation, and managing risks that manifest gradually at a systems level.

We believe that solving problems across our core focus areas will require sampling from both mindsets, moving between them as foxes rather than siloing as hedgehogs. For example, in strategic competition, the diffusion mindset helps weigh security objectives against their second-order effects, which may be less visible but more significant than their direct policy impact. And in automation, innovation, and economic policy, the security mindset may assist with understanding step-change effects resulting from surprising and fast-moving AI capability increases. And more broadly, managing the AI transition will require an interlocking strategy for solving problems across focus areas, all at once.

Maintaining a flexible mindset will help us to adapt as we learn more about the shape of AI’s trajectory. We live now in a state of profound uncertainty about the nature and scale of AI’s effects. This means we must adapt to new evidence, while pursuing policy options that are beneficial across different futures.

What’s next

In the coming weeks, we will publish our first major investigations into the AI supply chain, the drivers of AI capabilities, and China policy. If you are a researcher, technologist, or policymaker who is interested in these problems, we are building our team and would be excited to chat about these topics. You can follow us at techstatecraft.org, techstatecraft.substack.com, and x.com/techstatecraft.