"Asserting Human Control to Reduce the Dangers of AI and Nuclear War"

"Asserting Human Control to Reduce the Dangers of AI and Nuclear War"

Daryl G. Kimball, Executive Director

Prepared Remarks on the "Use of AI in Nuclear Weapon Systems: From Strategic Instability to Arms Control and Disarmament" at the Global Nobel Laureates Assembly on Artificial Intelligence and Nuclear War

Castel Gandolfo, Italy, July 14, 2026

Thank you to the organizers of this august assembly and to our hosts. It is an honor to join this important dialogue at this pivotal time in human history.

As the mission statement for this conference aptly notes: the great challenges of the twenty-first century --- including Artificial Intelligence (AI) and Nuclear Weapons -- can no longer be understood or governed in isolation.

AI integration in nuclear weapons operations is occurring at the same time we face a complex and growing array of nuclear weapons-related dangers.

Tensions between nuclear-armed states are on the rise. A qualitative nuclear arms race is already underway. The world’s nine nuclear-armed actors are squandering hundreds of billions of dollars to maintain and upgrade, and in the case of China, expand the size of their deadly arsenals.

Key agreements that have helped reduce nuclear danger -- so-called guardrails against nuclear catastrophe -- are gone or at risk due to inattention or incompetence. The last treaty capping the size of the two largest arsenals (New START) expired earlier this year.

Without new restraints, we could see the U.S. and Russia increase the size of their arsenals for the first time in decades, which could trigger an unconstrained global nuclear arms race.

More nuclear weapons will not make anyone, anywhere, safer. Each of the world’s nuclear-armed states have more than enough nuclear weapons to meet their subjective requirements for nuclear deterrence and to produce destruction on a global scale.

Meanwhile, some nuclear-armed states are once again threatening nuclear use to try to coerce and bully nonnuclear states.

In sum, the risk of nuclear war is greater than at any point since the end of the Cold War. Exacerbating these challenges and the risk of nuclear war is the rush to develop autonomous weapons and the integration of artificial intelligence into interconnected nuclear command, control, and communications systems (NC3), and possibly nuclear targeting.

AI integration in nuclear operations creates new uncertainties that will likely increase the risk of miscalculation due to the loss of effective human control over decisions involving the use of the world's most dangerous weapons.

Immediate and decisive action is required to counter these growing nuclear dangers before it is too late.

The Dangers of AI in Nuclear Weapons Systems

There is a growing body of research and scholarship that strongly suggests that unconstrained introduction of AI into nuclear weapons operations (command, control, communications, as well as nuclear planning and decision-making) will lead to greater nuclear instability and increase the risk of nuclear war.

Referring to both nuclear and conventional combat operations, the 2021 U.S. National Security Commission on Artificial Intelligence, predicted that “AI will make the process of finding and hitting targets of military value faster and more efficient.”¹

“Currently, this process generally involves passing data in a serial fashion from a sensor and through a series of humans, to a platform that can shoot at the target. AI will help automate some of the intermediate stages of the decision process.” As AI matures, it will propel “more advanced processes that would operate more akin to a web, fusing multiple sensors and platforms to manage complex data flows and transmitting actionable information to human operators and machines across all [combat] domains,” the report added.

In October 2024, the commander of U.S. nuclear forces said, “Advanced AI and robust data analytics capabilities provide decision advantage and improve our deterrence posture.” He called out the potential “for more effective integration of conventional and nuclear capabilities.”²

The current commander said in March this year, “Pursuant to the [Department of War’s] initiative to cultivate an “AI-first” workforce, AI and ML are now integral to USSTRATCOM’s mission, enabling the analysis and interpretation of vast data streams.”³

These and other developments create several serious risks:

