The real danger is not a machine independently deciding to launch. It is a human decision made from a machine-generated picture that is wrong, incomplete, or misunderstood.
The most plausible route by which artificial intelligence could contribute to a nuclear war does not begin with a computer taking control of a missile silo.
It begins with a screen.
An early-warning system reports a possible missile attack. Intelligence software combines satellite images, radar tracks, and intercepted communications. It then presents officials with a high-confidence assessment that an attack is under way. The national leader still decides, but nearly everything placed before that leader has already been selected, ranked, and interpreted by machines.
If the assessment is wrong, human control over the launch order may offer less protection than it appears.
Could this happen? In principle, yes. While there is no public evidence that any nuclear power has delegated the decision to use nuclear weapons to AI and the details of nuclear command systems are closely guarded, nevertheless, AI does not require launch authority to influence nuclear stability.
AI does not need access to the launch button
The term “AI” can be misleading. The systems most likely to enter nuclear operations are not necessarily chatbots or human-like computers. They are more likely to be specialised tools trained to recognise patterns, find anomalies, combine large quantities of data or suggest possible courses of action.
AI could assist with missile warning, intelligence analysis, cyber defence, equipment maintenance, operational planning and target identification. It might search satellite images for mobile missile launchers, distinguish an attack from a test or detect unusual activity on military networks.
This is not entirely theoretical. The United States Strategic Command stated in 2025 that it intended to use AI and machine learning for nuclear command, control, and communications cybersecurity, situational awareness and the acceleration of human decision-making. It also said the technology would remain subordinate to human authority.[1] Exactly how far other nuclear powers have proceeded is not publicly known.
Some uses, such as identifying a failing component, are far removed from a launch decision. Others sit much closer to the nuclear threshold. Software that ranks warnings, estimates an adversary’s intentions or places response options before a leader may shape a decision even if it cannot execute it.[2]
Could AI recognise a real attack more reliably than people?
Possibly, under the right conditions.
Machines can examine more information than human analysts and do so more quickly. An AI system could compare independent sensors, notice that one reading conflicts with the others and direct attention towards the discrepancy. It might therefore prevent a false alarm rather than cause one.
The difficulty is proving that it will remain reliable in the event for which it matters most. There is no record of a real modern strategic nuclear-missile attack from which a warning system can learn. Tests must rely on missile trials, exercises, simulations and historical incidents, none of which can reproduce every failure or deception that might occur during an unprecedented crisis.
AI is generally strongest when new information resembles its training data. A nuclear crisis may bring the opposite: unfamiliar weapons, damaged sensors, disrupted communications, and deliberate confusion. Success in testing would not guarantee sound judgement on the system’s worst day.
When false information looks real
Past warning failures show why context matters. In November 1979, a training scenario depicting a large Soviet missile attack was mistakenly introduced into an operational United States warning system. The displayed attack looked convincing because a realistic exercise was designed to imitate the sensor information expected during the real event.[2]
An AI examining only those inputs might reasonably — but wrongly — conclude that an attack was taking place. Detecting the mistake would require wider context: where the data originated, whether independent sensors confirmed it and whether the political situation made an attack credible.
AI also creates opportunities for deliberate deception. The United States National Institute of Standards and Technology describes “evasion”, in which inputs are altered to produce a wrong classification, and “poisoning”, in which training data or the model is corrupted.[3] An adversary might try to disguise a real object, manufacture a false signature, or corrupt information used in an assessment.
Public information is insufficient to judge the security of classified nuclear systems, and penetrating one may be extremely difficult. Not every deception would require such access, however. An intelligence model could be misled by false imagery, fabricated communications or staged military activity gathered outside the nuclear network.
AI could also strengthen cyber defence by detecting unusual behaviour rapidly. It may expose an intrusion humans would miss or add another conclusion that may have been overlooked.
The danger of a confident answer
The human response to automation may matter as much as the software.
“Automation bias” is the tendency to give excessive weight to a computer-generated recommendation. As a system establishes a record of success, operators may stop asking how it could be wrong and begin treating its output as fact.[4]
A precise figure can make this worse. An assessment such as “92 per cent confidence” looks authoritative, but it is still a model’s calculation based on assumptions and available data. It does not mean there is an objectively measured 92 per cent chance that war has begun.
This is why merely placing a human “in the loop” does not settle the issue. If the human sees only the machine’s summary, has little time to question it and cannot examine independent evidence, the final decision remains human in a formal sense while becoming heavily dependent upon automation.
