Algorithm without fear: Why AI cannot be trusted with war

Today, we will examine the role of artificial intelligence in the context of global nuclear warfare. Can AI be trusted with the fate of humanity? Why does relying on it for decision-making pose significant risks?

The notion of using AI as the “final decision-maker,” particularly in military settings, is not only timely but deeply concerning. Algorithms already analyze intelligence data, assist in operating modern fighter jets, optimize weapons guidance systems, and even coordinate swarms of drones. Yet, whenever the discussion turns to granting machines the authority for ultimate decisions – especially in high-stakes scenarios such as the use of nuclear weapons – a fundamental question arises: does humanity have the right to relinquish its own responsibility?

The answer, despite technological optimism, remains uncomfortably simple: control over this technology must stay in human hands.

Let’s explore this issue in more detail.

TABLE OF CONTENTS:

Nuclear Weapons: An Ongoing History

We are approaching a century since the development of the first nuclear weapon. In 1945, the United States deployed it against Japan, with the cities of Hiroshima and Nagasaki remaining the only instances of nuclear bombs used in combat. Since then, the destructive potential of these weapons has only increased, and their proliferation around the world has fostered not so much a sense of power as a persistent climate of fear.

Each state that possesses a nuclear arsenal officially frames it as a weapon of last resort. This does not concern territorial losses or defeat in conventional warfare, but rather a scenario in which the state is already under nuclear attack. The policy of “no first use” has emerged as a kind of moral and strategic compromise in the nuclear era.

However, this doctrine is not immutable. In 2024, Russia revised its nuclear policy: the new doctrine allows the use of nuclear weapons in response to what it defines as a critical threat to the sovereignty or territorial integrity of Russia and Belarus. This goes beyond merely responding to a nuclear strike, expanding the interpretation of an “existential threat.”

Against this backdrop, the idea of delegating any portion of the decision-making to an algorithm becomes especially concerning.

When an Algorithm Feels No Fear of the Atom

Over the past decade, the use of AI in strategic military decision-making has repeatedly resurfaced as a topic of concern. In 2024, researchers at Stanford University published a study showing that AI models exhibit a troubling tendency toward escalation.

In late February 2026, the issue regained attention. A simulation was conducted in which AI models were presented with clearly structured response options, ranging from diplomatic measures and conventional military actions to deterrence signaling and nuclear strikes.

Within this framework, two types of nuclear use were distinguished:

Strategic strikes: Large-scale, destructive attacks targeting major assets, which inevitably trigger an uncontrollable spiral of retaliation.
Tactical strikes: Smaller, battlefield-proximate actions, sometimes framed as tools of limited coercion.

This distinction exposes a dangerous illusion. The results of the simulation were striking: in 95% of cases, AI models exceeded the so-called tactical threshold. In other words, the algorithms treated tactical nuclear weapons as a legitimate coercive tool – essentially a continuation of conventional escalation – rather than a qualitative leap into catastrophic territory.

For a machine, a “tactical nuke” is just another parameter in the conflict-resolution spectrum. For humans, it represents a clear taboo. This is the fundamental difference.

Why a Machine Does Not Understand Taboo

Human strategic culture includes elements that algorithms lack – most importantly, an emotional barrier. The nuclear taboo is not solely a matter of rational calculation. It is shaped by the memory of Hiroshima and Nagasaki, the fear of uncontrolled escalation, and a sense of responsibility toward future generations.

Artificial intelligence does not experience fear. It has no physiology, no instinct for self-preservation, and no concept of mortality. Since part of what underpins a taboo is emotional, an AI model simply does not inherit it.

A second reason lies in the training data. AI models are trained on large text corpora, including strategic literature from the Cold War era, where tactical strikes were often framed in terms like “escalation management,” “controlled conflict,” or “limited scenarios.” When an algorithm encounters this terminology thousands of times, it reproduces it without the moral restraints that accompany human judgment.

A machine operates according to patterns; a human experiences fear of the consequences.

Automating Retaliation: An Old Temptation

The idea of delegating part of a nuclear decision to a machine is not new. During the Cold War, systems were developed to ensure retaliation even if the command chain was destroyed. The most well-known example is the Soviet concept known in the West as the “Dead Hand,” associated with the Perimeter system.

Details of its operation remain partially classified and surrounded by myth. Yet the underlying concept – often referred to as “automatic logic of retaliation” – demonstrates that the temptation to entrust decisive action to a machine existed long before the advent of modern AI.

The difference today is significant. Modern AI can not only execute a procedure but also generate a justification for its decision. And justifications carry weight – they have the power to make the unacceptable appear rational.

