As the world gears up for the 2026 Nuclear Non-Proliferation Treaty (NPT) Review Conference, a new and multifaceted factor is complicating the global strategic calculus: Artificial Intelligence (AI). The “nuclear-AI nexus” has evolved from a niche technical interest to a prominent feature in global security discussions, with implications for every aspect of the NPT’s three pillars: non-proliferation, disarmament, and peaceful uses of nuclear energy. However, as experts recently cautioned at the “Atoms for Algorithms” webinar, we need to cut through the speculative “AI hype” to ensure this technology remains a means for peace, not an avenue for unintentional escalation.

To regulate the nuclear-AI connection, we first need to understand the technology. As a United Nations Institute for Disarmament Research researcher, Dr. Yasmina Fina suggests, AI is not a monolithic entity, but a “construct” and a “system of systems,” made up of code, software, data, hardware, and sensors. It is a system used to fulfill certain functions, not a threat capable of usurping human decision-making. The risk is the “speed and scale” of AI, which can have myriad implications for performance and strategy. Moreover, Fina warns that comparing nuclear governance to AI governance is unhelpful because the technologies are not the same; nuclear materials are limited and tangible, whereas AI is ubiquitous and digital.

In the context of non-proliferation, AI is touted as a game-changing verification tool. The International Atomic Energy Agency (IAEA) could potentially use “AI agents,” semi-autonomous machines capable of processing large data streams and satellite images to verify the accuracy of declarations by states at a pace humans cannot match. However, Dr. Ian Stewart, Executive Director of the CNS Washington, states that AI will not help states develop nuclear weapons they could not otherwise build, for two physical reasons: AI cannot “magic up” fissile material, and there is no evidence that large language models can transfer the “tacit knowledge” necessary for weaponization.

When it comes to AI and nuclear weapons, political concerns are high. States have been reluctant to allow the IAEA to use open-source data or “black box” algorithms. Should an AI detect an event, the absence of “explainability” or how the machine arrived at its decision could lead to a crisis of political legitimacy for the safeguards system.

The most fraught part of the nexus is disarmament. We are now seeing a “race to adopt” AI in military strategies due to the perceived speed advantage it offers. AI can speed up threat identification and data integration, potentially freeing up more time for decision-making (or, on the other hand, reducing decision-making cycles to the point that humans are simply rubber-stamping decisions).

Aliche Sultini, senior research lead at the Rhode Island School of Design, explains that AI systems create new levels of uncertainty. If a state cannot grasp how an adversary’s AI operates in its decision-making, it might fall into worst-case scenarios, reinforcing alert postures that prevent disarmament. To support disarmament, AI must be used to enhance technical verification and confidence, not to shorten the path to war. Possibly the most disruptive aspect of this interaction is the NPT’s third pillar: peaceful nuclear energy applications. According to Mr. Shota Kamishima of the IAEA, an “affinity” is emerging between nuclear power and AI. We are now moving into a world where energy-hungry AI data centers need the clean, scalable, and reliable power offered by nuclear power, and where nuclear generation and maintenance are improved through AI.

This alliance is especially important for the rollout of Small Modular Reactors (SMRs), for which AI-optimized construction schedules and supply chains are critical. By enhancing predictability and avoiding cost overruns, a major issue for nuclear construction, AI could make nuclear projects more “bankable” and thus more attractive for the global shift towards clean energy. Despite technological progress, the experts agree: human responsibility is essential. Whether it is a safeguards inspector at the International Atomic Energy Agency (IAEA) or a human commander in a nuclear-armed nation, humans are the “last line of defence”.

“Black box” systems are incompatible with a strong safety culture. Governance policies must ensure that AI is implemented with transparency, traceability, and “explainability”. We must also be alert to the potential for AI to be employed by “agents” to monitor sites, which could result in disinformation and the distortion of threat perception through “anomaly detections.” As the 2026 Review Conference draws near, policymakers’ mission should be to “denoise.” Nuclear policy decisions must not be based on science fiction or fear of competition. Rather, we should prioritize “lower stakes” opportunities where AI can help us now, such as employing AI agents to navigate the overwhelming output of the NPT process — making it searchable and highlighting inconsistencies in delegation positions.

The relationship between nuclear and AI is not a bug to be fixed with more software, but a circumstance to be managed by the international community. By prioritizing evidence-based policy and law, human-in-the-loop systems, and the common ground of the peaceful “Atoms for Algorithms” alliance, we can ensure that the digital revolution supports, rather than undermines, the global nuclear order. In the end, the fate of the NPT will not be determined by algorithms, but by human intelligence.

Muhammad Shahzad Akram is a Research Officer at the Centre for International Strategic Studies, AJK. He holds an MPhil in International Relations from Quaid-i-Azam University, Islamabad. He is an alumnus of the Near East South Asia (NESA) Centre for Strategic Studies, National Defence University (NDU), and Washington, DC. His expertise includes cyber warfare and strategy, arms control, and disarmament. Views expressed in this article are the author’s own. 

Navigating the AI and Nuclear Nexus was originally published on Global Security Review.

Focusing on information operations is critical. Drones, for example, use artificial intelligence capabilities, as do defensive systems. However, conflict between near-peer adversaries and competitors is still unlikely as gray zone and hybrid conflict are the dominant avenues for competition. With the information environment transcending all domains of warfare, artificial intelligence capabilities become the go-to capability to ensure and maintain information advantage.