The opposite failure is also possible. A system that produces too many false warnings may teach its operators to disregard it. AI could therefore contribute to a mistaken launch by being trusted too much or contribute to a failure to recognise a real attack by being trusted too little.
When faster is not necessarily safer
Speed is one of AI’s principal military attractions. It is also one of its principal nuclear risks.
Intercontinental ballistic missiles can reach their targets in roughly 30 minutes, leaving leaders considerably less time after detection, confirmation and consultation.[5] AI could use that limited period more efficiently. If it resolves conflicting reports and establishes that warning data are genuine, it may improve stability by making retaliation more certain and discouraging an adversary from attempting a first strike.
Yet faster analysis can become an expectation of faster action. A state that believes its opponent can identify mobile missiles, track nuclear submarines, or attack command systems with AI may fear that delay will leave it unable to respond. Each side’s attempt to gain decision time could leave both with less.
Deterrence does not depend on speed alone. It also depends on survivable forces, credible communications and leaders understanding what an opponent is likely to do. AI may strengthen one part of that system while weakening another. A quicker decision is useful only if it is the correct decision.
What does “meaningful human control” mean?
Several nuclear powers have publicly recognised the issue. United States Strategic Command has promised a human role in all actions critical to informing and carrying out presidential decisions on nuclear use.[1] In November 2024, China and the United States jointly affirmed that the decision to use nuclear weapons should remain under human control.[6] France, the United Kingdom and the United States repeated a broader commitment to human control and involvement in critical nuclear-employment actions at the 2026 Nuclear Non-Proliferation Treaty Review Conference.[7]
The United Nations General Assembly also addressed the subject directly in December 2025. A resolution calling for human control and oversight over nuclear command, control, and communications systems, including those incorporating AI, passed by 118 votes to nine, with 44 abstentions.[8] The divided vote and the non-binding nature of the resolution show that international concern does not yet amount to a universal, enforceable standard.
A formal guarantee could establish an expectation, but its value would depend upon what it covered. Does human control apply only to the launch order, or also to attack assessment, target selection and the presentation of options? How much information and time must a leader have to reject the machine’s conclusion? Such provisions would be difficult to verify inside classified national systems.[2]
The central issue is therefore not simply whether a person gives the final order. It is whether that person retains enough independent information, time and authority to exercise genuine judgement.
So, could AI accidentally start a nuclear war?
AI is unlikely to “start” a nuclear war as an independent actor like we’ve seen in movies such as Wargames or The Terminator. A more credible danger is a chain of events: ambiguous or manipulated data enters a system, the system interprets it as an attack, officials trust that conclusion, and a leader authorises an irreversible response before the error is discovered.
That chain is plausible, but it is not inevitable, and its probability cannot be calculated from public information. AI could also interrupt the chain by finding a sensor malfunction, exposing falsified data, or showing that an apparent attack is inconsistent with other evidence.
The effect will depend on where AI is placed, how its uncertainty is communicated and how people behave when time is short. Humans may remain in legal control while becoming practically dependent on a machine-generated view of the crisis.
That is less dramatic than a computer seizing the launch codes. It is also the more realistic risk.
- Sources and endnotes
- United States Strategic Command, 2025 Congressional Posture Statement, 26 March 2025, pp. 17–18.
- Herbert S. Lin, “Artificial Intelligence and Nuclear Weapons: A Commonsense Approach to Understanding Costs and Benefits”, Texas National Security Review, 12 June 2025.
- Apostol Vassilev et al., Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations, National Institute of Standards and Technology, March 2025.
- United States Government Accountability Office, Artificial Intelligence: DOD Needs Department-Wide Guidance to Inform Acquisitions, GAO-23-105850, June 2023, pp. 20–21; United Nations Secretary-General’s Advisory Board on Disarmament Matters, Report on the Work of its Eighty-third Session, A/80/240, 2025.
- United States Government Accountability Office, Nuclear Triad: DOD and DOE Face Challenges Mitigating Risks to U.S. Deterrence Efforts, GAO-21-210, 6 May 2021, p. 8.
- Ministry of Foreign Affairs of the People’s Republic of China, “An Overview of the Meeting Between Chinese and US Presidents in Lima”, 17 November 2024.
- France, the United Kingdom and the United States, P3 Joint Statement for Main Committee I, 2026 NPT Review Conference, 1 May 2026.
- United Nations General Assembly, Possible Risks of the Integration of Artificial Intelligence into Command, Control and Communications Systems of Nuclear Weapons, A/RES/80/23, adopted 1 December 2025.
