September 2025

Rethinking a Political Approach to Nuclear Abolition
George Perkovich, Fumihiko Yoshida, and Michiru Nishida
Carnegie Endowment for International Peace, 2025

This book asserts that addressing the accelerating expansion of nuclear arms, growing nuclear risks and deepening geopolitical divisions requires strategies grounded in political-security realities more than the qualities or quantities of weapons. Notably, the authors acknowledge that their proposal may frustrate both disarmament and deterrence proponents. As they write, “In focusing on people and politics first, rather than the numbers and types of weapons and plans for their use, [… the book] will assume that preventing nuclear war and achieving nuclear disarmament are politically more difficult than the most avid proponents and opponents of nuclear weapons tend to recognize.”

Noting the difficulty of preventing nuclear war and further nuclear armament, the authors stress the “vital role of high-level political dialogue in advancing nuclear disarmament.” They say that the pragmatic approach is to acknowledge the “reality of nuclear deterrence while working to reduce both intentional and accidental risks.” The book outlines the unstable dynamics of the current nuclear landscape. It also explores three paths forward—abolition of nuclear weapons, “unstabilized” deterrence, or stabilized competition—and makes the case for the middle-ground option, stabilized competition, as the most pragmatic near-term solution “so long as politics preclude abolition.” Drawing on past arms control lessons, the authors articulate one large goal—to end nuclear overkill—and six guidelines to make deterrence more stable and accountable.

Despite little hope for disarmament, the authors argue that “it is possible that societies and some leaders will recognize the unnecessary danger of current trends and begin laying the groundwork for mutual restraints to 
be built when political changes allow more reasonable policies.”—SHIZUKA KURAMITSU

Beyond the Euromissile Crisis: Global Histories of Anti-Nuclear 
Activism in the Cold War
Luc-André Brunet and Eirini Karamouzi (eds.)
Berghahn Books, 2025

This open-access volume, funded by the UK Arts and Humanities Research Council, expands the scope of historical research on anti-nuclear activism during the Cold War beyond the well-known North Atlantic movements. The editors argue that the focus in existing literature on the Euromissile crisis of the 1980s and Intermediate-Range Nuclear Forces Treaty insufficiently represents the international reach of anti-nuclear movements in the late 20th century.

The edited volume collates 13 chapters and contributes to the field in four main ways: It broadens the geographical scope and includes regions that have been underrepresented in the discourse, such as Yugoslavia and French Polynesia; it highlights the connections and divisions across transnational movements, including disagreements on the definition of peace; it reassesses the established Cold War chronology by questioning the relevance of the Euromissile crisis for global movements; and it analyzes the relations between state and nonstate actors in the form of policymakers and peace movements. Overall, the book shows the diversity of movements, demonstrating that anti-nuclear activists followed no uniform practice in their campaigns.—LENA KROEPKE

September 2025
By Lt. Gen. John "Jack" N.T. Shanahan

As evidenced by the November 2024 agreement between Chinese President Xi Jinping and U.S. President Joe Biden, there is growing international consensus that artificial intelligence (AI) must never supplant human judgment in the authorization or execution of nuclear weapon launches.¹ Such momentous decisions should remain the sole domain of human leaders.

Although this nonbinding consensus is encouraging, it overshadows a more complex and less understood challenge, namely, that the integration of AI across the nuclear command and control/nuclear command, control, and communications (NC2/NC3) enterprise and beyond may create false confidence in the information that is shaping leaders’ situational awareness and influencing nuclear-related decisions.

Leaders of nuclear-weapon states must demand a thorough examination of the risks resulting from AI integration into all elements of the nuclear decision-making process. This analysis should encompass not just the traditional, well-documented components of the NC2/NC3 enterprise, but also all ancillary conventional platforms, sensors, intelligence systems, and information technology networks. Even seemingly unrelated applications, such as computer vision models deployed in Project Maven² or large language models (LLMs) used for problem scoping and mission analysis, can amplify the effects of other inputs that shape nuclear-related decisions.

Some incremental effects might not differ substantively from the pre-AI era, but the consequences of compounding failures in worst-case scenarios are potentially catastrophic. This reality underscores the need for a deep understanding of how AI affects every system, platform, and process that intersects with NC2/NC3. As no current modeling or simulation can capture the complexity of the entire NC2/NC3 enterprise during a real crisis, exercises, wargames, and tabletop simulations are crucial, yet insufficient. There may be no better case for applying the precautionary principle than when considering the role of AI in nuclear operations.