China’s AI-enhanced information operations are becoming increasingly sophisticated. For example, the Chinese advanced persistent threat actor Spamouflage Dragon uses generative AI to create online personas to influence public opinion. China and its proxy companies seek to develop or compete for AI supremacy within the information environment.

Of course, China will use anything within its arsenal to shape strategic, operational, and tactical levels of war to its advantage, expand its influence, and create an ecosystem that is dependent on its technologies. For example, Baidu, known as the “Google of China,” invested billions into AI capabilities, creating the proprietary ERNIE model, which has been trained on billions of parameters, increasing the output’s quality and complexity.

However, China is also leveraging open-source AI models to shape the information environment. With the recent release of open-source large language models such as DeepSeek and Qwen, Chinese-linked subsidiaries, High-Flyer and Alibaba Group created a way to expand their influence, revise history, and likely create a dependent ecosystem for target countries. Unlike the much more expensive ChatGPT, for which the more basic model is free, China’s investment in generative AI models is free for the public and even surpasses ChatGPT’s in some respects.

There is a debate about how China’s proxy state companies were able to create these advanced models without US-based critical components. China allegedly did not have access to the advanced critical Nvidia chips for which most AI models are dependent. China seems to have created generative models just as suitable or even better than that of ChatGPT, but allegedly at a fraction of the cost and free of charge to the public. The US limited Nvidia chip exports to China, a market predicted to top $1 trillion in revenue within a decade. Still, the accusation is that the Chinese subsidiary leased or bought the more advanced Nvidia chips from Singapore, circumventing restrictions, and used ChatGPT to train its model.

Regardless of how China secured these critical technologies, the cat is indeed out of the bag. China has shown that it has the capability to develop new and emerging AI technologies. From the capabilities already built, it now has a baseline to create even more capabilities to develop its own AI chip ecosystem. With such capabilities, China will become more active within the information environment with the help of AI capabilities, and its motives are far from benevolent.

Why is the Chinese model free? China has several motives, but it is likely in hopes that data and information across the globe are the price tag for using the model while lessening a dependency on Western technologies and becoming a global leader in AI by 2030. Whatever data is obtained by the United States is icing on the cake. The West is not the primary target audience. Both models have servers in Singapore and China, where information is likely subject to Chinese laws, and terms and conditions are meaningless.

The Chinese will use AI technologies to gain an advantage in the information environment and seek to expand influence by creating an ecosystem for which other countries are dependent on their models. The incentive is to give countries this technology to foster dependency. The idea is similar to China’s debt-trap diplomacy—the Belt and Road Initiative. While ChatGPT’s basic model is free, China seeks to develop better models at a cheaper price to serve as leverage over countries that cannot afford the higher-end US-based models.

The United States is taking the right approach to maintaining its information advantage through AI development and investment. The billions pouring into creating AI data centers will play an important role in ensuring the United States has the edge in AI.

These data centers remain critical for identifying and countering any malign information operations against the United States, its partners, and its allies. When Iran attempted to influence the 2024 presidential election using the generative model GPT, OpenAI detected and shut it down. Without this expansive investment in AI data centers that keep information within the letter of US law and oversight, these interventions would be out of reach, and information operations may be even more challenging to detect.

However, this approach is insufficient without incorporating artificial Intelligence into all aspects of military operations. The DoD uses artificial intelligence within some branches, but given the expansive nature of AI, this is not enough. AI is expected to touch nearly all aspects of military operations, especially information operations, and may not have time to wait for its major AI initiative, Project Maven, to fully develop.

Some military scholars have called something like this a revolution in military affairs, but perhaps, given the impact of war, it could be classified as such. The concept is somewhat antiquated and outdated without some context, but it remains the best way to describe what should take place within the DoD. The foundation is already in place through the conceptual framework of multidomain operations.

Artificial capabilities are widely available through graphical user interfaces in deployable, ready-to-use form, such as ChatGPT or even internal large language models. The joint force should use these capabilities to the broadest extent possible. If anything, artificial intelligence, including large language models, will make joint and combined forces more lethal and accurate as they counter Chinese efforts within the information environment.

The DoD must adopt incentives for service members to understand the capabilities of AI and incorporate them in all training environments. These incentives can include bonuses for taking AI-driven courses. The DoD can also increase awareness and accessibility of AI courses on its education platforms which now have a paucity of artificial intelligence courses.

The DoD must also improve the training environment. With proprietary or off-the-shelf software, the DoD can incorporate AI offensive and defensive platforms within all training and mission-critical tasks. Even simply assisting with identifying generative outputs, e.g., deepfakes, will counter Chinese influence within the information environment, especially during hybrid conflict. Furthermore, military doctrine should recognize the importance of AI, especially information operations, with an emphasis on psychological operations.

While AI investment is critical to countering Chinese influence within the information environment, the only way to truly embrace multidomain operations is to ensure service members have the AI technical competency necessary to maneuver within the information environment deterring Chinese aggression.

US Army Lieutenant Colonel Matthew J. Fecteau is a PhD researcher at King’s College London studying how artificial Intelligence will impact conflict. He can be reached at matthew.fecteau.alumni@armywarcollege.edu.

China’s AI-Driven Information Operations Are Here: The US Needs an AI RMA was originally published on Global Security Review.