AI integration in nuclear operations may seem attractive because it can help process more information more quickly, enabling a higher operational tempo. But the speed made possible may endanger appropriate human oversight of not only key decisions, but the plans that support them.
Automation could be even more useful, some AI advocates argue, by helping military commanders—up to and including the president—select nuclear and non-nuclear responses to confirmed indications of an enemy attack. With little time to act, human decision-makers could receive a menu of possible "adaptive targeting" countermoves devised by algorithms.
Another danger is "automation bias," the tendency for stressed-out decision-makers to trust the information and advice supplied by advanced computers rather than their own judgement.
The application of AI to detect and continuously monitor the ground, sea, and airborne nuclear forces of other states, and the potential threat posed by loitering autonomous weapons to these forces, could erode confidence in other states’ nuclear retaliatory capabilities, which could in turn increase incentives for them to launch preemptive strikes in a crisis or increase the size of their nuclear forces.
The use of AI to integrate disparate data sources to build a cohesive picture of ongoing military operations “may create false confidence in the information that is shaping leaders’ situational awareness and influencing nuclear-related decisions,” as the inaugural director of the U.S. Department of Defense Joint Artificial Intelligence Center, Lt. Gen. John "Jack" N.T. Shanahan warned in a 2025 article in Arms Control Today.
The integration of AI in nuclear operations risks creating errors resulting from flawed large language models or adversarial manipulation (a.k.a. data poisoning) or expanding the attack surface for targeted cyberattacks by an adversary on NC3 systems in a crisis.
Integration of nuclear and conventional C3 systems, as the Pentagon plans to do, creates "entanglement risks," including the possibility that a cyberattack on command-and-control systems related to conventional operations could be interpreted as endangering a state’s nuclear arsenal if those systems happened to be integrated. Nuclear-conventional integration also creates the risk of accidental use of nuclear weapons due to misconfiguration of software primarily intended for use with conventional systems and operations.
Finally, as many experts have noted, today's rapidly advancing "open- and closed-source LLMs remain too brittle, skewed, and opaque to be trusted with mission-critical tasks like intelligence analysis for nuclear decision-making." Such systems are capable of "emergent" behavior -- novel and unpredictable outcomes that arise from the interaction of relatively simple components—outcomes not explicitly programmed into the system.⁴

Such unpredictability cannot be tolerated in nuclear operations. We are already vulnerable to potentially irrational behavior from the handful of humans who ostensibly have the sole authority to order the use of nuclear weapons; we cannot allow computer hallucinations to complicate the equation.

Inadequate Responses to the AI-Nuclear Weapons Nexus

As the major nuclear-armed states, including the United States, move rapidly to integrate AI tools across the nuclear and conventional military systems and machinery, our political, diplomatic, and legislative responses to regulate these technologies are failing to keep pace.

The best we have is the growing international consensus that artificial intelligence (AI) must never supplant human judgment in the authorization or execution of nuclear weapon launches.⁵

The "human in the loop" principle, while necessary, is clearly insufficient to prevent all conceivable categories of accidents that could arise from AI integration in nuclear command, control, communications, and possibly nuclear targeting.

So far, other national and multilateral efforts to establish rules of the road and/or legally-binding restrictions on AI uses in the military domain have fallen well short of what is necessary to address the risk of AI in nuclear operations.

Furthermore, the major nuclear weapon states have, so far, failed to engage in sustained bilateral or multilateral discussions with one another on how they each plan to put the "human in the loop" principle into operational practice.

The world's major military powers have generally resisted efforts to negotiate legally-binding measures -- as many states have called for and as recommended by UN Secretary General Guterres -- to ensure that autonomous weapons and AI integration in military operations are consistent with the Law of War and International Humanitarian Law.

Instead, the United States and its allies have favored the adoption of voluntary guidelines through a "Political Declaration on the Responsible Military Use of Artificial Intelligence and Autonomy," first launched in 2023.

However, that effort dropped references to AI and nuclear weapons and human control. Diplomatic efforts to advance these principles have stalled under the second Trump administration, which is seeking to accelerate the adoption of autonomous weapons and AI in U.S. military operations.

Other multilateral efforts, particularly the “Responsible Artificial Intelligence in the Military Domain" (REAIM) process—with summits in The Hague in 2023, Seoul in 2024, and A Coruña in 2026—have done important work. But U.S., Russian and Chinese support is lacking and the wide scope of the discussions -- spanning AI safety, innovation, autonomous weapons, and more -- limit their ability to address the troublesome implications of AI in nuclear weapons operations.

Recent “informal exchanges” in accordance with General Assembly Resolution 80/58 of 2025, which proposed the consultation to address the risks and opportunities posed by the accelerating application of AI to military targeting and decision-making tools, have been similarly inconclusive.

The Nuclear Disarmament Deficit

For more than a decade, progress on disarmament diplomacy that could lead to new forms of restraint has stalled due to inattention, infighting and intransigence among major nuclear-armed states.

Earlier this year, the 2026 Review Conference for the Nuclear Nonproliferation Treaty (NPT) ended without reaching a consensus outcome or action plan for the third straight time. The NPT’s nuclear five methodically forced the removal of any passages that would have committed them to begin “negotiations on disarmament” or even simply to pursue discussions on an “urgent” basis.

Instead, the final draft outcome document vaguely “calls on the nuclear-weapon States to engage in … constructive dialogue on the basis of mutual respect … and notes that such engagement could facilitate future arms control discussions, and help progress towards nuclear disarmament”

The exchanges at the conference also revealed the discord among the nuclear five on how and whether to reduce nuclear risks and nuclear arsenals.