Definitions

Experts in nuclear policy, AI, and arms control have begun to analyze the opportunities, risks, and potential benefits of AI integration in the nuclear enterprise,³ but few studies offer a comprehensive assessment across the full spectrum of systems and networks that influence nuclear decision-making. Given AI’s still-nascent role in national security across nuclear-armed states, this gap reflects both the complexity of such analysis and the vast number of unknowns, to include many “unknown unknowns.” Nevertheless, the accelerating pace of AI integration underscores the urgency of such assessments.

The act of employing a nuclear weapon is often portrayed as a singular, dramatic decision made by the head of state during an escalating crisis or conflict. In reality, this decision is intended to be the culmination of a complex, all-domain process involving numerous interdependent systems. In some scenarios, such as launch-on-warning, timelines may compress to mere minutes. In others, leaders may have days or weeks to deliberate.

A critical first step in understanding this process is to distinguish clearly between NC2 and NC3, two terms often conflated under the broader “nuclear enterprise” umbrella. According to U.S. doctrine, NC2 is defined as the “exercise of authority and direction, through established command lines, over nuclear weapon operations by the President as the chief executive and head of state.”⁴ Its core requirements are that it must be assured, timely, secure, survivable, and enduring—delivering information and communications that enable presidential decision-making throughout a crisis.⁵ As in all U.S. military interpretations of command and control, the tenets of authority, direction, and control reside at the core of NC2.

NC2 comprises five mission-essential functions: force management, planning, situation monitoring, decision-making, and force direction. Force management refers to the assignment, training, deployment, maintenance, and logistic support of nuclear forces and weapons. Planning is defined as the development and modification of plans for the employment of nuclear weapons. Situation monitoring is defined as the collection, maintenance, assessment, and dissemination of information on forces and possible targets, emerging nuclear powers, and worldwide events of interest. Decision-making comprises the assessment, review, and consultation for employment of nuclear weapons. Force direction refers to the implementation of decisions regarding the execution, termination, destruction, and disablement of nuclear weapons.⁶

NC3, by contrast, refers to how those functions are executed: the means “through which Presidential authority is exercised and operational command and control of nuclear operations is conducted.”⁷ NC3 is part of the broader National Leadership Command Capability., It consists of the three broad mission areas of presidential and senior leader communications, NC3, and continuity of operations/continuity of government. NC3 encompasses facilities, equipment, communications, procedures, and people.

Together, NC2 (the ways) and NC3 (the means) are critical to the proper functioning of the U.S. National Military Command System (NMCS), which supports the desired ends: continuous, survivable, and secure nuclear command and control. The NMCS relies upon ground, airborne, maritime, space, and cyber systems to provide “unambiguous, reliable, accurate, timely, survivable, and enduring” warning about attacks on the United States, its allies, and its forces overseas.⁸

Despite these clear doctrinal definitions, even the most comprehensive descriptions of the NC2/NC3 enterprise fail to capture the full breadth of systems involved; the U.S. Air Force alone owns over 100 separate NC3 systems. They also neglect the staggering number of non-NC3 networks, platforms, sensors, and systems across all the military services that, in various ways, feed into, interface with, or could interface with the wider nuclear enterprise. This blurring of lines is even more pronounced in some other nuclear-armed states. Given the current trajectory, over the next decade AI is likely to be integrated not only into conventional military and intelligence platforms and networks, but also into NC3 systems. AI also is expected to play a larger role in supporting NC2 decision-making processes. A clear understanding of these definitions and their interconnections is essential to evaluating the implications of AI integration across the nuclear enterprise.

AI, NC2/NC3, and Strategic Stability

Scholars have argued that incorporating AI into NC2 processes and NC3 systems will introduce destabilizing dynamics among nuclear-armed states, particularly by diminishing direct human involvement in decisions about nuclear weapons use. Yet many of the same experts also recognize potential stabilizing effects in certain contexts. For example, by enhancing early warning and situational awareness, enabling more consistent detection and tracking of adversary systems, providing advanced tools for monitoring and verifying arms control and disarmament, and improving warhead and weapon safety, security, and reliability.

