The risks of artificial intelligence are no longer limited to technical errors or misinformation, but have extended to the military sphere, where an unresolved automated estimate can influence fateful decisions and potentially contribute to igniting a military confrontation that no party had planned for.
Imagine an AI system detecting unusual movements of military aircraft or ships and classifying them as indicators of an imminent attack. The detection might be accurate, but the interpretation of those movements as preparation for military action remains an uncertain possibility. If this result pushes leadership to raise the alert status or move its forces, the other side may interpret those actions as a threat and respond with similar steps, starting a chain of mutual reactions.
In such circumstances, AI has not declared war or made the decision to attack, but it may have contributed to creating an environment that pushes parties toward escalation. Herein lies the dilemma: what happens when the speed of the warning outpaces the ability to understand the reality of the threat?
A correct warning... and a wrong conclusion
Detecting unusual military activity does not mean knowing its purpose. The observed activity could be a military drill, a precautionary measure, or preparation for an offensive operation. But in an atmosphere of tension and distrust, a possibility may turn into an estimate that leadership treats as a confirmed fact.
Dennis Fua, a specialist in strategic and security studies at the S. Rajaratnam School of International Studies in Singapore, addressed this problem in a paper titled "Artificial Intelligence and Military Decision-Making: When Warning Outpaces Understanding."
Fua explains that an automated warning does not have to be entirely wrong to pose a risk; the system may succeed in detecting unusual air or maritime movements, but it cannot necessarily confirm that an attack is imminent.
The problem begins when these estimates turn into field decisions. Raising a country's alert level may appear to the opposing party as preparation for an attack, prompting it to take its own precautionary measures, which are then interpreted as additional proof of hostile intent.
Thus, responding to an unresolved warning can turn into a cycle of escalation, where each party becomes more convinced that the other is preparing for confrontation.
Lessons from warnings that almost sparked crises
The risks of false military warnings are not new. In 1983, Soviet officer Stanislav Petrov dealt cautiously with a warning from an early warning system that indicated the launch of American missiles, and did not consider it sufficient evidence of an attack.
In 1995, the launch of a Norwegian scientific rocket raised concerns within Russia about the possibility of a missile attack.
Although these two incidents preceded the use of AI in modern military systems, they reveal the danger of making fateful decisions based on incomplete signals or signals subject to multiple interpretations.
The problem is not just the issuance of the warning, but how it is handled and the extent to which it is verified before taking actions that might be difficult to reverse.
From data analysis to influencing military decision-making
The use of artificial intelligence is increasing in military systems that handle information analysis, target detection, and tracking of movements. In 2024, countries participating in the AUKUS agreement conducted tests to develop AI technologies supported by autonomous systems.
The United States also continues to develop the "Maven" intelligent system, which integrates sensor data and helps detect and track objects and support military decisions.
But the speed of information processing does not necessarily mean the accuracy of the estimates. Systems may face incomplete data or unfamiliar circumstances, and may also produce conclusions that seem convincing despite lacking sufficient evidence.
A study published by the journal "Nature Machine Intelligence" in 2026, titled "Rethinking Human-Machine Interaction and AI Governance in the Military Domain," emphasizes that human oversight must cover all stages of a system's life, starting from its development and testing, all the way to evaluating its performance and verifying the validity of its use.
This means that the presence of a human official to approve the final decision is not enough alone to guarantee control; rather, this official must have the ability to review the recommendation, understand its reasons, and object to it when necessary.
When speed turns into a source of danger
Military decisions become more sensitive when leaders receive a warning in a tight timeframe while available information remains incomplete. Waiting to verify the threat may seem like a risk, while immediate action might lead to an escalation based on a false estimate.
In these conditions, the recommendations of smart systems may influence how officials think, especially if the results seem accurate or are supported by complex simulations.
A study titled "Black Box War: Human Judgment and Military Decision-Making in the Age of Artificial Intelligence" discusses the nature of the relationship between algorithmic recommendations and human decisions, and the importance of understanding how decision-makers handle the output of automated systems.
Fua notes that human approval does not necessarily mean there is genuine oversight, as the official must be given the opportunity to review estimates, test them, and request additional information before adopting them.
In this context, it is essential to distinguish between what the system actually detected and what it inferred about the other party's intentions, in addition to seeking independent sources that confirm the warning or provide alternative interpretations.
Artificial intelligence and nuclear crises
The risks multiply when automated estimates enter into nuclear crisis calculations, where any misunderstanding can have very serious consequences.
A study titled "Generative AI and Artificial Foresight at the Edge of the Abyss: The Future of Decision-Making in Nuclear Crises" addresses the role of generative systems in shaping leaders' perceptions of adversary intentions and potential escalation paths.
The study points to the concept of "artificial foresight," which describes the use of AI to produce future scenarios that may influence preparedness and planning decisions.
However, repeated simulation of the most dangerous scenarios may give them greater weight in leaders' calculations and make ambiguous military movements look like indicators of hostile intent.
The study illustrates this through a hypothetical scenario of a crisis between India and Pakistan in 2029, starting with an explosion in Kashmir. In this perception, an Indian system interprets Pakistani movements as preparation for a nuclear signal, while a Pakistani system reads the Indian movements as preparation for a preemptive strike.
As both sides raise their military readiness, defensive measures turn into mutual threat indicators, and opportunities for diplomatic communication and crisis containment recede.
The study emphasizes that this scenario is hypothetical and does not represent an event that actually happened, but it illustrates how automated estimates can influence the behavior of parties and even contribute to producing developments that later appear to confirm those estimates.
Who bears the responsibility for escalation?
If a warning from an AI system contributes to driving a military decision toward escalation, responsibility does not stop at determining whether the system made a mistake, but extends to the entities that designed and tested it, the leaders who decided to use it, and the mechanisms for reviewing its outputs.
In this framework, a report published by Reuters in September 2026 discussed American and Chinese proposals to establish safeguards to limit the risks of AI, inspired by some principles of nuclear arms control.
Among the proposals are establishing controls on integrating AI into nuclear systems, maintaining human control over cyberattacks with serious consequences, and establishing communication channels to deal with incidents related to autonomous systems.
The principle of keeping the decision to use nuclear weapons in human hands, not automated systems, also stands out as one of the basic safeguards to limit the risks of loss of control.
Fua believes that military institutions need to establish clear procedures in advance for dealing with warnings, identify the entities authorized to review or stop a response, and the evidence that warrants re-evaluating the situation.
He asserts that faster warnings may give leaders more time to make better decisions, but this only happens if they have the ability to question the results, review them, and choose between multiple alternatives.
Could the machine become the spark of war?
The increasing use of AI does not mean that the outbreak of a war due to a false warning has become inevitable, nor is there any indication that these systems have actually caused a war in this way.
But the danger lies in a linked chain of events: detecting a real military move, then interpreting it in an uncertain way, then making a quick decision based on this interpretation, leading to a reaction from the other side that increases tension and fuels mutual suspicions.
Thus, war might not start with a decision made by the machine, but with human decisions based on automated estimates that have not undergone sufficient verification.
The real challenge is not just making AI faster at issuing warnings, but ensuring that the speed of analysis does not turn into pressure that drives leaders to make fateful decisions before understanding what is actually happening.