The U.S. delegation sought to focus attention on China’s nuclear buildup and claimed it had proposed “multilateral strategic stability talks” on “transparency, risk reduction, and nuclear testing.”

However, given the vagueness of the U.S. offer and the complexities of a five-sided negotiation involving states with different force sizes, force structures, nuclear postures, and strategic cultures, this could be a formula for further inaction on disarmament.

Multilateral negotiations on disarmament involving all five nuclear-armed NPT states, if launched, might produce meaningful results.

But such an initiative should not be allowed to substitute for the immediate commencement of serious bilateral talks between the United States and Russia and the United States and China on nuclear risk reduction, strategic stability, and nuclear arms reductions that could also yield concrete arms control and risk reduction outcomes, and perhaps more quickly.

To improve the chances of success in future arms reduction talks, whether multilateral or bilateral or both, and to prevent unconstrained nuclear competition, all five NPT-recognized nuclear-armed states should agree to a mutual and verifiable freeze of their strategic launchers at their current numbers.⁶

This first step could reduce tensions and improve the environment for sustained talks on concrete nuclear risk reduction measures, including understandings that mitigate the risk of AI in nuclear operations, and on disarmament.

Moving Beyond the "Human In the Loop" and Reversing the Nuclear Arms Race

The situation requires urgent action in the pursuit of a new, human-centered paradigm of peace and cooperative security.

First, we must recognize the limitations and risks of nuclear deterrence and the destabilizing effects that AI integration into nuclear systems and nonnuclear military systems may create.

Nuclear deterrence requires the capability to impose unacceptable costs and the resolve to use nuclear weapons. Without the belief in this resolve, nuclear deterrence theory does not work. Deterrence proponents argue that to deter nuclear attack, one must constantly strive to make the threat of nuclear weapons use more credible in all possible scenarios.

By this logic, nuclear-armed states must develop "more usable battlefield nuclear weapons" to fill every perceived deterrence "gap”, engage in exercises to wage and win a "limited" nuclear war, and from time to time make implied or explicit nuclear threats against adversaries in crisis situations.

In other words, nuclear deterrence is a scheme for making war less probable by making nuclear war more probable.

The reality is that when nuclear deterrence fails it will fail catastrophically. If and when nuclear weapons are used in a conflict involving another nuclear-armed adversary, there is no guarantee it will not result in all-out global thermonuclear war.

Pope Leo is correct when he calls out “the widespread yet erroneous belief that nuclear deterrence is an indispensable prerequisite for security” in Paragraph 194 of his new encyclical Manifica Humanitas.

AI introduces new uncertainties to the already uncertain deterrence equation.

Second, we must demand urgent action on concrete steps that:

move the nuclear weapon states and their allies away from their unhealthy and unsustainable reliance on the doctrine of nuclear deterrence
toward new agreements and understandings that halt and reverse the global nuclear arms race, and
put in place effective measures to head off the dangerous consequences of integration of AI in nuclear weapons systems and the conventional military systems that are entangled with their nuclear forces.

Regarding the use of AI for nuclear and conventional military purposes, we need enforceable and clear prohibitions that ensure meaningful human control, accountability, and democratic oversight.

Changeable national policies and executive orders are inadequate. The men who lead the small number of powerful companies who are profiting from the drive for AI innovation and implementation cannot be counted on, though they have a moral responsibility to be part of the solution.⁷

This requires new national legislation, effective regulation, and new international agreements and understandings. Some options include:

Nuclear-armed states must commit—either through coordinated action or in a binding agreement—to retain human control over any decision to use nuclear weapons and to insert automated, failsafe “tripwires” or "firebreaks" in advanced command-and-control systems to disallow action leading to or resulting in nuclear weapons employment or escalation without human approval. For example: AI algorithms to assist nuclear early warning and launch systems should remain physically disconnected from nuclear launch authority to prevent any possibility of an unintended AI-triggered nuclear exchange; AI-assisted software should not be used for the selection of nuclear war plan targets or for "adaptive nuclear targeting" in the event that nuclear weapons are used in a conflict.
Nuclear-armed states should agree to take concrete steps to move away from nuclear postures that call for prompt retaliatory nuclear counterattack and agree that the sole purpose of nuclear weapons is to deter nuclear attack.
Nuclear-armed states should regularly report to and engage with an expert, representative multilateral body (i.e. a standing group of governmental experts) on their use of AI-assisted tools in nuclear command and control systems, the measures they have implemented to mitigate nuclear risks, and to discuss the dangers these could pose and to consider further limitations on the use of such tools.
The international community should adopt binding rules requiring human oversight of autonomous weapons systems and strict compliance with applicable international law at all times. States with advanced militaries cannot be allowed to delay or defer such efforts any longer.

