By Jack Higgins, Policy Intern, Summer 2025
The United States plans to spend $1.7 trillion to modernize its nuclear forces over the next 30 years with artificial intelligence set to play a major role in the effort, but will it be a worthwhile investment? While AI has the potential to improve safety-critical early warning and detection systems, its implementation is accompanied by serious risks. The United States must ensure it uses AI responsibly, pursuing policies and working with other nuclear powers to mitigate the risks.
Early warning and detection systems play a critical role in nuclear deterrence, using a combination of sensors and satellites to watch for incoming attacks. AI can analyze the vast amount of data involved and identify patterns more quickly than a human being, giving it the potential to improve situational awareness and make predictions about production, deployment and use of nuclear forces. In fact, AI is already playing a role in early warning systems through AI-enabled detection systems used to monitor airspace and track threats.
Benefits and Risks of AI Implementation
If applied correctly, AI tools could present decision-makers with relevant information quickly and accurately, granting more time to make critical decisions. This time is especially important for a country like the United States which employs a launch on warning policy for deterrence, leaving potentially less than 20 minutes to decide whether to retaliate when a launch is detected. By improving early warning and detection systems, AI can help decision-makers assess threats and reduce the risk of close calls like those that brought the world to the brink of nuclear conflict during the Cold War.
While AI has undeniable potential, it is also inherently limited, making it risky to use in safety-critical tasks like early warning and detection. Machine-learning programs must be trained on large amounts of accurate data. The lack of real-world data from nuclear conflict could cause AI to fail when confronted with real-world scenarios. On top of this, evidence shows AI often makes mistakes even in controlled environments, and we cannot observe the internal process through which AI draws conclusions, making its outputs unpredictable.
All of this results in the possibility that AI early warning and detection systems might get it wrong and draw false conclusions that appear accurate. In a worst-case scenario, an AI system could issue a false positive, warning decision-makers of a non-existent attack leading to catastrophic nuclear conflict. Concern over this exact scenario has prompted nuclear powers including the United States, France, the United Kingdom and China to make policy statements promising human control over the ultimate decision to launch nuclear weapons. While these statements are a step in the right direction, much more needs to be done to mitigate risks.
Maximizing Benefits While Mitigating Risks
As AI is implemented in early warning and detection systems, there is much the United States can do on its own and in collaboration with other nuclear powers to mitigate risks. The goal should be to develop a tool that can support human decision-making without granting AI too much autonomy.
-
-
- The United States must work to ensure active human control over AI and provide decision-makers with the tools to assess AI conclusions. Even if humans are in control of the final decision to launch nuclear weapons, it may not be enough to prevent an AI mistake from leading to disaster. People who work with AI frequently can suffer from automation bias, becoming over-reliant on AI so that they fail to monitor its decisions closely and hesitate to override it. The United States should take steps to ensure decision-makers remain engaged in the process and have the tools to accurately and effectively evaluate AI analysis. It is important that decision-makers are never solely reliant on AI for information and are always able to check AI conclusions against other independent sources, especially when probabilistic AI is used to draw conclusions from patterns in data. One prudent measure might be to employ multiple AI algorithms trained on different sets of data, helping mitigate the risk of bad data leading to mistakes. It is also important to ensure that early warning and detection systems can still function without AI when it is implemented, in case it fails or becomes unavailable. These steps and others should be pursued with the goal of ensuring AI functions as a useful tool to support decisions, rather than making decisions itself.
- As the United States works to implement AI responsibly and safely into its own systems, it should engage with other nuclear powers to reduce risks. The United States and other nuclear powers should continue Track One dialogue like the discussions between the United States and China that resulted in a joint statement on the need for humans to maintain control over nuclear weapons, and Track Two efforts like the Normandy P5 Initiative, bringing together experts from the P5 members of the UN Security Council for a series of dialogues focused on mitigating nuclear risk and the nexus between AI and nuclear weapons. Steps like publicly defining AI terms and stating public AI-nuclear doctrines could help reduce tensions and avoid conflict. Although it may be uncomfortable in a field with so many closely guarded secrets, transparency about AI implementation can help avoid a dangerous race to over-automate nuclear systems.
- AI tools should be leveraged to support the decision-making process but not necessarily speed it up. There is a risk that AI use in early detection systems could shorten decision timelines as states rush to automate and quicken responses. To avoid this, nuclear powers should strengthen crisis communication channels focused on AI to help prevent mistakes leading to catastrophic consequences. AI should be used to create more time for accurate and informed decisions, not rush them.
-
The investment that will be made in AI early warning and detection systems is too large, and the stakes are far too high, to get it wrong. While steps have been made, the United States and other nuclear powers must ensure AI is implemented responsibly to make the world safer, not more dangerous.