On balance, however, the risks to strategic stability⁹ from significantly accelerating nuclear decision timelines or reducing human involvement in launch decisions and execution are likely to outweigh the potential benefits.¹⁰

One especially destabilizing scenario involves the widespread application of AI to detect and continuously monitor the ground, sea, and airborne nuclear forces of other states. Adversary integration of AI-enhanced systems that can reliably locate mobile missile launchers or submerged ballistic missile submarines, which are considered the most survivable elements of a second-strike capability, could erode confidence in a state’s assured retaliation posture. This loss of confidence could in turn increase incentives for preemptive strikes, weakening the logic of deterrence and undermining strategic stability. The problem becomes even more acute if such surveillance and tracking technologies are paired with AI-enhanced automation of decision-making or launch processes, heightening first-strike incentives during a crisis.

At present, there is no evidence that any nuclear state, including the United States, intends to integrate AI directly in its NC3 systems or formal NC2 decision-making procedures. Nevertheless, in the absence of bilateral or multilateral agreements or established international norms circumscribing AI use, the likelihood that a state will take such steps will increase, while indirect integration into nuclear-adjacent systems is all but certain. Such integration may generate cascading effects that subtly and dangerously influence nuclear decisions, particularly if they remain undetected. Moreover, the mere perception that a state is integrating AI into its NC2/NC3 enterprise, or contemplating reduced human oversight over nuclear decisions, could itself be destabilizing. Given the uncertainty surrounding current and emerging AI capabilities, managing not only deployment but also perception will be a crucial component of strategic stability in the years ahead.

Indirect Pathways and Unpredictable Behaviors

The potential for AI to augment, accelerate, and automate tasks across the national security domain will lead inexorably to its expanded adoption. Although the expected benefits, such as greater speed, accuracy, effectiveness, and efficiency, may improve military operations, AI’s novel characteristics also introduce unique risks. These go well beyond automation of nuclear weapon launch decisions. In addition to the concerns already outlined, four compounding and interconnected areas demand urgent attention to avoid unintended or mistaken use of nuclear weapons.¹¹

First, the integration of AI across the NC2/NC3 enterprise introduces the risk of cascading effects and emergent behaviors. Errors resulting from flawed models or adversarial manipulation could propagate across interconnected systems, especially in the absence of sufficient human oversight.¹² Even seemingly minor downstream errors could produce disproportionate upstream consequences once multiple AI models are embedded across interconnected platforms and decision-support systems.

Compounding the danger is automation bias: the tendency to over-trust machines, particularly under crisis conditions marked by time compression, ambiguity, and extreme stress.¹³ These pressures could amplify the effects of unnoticed errors, producing tightly coupled feedback loops that distort the decision-making environment. Adversaries likely will exploit these vulnerabilities through AI-enhanced information operations designed to influence, disrupt, and corrupt the judgment of senior political and military leaders. For example, adversaries could use deepfake audio or video to fabricate crisis-related statements by senior political or military leaders or deploy LLMs to generate tailored disinformation campaigns that manipulate strategic warning indicators, public sentiment, or even internal military communications.

Emergent behavior poses a related concern. In AI, emergence refers to novel and unpredictable outcomes that arise from the interaction of relatively simple components—outcomes not explicitly programmed into the system.¹⁴ Ensembles of advanced AI models operating in dynamic, nonlinear systems may produce opaque and surprising results. Such unpredictability is particularly dangerous in nuclear operations, where the standard for accuracy and reliability is absolute. Former U.S. Navy Secretary Richard Danzig warned in Technology Roulette that “the introduction of complex, opaque, novel, and interactive technologies will produce accidents, emergent effects, and sabotage.” He cautioned that U.S. national security institutions risk losing control over their own creations.¹⁵ Although there is no evidence that any true self-learning AI systems have yet been fielded, developments are advancing rapidly, and the demonstration of emergent behaviors may only be a matter of time.¹⁶

Second, even if AI is monitored carefully within the formal NC2/NC3 enterprise, its use in adjacent systems—such as intelligence and surveillance sensors, conventional weapon systems, related information networks, and decision support systems—will indirectly influence nuclear decision-making. The same risks of cascading effects and automation bias apply, but with an added concern: Nuclear decision-makers may be unaware that AI is shaping their understanding of the operational environment. Comprehensive assessments of AI’s integration into conventional systems are essential, however complex those analyses may be. These AI-enabled assessments should map potential influences, ranging from most likely to most dangerous, and be shared with senior civilian and military leaders.

Third, large LLMs such as those powering ChatGPT or Perplexity may seem distant from the nuclear enterprise, yet arguably they pose a more immediate and underestimated risk. Generative AI (GenAI) tools are now being piloted across U.S. federal agencies. The Trump administration’s AI executive order calls for decisive action to ensure U.S. leadership in this field.¹⁷ Some departments and agencies have approved some models, tailored for government use, for experimentation and pilot projects. Wider adoption is inevitable.

