The biggest risks involve technical flaws and the way AI changes military strategy. Nuclear armed countries are considering the integration of artificial intelligence into existing nuclear command, control, and communications structures as a way to increase speed and efficiency.  However, this integration brings major risks:

• Reduced decision-making time and rapid escalation

The application of AI into nuclear-decision making processes risks drastically shortening the “sensor-to-shooter” window, forcing existential decisions to be made much more rapidly than the minutes currently allotted. (SIPRI). 

AI-enabled systems increase the pace of conflict, which can trigger escalation in which systems will respond more aggressively due to misinterpretation of data or lack of context

• Black Box, Hallucinations and Reward hacking - Technical incomprehensibility and lack of transparency

There are inherent technical flaws in AI models, particularly those based on deep learning that are incompatible with nuclear command and control. Advanced AI processes are in a ‘black box’ meaning even their programmers do not fully understand the reasoning behind recommendations, making them impossible to verify during a crisis. AI systems can also suffer from "hallucinations" generating confident but entirely false information, and for situations like a nuclear crisis for which there is little or no real-world training data, generating false information (for example identifying an incoming attack when there is none). Finally, there is a risk that the AI might misunderstand exactly what it is supposed to do therefore it could try to achieve a goal in a way that disregards consequences, also known as “reward hacking”.

• Perceived increases in vulnerability incentivising nuclear weapon use

AI-enhanced intelligence, surveillance, and reconnaissance can analyse massive amounts of data from satellite imagery and other sensors, increasing the ability to ‘see’ things that were previously considered ‘invisible’ or hidden, such as nuclear-armed  submarines. This might incite decision-makers to launch their nuclear weapons in pre-emptive strikes rather than risk their loss.

• Cyber risks, data poisoning and spoofing

Adversaries could use AI to manipulate the information reaching decision-makers, creating a false picture of the battlefield or a fabricated threat of an incoming attack. It is also impossible to completely shield nuclear weapon systems from being hacked or compromised by AI-enabled cyber operations. 

[Last revised February 2026]