EVIDENCE FILE 19 // NUCLEAR COMMAND
AI and Nuclear Command Risk: The Last Human Decision
There is no restart button after a nuclear launch. That is why the most terrifying role for AI is not pulling the trigger itself. It is shaping what a frightened human sees in the final minutes before the decision.
SIPRI research warns that military AI can compress nuclear decision timelines, increase miscalculation risk and destabilise confidence even when the AI is outside the nuclear weapon system itself. The UN Secretary-General has repeatedly argued that decisions on nuclear use must remain with humans, not machines.
TWELVE MINUTES TO DECIDE WHETHER THE WORLD ENDS
Warning systems report an attack. Some sensors agree. Others do not. An AI decision aid insists the pattern is real and recommends immediate action. A deepfake military communication appears to confirm it.
The nightmare is not an AI pressing a fictional red button by itself. It is human leaders being pushed toward irreversible decisions by machine-speed information they cannot verify in time.
Scenario: this is a deliberately extreme “what if?” exercise, not a claim that these events are happening or certain to happen.
AI can affect nuclear risk without controlling a missile
Nuclear command depends on warning systems, intelligence, communications, planning and assessments of adversary intent. AI can enter those surrounding layers long before anyone proposes giving a model direct launch authority. A system that filters sensor data or recommends how to interpret an adversary’s movement can influence the decision while the legal authority remains entirely human.
SIPRI’s 2025 work on military AI and nuclear escalation argues that non-nuclear uses of AI can still compress decision timelines, bias decision-makers and threaten strategic stability. Its 2026 research also documents growing integration of AI into nuclear-related systems and continuing debate about what meaningful human control should require.
Speed is both the advantage and the trap
The military case for AI is speed: detect threats sooner, fuse more data and give commanders more time. But faster analysis can also create pressure to respond faster. If an opponent is believed to be using AI to act at machine speed, waiting for slower human verification may feel dangerous. Each side can therefore accelerate because it fears the other side is accelerating.
That dynamic is especially dangerous in nuclear strategy, where false alarms and misread signals already have a documented history. AI could reduce some errors while introducing new ones that are harder for operators to understand.
WORST-CASE SCENARIO — The final seven minutes
SCENARIO: During a major cyber conflict, early-warning networks begin producing inconsistent data. A military AI correlates satellite changes, communications outages and unusual aircraft movements into a high-confidence warning of an imminent attack. Human analysts disagree, but the model has outperformed them repeatedly in exercises. Commanders are told that delaying a response could eliminate their retaliatory capability.
At the same time, a deepfake appears to show an enemy leader announcing mobilisation. Diplomatic hotlines fail under cyberattack. The opposing state sees defensive dispersal and receives its own AI-generated warning. Seven minutes later, both leadership teams are looking at different machine-generated versions of reality. The machines do not possess launch authority. They do something almost as dangerous: they convince humans that there is no time left to doubt them.
Why human control has become an international principle
The UN Secretary-General has repeatedly stated that decisions on the use of nuclear weapons must remain with humans. SIPRI notes that major nuclear states have made various commitments around human control, while the details and transparency differ. The concern is not symbolic. Human judgement is the final barrier against a chain of automated interpretation becoming an irreversible act.
The ICRC goes further in its 2026 military-AI guidance, arguing that certain AI uses should be prohibited, including in nuclear command and control systems. These positions show that the AI-nuclear intersection is already treated as a concrete governance problem.
The scenario does not require a sentient villain
Nothing in this pathway requires consciousness or an AI deciding it hates humanity. A model can simply be wrong, manipulated or over-trusted. Alternatively, a malicious actor could deliberately poison data or generate false signals. The catastrophic variable is the coupling between machine-speed analysis and weapons whose consequences cannot be reversed.
That is why nuclear risk is one of the clearest examples of the broader theme of this site: the worst AI disaster may emerge from systems interacting with human fear and existing infrastructure rather than from a cinematic declaration of war by machines.
If strategic crisis reaches civilians
Households cannot meaningfully prepare for every nuclear outcome, and no survival product should pretend otherwise. What people can prepare for are the earlier stages of national crisis: communications failure, power disruption, transport restrictions, shortages, payment problems and rapidly changing official instructions.
The 72-hour manual focuses on those survivable opening failures. It is not a promise that preparation defeats the end of the world. It is a plan for the period when decisions still matter.
Continue from here
Connect the documented risk discussion to the scenario it informs and the practical preparation it changes.