Yet despite their rapid improvement, today’s open- and closed-source LLMs, to include the more advanced frontier and reasoning models, remain too brittle, skewed, and opaque to be trusted with mission-critical tasks like intelligence analysis for nuclear decision-making. Although this may seem obvious, there is a real risk that GenAI systems, used outside the NC2/NC3 enterprise, could influence assessments that ultimately shape nuclear decisions.¹⁸ At a minimum, U.S. national security leaders should require explicit disclosure of any use of GenAI in assessments that support nuclear course-of-action development. Even when GenAI is used for non-nuclear purposes, its outputs may shape broader assessments that feed into nuclear planning processes, sometimes without explicit attribution.

Finally, the integration of AI agents into commercial systems is advancing rapidly, and their eventual adoption in military operations and intelligence analysis is highly likely.¹⁹ These agents, capable of operating autonomously or semi-autonomously across systems and networks, pose enormous risks to nuclear operations. Their ability to take initiative, act without direct human intervention or even oversight, and learn over time creates scenarios in which behaviors become difficult to predict or control.

Cyber vulnerabilities and adversarial attack only compound the threat. As with cascading effects and emergent behavior, the risk posed by AI agents demands rigorous and detailed study. Mitigation must begin with the imposition of clear, human-defined constraints. With advanced AI, it will be far more difficult to enforce what agents must not be allowed to do than to direct them toward specific tasks.²⁰ Yet both types of constraints are essential if agents are to be integrated safely, if ever, within the NC2/NC3 ecosystem.

Circumscribing AI’s Nuclear Role

Although formal international agreements proscribing AI from authorizing or executing nuclear launches are unlikely in the near term, the 2024 Biden-Xi agreement and the emerging informal global consensus mark an important beginning. Still, these are only opening moves. It remains to be seen whether and how current or future administrations will follow through.²¹ These issues warrant candid discussion across Track 2, Track 1.5, and Track 1 dialogues, as well as in broader international forums. Depending on the state of relations between nations, these discussions could take place under the auspices of AI dialogues, arms control dialogues, or broader state-to-state dialogues that include an AI-nuclear nexus.

Leaders of nuclear-weapon states should issue clear public statements affirming the necessity of human oversight and control in all aspects of nuclear weapon deployment and employment. For example, as Gen. Anthony Cotton, the commander of U.S. Strategic Command, stated in his testimony to the Senate Armed Services Committee Strategic Forces Subcommittee in March 2025, “USSTRATCOM will use AI/ML to enable and accelerate decision-making.… AI will remain subordinate to the authority and accountability vested in humans.”²²

The potential benefits of AI for national security are significant, and its diffusion rate will only accelerate. The gravest risks of AI in the nuclear enterprise, such as automation of launch authority, are well understood and relatively easy to proscribe. But other risks are more subtle and systemic. The goal is not to halt AI adoption, but to deepen understanding of its unintended consequences, especially cascading effects, emergent behaviors, and hidden influences on intelligence and decision-making. Accompanying analyses, to include rigorous red teaming, should drive actions to mitigate the risks of indirect pathways and unpredictable behaviors while setting the stage for rapid human intervention as soon as unexpected or undesirable behaviors are detected.

In a real crisis, there may be no time to question how AI shaped the intelligence and analysis that informed nuclear recommendations. It is imperative, therefore, that senior military and intelligence leaders initiate detailed studies now, with the expectation that the studies will need continual refinement as AI continues to rapidly advance. These studies should be supported by experts within commercial technology companies and academic institutions and must address not only AI within NC2/NC3, but also within adjacent systems that may exert indirect influence. They also should assess the current state of human-machine integration, especially at the decision-support function, emphasizing the importance of the interfaces and interdependencies between humans and smart machines across the NC2/NC3 ecosystem.

In many instances, AI may add clear value with minimal risk. For instance, AI will be highly valuable in the behind-the-scenes work of warhead design, weapon safety and security, and scientific discovery. However, thorough evaluations might reveal instances where the potential cumulative effects of AI integration are so adverse that guardrails, other proactive actions, or outright prohibitions are necessary to preclude any possibility that a nuclear weapon is launched in error.

As AI continues to advance, it will match or exceed human performance across more national security functions. At the same time, humans are and will forever remain highly fallible. The juxtaposition of these two claims should not lead reflexively to the conclusion that AI is always the right answer. War is the ultimate human endeavor, and nuclear war its most consequential form. What sets humans apart—deductive and causal reasoning, contextual understanding, common sense, emotional awareness, and moral judgment—cannot be replicated by machines.