To address the growing danger of nuclear war and nuclear arms racing, we need serious sustained disarmament diplomacy. Tensions between the major powers are no excuse for further inaction.

Consistent with their NPT obligations, the five major nuclear-armed states must immediately engage in serious, sustained, bilateral and/or multilateral dialogue and negotiations on binding and voluntary measures to reduce nuclear risk, reduce the salience of nuclear weapons, and irreversibly and verifiably reduce their nuclear arsenals.
All states must call upon all nuclear-armed states to refrain from actions that further exacerbate tensions or lead to the unraveling of the global nuclear nonproliferation architecture, such as resuming nuclear explosive testing, building additional nuclear weapons and delivery systems, or withdrawing from nonproliferation agreements.

If we are to hold the increasingly multipolar world together and prevent it from sliding toward nuclear chaos, we must strengthen and rejuvenate the indispensable but beleaguered global nonproliferation and disarmament enterprise, avert a new nuclear arms race, jumpstart dialogue and diplomacy between the major powers on nuclear disarmament and the destabilizing effects of AI.

Thank you for your attention.⁸

Download the statement (in MS Word) here.

______________

ENDNOTES

1. National Security Commission on Artificial Security (NSCAI), Final Report, 2021, p. 80, https://www.nscai.gov/wp-content/uploads/2021/03/Full-Report-Digital-1.pdf

2. Carley Welch, “AI has role to play in protecting American nuclear C2 systems: STRATCOM head,” Breaking Defense, Oct. 28, 2024, https://breakingdefense.com/2024/10/america-needs-ai-in-its-nuclear-c2-systems-to-stay-ahead-of-adversaries-stratcom-head/

3. “Statement of Richard A. Correll, Commander, U.S. Strategic Command before the House Armed Services Committee on Strategic Forces,” March 17, 2026, https://armedservices.house.gov/uploadedfiles/2026usstratcomcongressionalposturestatement2026.pdf

4. Artificial Intelligence and Nuclear Command and Control: It’s Even More Complicated Than You Think, Arms Control Today, September 2025, by Lt. Gen. John "Jack" N.T. Shanahan.

5. For example, see the 2022 working paper for the UN First Committee titled "The Principles and responsible practices for Nuclear Weapon States," submitted by France, the United Kingdom, and the United States that declares 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;" the original 2023 draft of the State Department’s "Political Declaration on Responsible Military Use of Artificial Intelligence and Autonomy," which proposed that governments “should maintain human control and involvement for all actions critical to informing and executing sovereign decisions concerning nuclear weapons employment.” (the current published versions does not contain such language); the respective read-outs from the November 16, 2024 meeting between Presidents Biden and Xi, which reported that "The two leaders affirmed the need to maintain human control over the decision to use nuclear weapons" and "also stressed the need to consider carefully the potential risks and develop AI technology in the military field in a prudent and responsible manner;" the 2024 U.S. Congressional legislation that stated that it is U.S. policy to ensure that AI integration does not compromise “the principle of requiring positive human actions in execution of decisions by the President;” the Russian Federation's May of 2026 report on national implementation for the NPT Review Conference, which said: "Overall, the Russian Federation ensures, on an ongoing basis, that human involvement and judgment play a decisive role in the field of nuclear weapons, including strict oversight at the level of the country’s highest military and political leadership. Against the backdrop of technological progress, this approach remains an immutable constant; " and the national implementation report of the United States for the NPT Review Conference, which declared it would "Retain – in all cases – a human "in the loop" for informing and executing the President’s decisions to initiate or terminate nuclear weapons employment."

6. As of now, Russia and the United States each have fewer than 800 total strategic launchers; China has an estimated 550, including unfilled strategic missile silos; and France and the United Kingdom have a combined total of about 96. A freeze on strategic nuclear launchers at these levels would not adversely affect any one country’s ability to deter nuclear attack, and could be more easily verified through remote, national means of intelligence.

7. See: "Want to prevent nuclear war caused by AI? Count the private sector out," commentary by Oliver Meier, European Leadership Network, July 8, 2026.

8. I want to acknowledge the invaluable research and analysis of my Arms Control Association colleagues Xiaodon Liang, Senior Policy Analyst and project coordinator for our AI and nuclear "firebreaks project, and our Senior Fellow, Michael Klare, which very much informed and helped to improve this presentation.