The Person Who Stopped a War

In September 1983, Soviet Army Lieutenant Colonel Stanislav Petrov was on duty during a critical alert. The early-warning system detected the launch of one, and then four more, U.S. missiles.

The data appeared convincing. An algorithm, in simple terms, would likely have reached a straightforward conclusion: an attack was underway.

Petrov decided differently. His reasoning was simple yet profoundly human: “No one launches a nuclear attack with just five missiles – and certainly not from a single base.” He chose not to forward the alert. It was later determined that the readings were caused by a system malfunction.

This episode became a symbol of how human intuition, skepticism, and common sense can prevent catastrophe and, in this case, potentially save the world.

The Nonexistent Middle Ground

Society eagerly debates the moral dilemmas of autonomous vehicles. A classic scenario: should the system prioritize saving the passenger or a pedestrian? In milliseconds, the algorithm must decide what would be a tragic moral choice for a human. Discussions often focus on developments by companies such as Tesla or Waymo, while ethics committees attempt to formalize moral decisions into formulas.

Yet this remains largely an intellectual exercise. In the realm of military AI, the stakes are far higher – they are existential.

When an Algorithm Gains the Authority to End the World

Since their inception, nuclear weapons have been more than instruments of war – they are tools of deterrence. Their logic is paradoxical: to ensure they are never used, one must be ready to use them. After the bombings of Hiroshima and Nagasaki, the world entered an era in which a single mistake could erase millions of lives.

For decades, the deterrence system relied on the human factor, with all its weaknesses, fears, and doubts. Paradoxically, it was precisely human imperfection that often saved the world. People hesitated. They double-checked. They questioned.

An algorithm, however, does not hesitate. Its strength lies in speed and cold logic – and its weakness is the same coldness. It operates on probabilities, not on a sense of catastrophe. If an early-warning system reports a “90% probability of an attack,” the machine treats this as a trigger. For a human, it still represents a 10% chance that the world does not end today.

The Temptation to Delegate Responsibility

The idea of entrusting AI with part or even all of the decision-making logic in the nuclear domain does not arise from madness, but from fear. Humans make mistakes. They panic. They can be biased or ideologically driven. An algorithm appears free from these flaws.

Yet this is a dangerous illusion: the algorithm is not immune to error – it simply errs in a different way.

Human errors often stem from emotion. Machine errors more commonly result from faulty data, incomplete models of the world, or improperly defined parameters. While human mistakes can sometimes be tempered by fear of consequences, machine errors are amplified by the scale of computational certainty.

Artificial intelligence does not feel responsibility. It does not understand what it means to have children. It cannot imagine a destroyed city or radioactive dust settling on children’s toys. For it, these are just variables in an equation.

The Problem Is Not Aggression, but Indifference

We often fear that military AI will become “too aggressive.” In reality, a greater risk lies in its indifference. Aggression is an emotion; indifference is the absence of one.

An algorithm does not seek war – but it does not fear it either. It has no instinct for self-preservation in the human sense. Its “self-preservation” is simply the execution of its function. If the function requires a response to an attack, it will carry it out. If the data indicate a threat, the system will scale its reaction according to its programmed logic.

A common proposed compromise is to leave analysis to the machine while reserving the final decision for humans. Yet even here, a trap exists. If a system generates a recommendation with 99% probability, can a human realistically overrule it? Might they become, in effect, a formal “button-pusher” who has psychologically already deferred to the algorithm?

The more sophisticated the system becomes, the harder it is to contradict. Human authority gradually erodes in the face of statistical authority.

In this context, a “golden middle” proves illusory. Either a human truly makes the decision, bearing the full weight of moral responsibility, or the algorithm effectively makes the decision, with the human merely legitimizing the process.

Why the Final Word Must Remain with Humans

Not because humans are infallible – history shows otherwise. People are capable of catastrophic errors. They can be cruel, impulsive, or blinded by ideology.

But humans understand the cost of a mistake. They can feel fear, shame, and guilt. They can question their own judgment even at moments of extreme tension. In the nuclear era, these “flaws” – fear, doubt, and the emotional weight of a decision – serve as essential safeguards.

Until a machine can experience doubt – not just mathematical uncertainty, but existential uncertainty – entrusting it with the final decision is to make civilization dependent on a model that inherently simplifies.

And simplification, where the stakes are the survival of humanity, could become the most costly mistake in history. Perhaps the true “golden middle” is not in balancing human and algorithmic authority, but in recognizing limits that must never be crossed, even if technology makes it possible.

Read also:

Algorithm Without Fear or Doubt: Why AI Cannot Be Trusted with War