Ultimately, the decision to launch nuclear weapons must remain a distinctly human responsibility. National decision-makers, whether in the United States or any other nuclear- weapon state, must understand what role, if any, AI played in shaping the intelligence and recommendations presented to them. In that sense, the 2024 Biden-Xi agreement can be viewed not as the end of the debate, but as an inflection point marking the start of a global reckoning with AI’s role in nuclear stability.

ENDNOTES

1. The White House, Readout of President Joe Biden’s Meeting with President Xi Jinping of the People’s Republic of China, November 16, 2024. Beyond the Biden-Xi agreement, the broad international consensus remains an agreement in principle rather than in practice. For instance, the original draft of the State Department’s Political Declaration on Responsible Military Use of Artificial Intelligence and Autonomy contained language asserting that, as a best practice, governments “should maintain human control and involvement for all actions critical to informing and executing sovereign decisions concerning nuclear weapons employment.” This language does not appear in the final published version of the Declaration; in fact, there are no references at all to nuclear weapons or NC2/NC3. The Principles and responsible practices for Nuclear Weapon States, a 2022 U.N. working paper submitted by France, the United Kingdom, and the United States, states that “Consistent with long-standing policy, we will maintain human control and involvement for all actions critical to informing and executing sovereign decisions concerning nuclear weapons employment.” AI is not mentioned.

2. Project Maven is the name commonly used for the U.S. Department of Defense Algorithmic Warfare Cross-Functional Team, established in 2017.

3. See, for example, Johnson; Qi; Boulanin, et al.; Li; Kania; Depp and Scharre; Bajema; Schwartz and Horowitz; and Stokes, Kahl, Kendall-Taylor, and Locker:

James Johnson, AI and the Bomb: Nuclear Strategy and Risk in the Digital Age, Oxford University Press, 2023. Haotian Qi, “Chinese Perspective,” in Enhancing US-China Strategic Stability in an Era of Strategic Competition: US and Chinese Perspectives, edited by Patricia M. Kim, United States Institute of Peace, April 2021. Vincent Boulanin, Lora Saalman, Petr Topychkanov, Fei Su, and Moa Peldán Carlsson, “Artificial Intelligence, Strategic Stability and Nuclear Risk,” Stockholm International Peace Research Institute, June 2020. Xiang Li, “Artificial Intelligence and its Impact on Weaponization and Arms Control,” in The Impact of Artificial Intelligence on Strategic Stability and Nuclear Risk, edited by Lora Saalman, Stockholm International Peace Research Institute, October 2019. Elsa Kania, “The AI Titans’ Security Dilemma,” Hoover Institution, October 29, 2018. Michael Depp and Paul Scharre, “Artificial Intelligence and Nuclear Stability,” War on the Rocks, January 16, 2024. Natasha Bajema, “Will AI Steal Submarines’ Stealth?” IEEE Spectrum, July 16, 2022. Joshua Schwartz and Michael Horowitz, “Out of the Loop: How Dangerous is Weaponizing Automated Nuclear Systems?” arXiv, May 1, 2025. Jacob Stokes, Colin Kahl, Andrea Kendall-Taylor, and Nicholas Lokker, “Averting AI Armageddon: U.S.-China-Russia Rivalry at the Nexus of Nuclear Weapons and Artificial Intelligence,” Center for a New American Security, February 2025.

4. Deputy Assistant Secretary of Defense of Defense for Nuclear Matters, Nuclear Matters Handbook 2020 [Revised] (unofficial).

5. Ibid.

6. Department of the Air Force Curtis E. Lemay Center for Doctrine Development and Education, Air Force Doctrine Publication (AFPD) 3-72 Nuclear Operations, Nuclear Command, Control, and Communications, December 18, 2020.

7. Secretary of the Air Force Air Force Instruction 13-550, Air Force Nuclear Command, Control, and Communications (NC3), April 16, 2019.

8. Nuclear Matters Handbook 2020

9. For this paper, I use James Acton’s definition of strategic stability: “A deterrence relationship is stable if neither party has or perceives an incentive to change its force posture out of concern that an adversary might use nuclear weapons in a crisis.” I also accept his three broad interpretations of strategic stability: the absence of incentives to use nuclear weapons first (crisis stability) and the absence of incentives to build up a nuclear force (arms race stability); the absence of armed conflict between nuclear-armed states; and a regional or global security environment in which states enjoy peaceful and harmonious relations. (James Acton in Strategic Stability: Contending Interpretations, edited by Elbridge Colby and Michael Gerson, Carlisle, PA: Strategic Studies Institute, 2013, pp. 117-146.)

10. Colin Gray provides an alternative argument on the effects of technology on strategic stability. Although Gray did not address AI and the security dilemma, he analyzed the impacts of earlier technologies on strategic stability. He notes that strategic stability can result “when there is a rapid change in technological generations and considerable unpredictability concerning the building programs of rivals, yet where a tolerable balance of military power is maintained – albeit almost exclusively through competition.” My counter to Gray’s argument is based on AI as an unproven technology, especially in military applications: the high levels of uncertainty associated with AI’s actual performance, when juxtaposed against the inflated statements about AI emanating from leaders in China and the United States, suggest that the kind of “unrestrained competition” associated with well-understood and well-established naval competitions in the later decades of the 19th century will not lead to strategic stability in the AI competition (at least not for the immediate future). When it comes to AI, we still do not know how much we do not know. (Colin Gray, “Strategic Stability Reconsidered,” Daedalus, Fall 1980, 109(4), pp. 135-154.)

11. There are additional novel differences between traditional and AI-enhanced military systems that must be considered when assessing the possible effects of the introduction of AI across the NC2/NC3 enterprise. These include the control challenge, black box challenge, and the accountability challenge (I deliberately use “challenge” rather than “problem,” since all three must be treated as inherent, systemic features of all AI).

12. These are both examples of AI “corruption,” defined as the deliberate or unintentional manipulation of the data, hardware, or software of an AI-enabled system that causes the system to produce missing, inaccurate, or misleading results, to deny or degrade the use of the system, or to force the system to expose hidden information used in the training or configuration of the AI component. (National Academies of Sciences, Engineering, and Medicine, Test and Evaluation Challenges in Artificial Intelligence-Enabled Systems for the Department of the Air Force, Washington, DC: The National Academies Press, 2023.)

13. However, I am equally wary of human bias, manifested in the form of disregarding or dismissing the contributions of machines in favor of human judgment, heuristics, or gut instinct. In the nuclear context, these kinds of biases can be equally calamitous.

14. TedAI San Francisco, October 21-22, 2025.

15. Richard Danzig, “Technology Roulette: Managing Loss of Control as Many Militaries Pursue Technological Superiority,” Center for a New American Security, June 2018.

16. Although LLMs demonstrate behavior that can be construed as “continual learning,” they are not true online learning systems. The debate continues as to whether they are stochastic parrots or something more advanced. See, for example, Steffen Koch, “Babbling Stochastic Parrots? A Kripkean Argument for Reference in Large Language Models,” Philosophy of AI, Vol. 1, 2025.

17. The White House, “Removing Barriers to American Leadership in Artificial Intelligence,” January 23, 2025. Examples of approved models include Claude Gov, Defense Llama, ChatGPT, NIPRGPT, and CamoGPT.

18. There have been only two instances of nuclear weapon use in conflict and relatively few global crises in which their use may have been seriously considered. As a result, the volume of real-world data available to train LLMs on nuclear decision-making is extremely limited. Although data from war games, simulations, experiments, and decades of nuclear deterrence and escalation theory will inform such models, any GenAI-derived responses to prompts related to nuclear weapons employment should be interpreted with great caution. For the foreseeable future, the human expert-smart machine combination will remain the only solution.

19. An AI agent is a software-based system capable of perceiving its environment, making decisions, and acting to achieve specified goals. They may combine sensing, reasoning, and learning capabilities to operate autonomously or semi-autonomously.

20. In the nuclear enterprise, the combination of emergent behavior and misalignment, in which LLMs and agents collaborate to achieve objectives beyond or contrary to human-defined boundaries, will be highly problematic. See, for example, Miles Wang, Tom Dupré la Tour, Olivia Watkins, Alex Makelov, Ryan A. Chi, Samuel Miserendino, Johannes Heidecke, Tejal Patwardhan, and Dan Mossing, “Persona Features Control Emergent Misalignment,” arXiv, June 2025.

21. The 2025 White House AI Action Plan and accompanying executive orders do not address the potential for follow-on AI dialogues between the United States and China, or with any other state.

22. Testimony of Gen. Anthony J. Cotton before the Senate Armed Services Committee Subcommittee on Strategic Forces, March 26, 2025.