AI 无需超级智能或恶意也能引发核战争

Current AI systems could trigger nuclear war by amplifying human cognitive blind spots, speeding up escalation spirals, and overwhelming decision-makers with information during high-stakes crises (Illustration by François Diaz-Maurin; source images from The White House and canva.com)

When AI researcher Jacob Coxon announced his resignation from Anthropic this month, he warned that people at companies racing toward self-improving superintelligence “earnestly believe that it could kill us all by the end of the decade.” Such warnings raise fears that AI could escape human control and start a nuclear war without human authorization.

Although superintelligence is not precisely defined, AI developers agree it is not yet here. But today’s AI systems can cause nuclear war even if people nominally remain in charge. AI need not launch a weapon, recommend a strike, or intend harm. Ordinary human fallibility—amplified by machines—may be enough to lead to catastrophe.

AI can shape the information reaching leaders, whose minds are already struggling to comprehend the consequences of nuclear war. AI can exploit known vulnerabilities of human judgment compounded by false, misleading, or out-of-context machine outputs and amplified by the speed, scale, and volume of AI-generated information. Even accurate information can be interpreted improperly or unwisely when presented without proper context; errors and ambiguities introduced by AI can make matters worse.

Neither extraordinary machine intelligence nor malicious intent is required for catastrophe to result. To avoid nuclear war, institutional guardrails must be put in place that protect leaders from their worst impulses, even when AI pressures them to act.

Moral relativism. One major vulnerability, psychic numbing, results from relying on emotions to recognize and respond to risk. Human feelings fail to keep pace with the number of people suffering. Sometimes concern even diminishes as the numbers grow. Humans can recognize a million deaths dispassionately as a number—without comprehending the reality of the lost lives that this number represents. The saying, “the death of one man is a tragedy, while the death of millions is a statistic,” illustrates the fundamental problem of psychic numbing.

Human attention is severely limited as well. Saving troops, stopping an attack, or ending a war can become overwhelmingly prominent in the thinking of decision makers. These are legitimate goals, but they are not the only important ones—averting global catastrophe is arguably far more important. While immediate benefits are readily imagined and carry an emotional punch, the longer-term consequences of nuclear war—climatic changes, ruined agriculture, mass population displacement and starvation, and the total collapse of healthcare and other services around the world—often feel too vague and emotionally remote to adequately grasp. These long-term threats could easily be eclipsed by more immediate considerations during a crisis.

In a recent survey experiment involving nearly 3,000 American adults, researchers described a hypothetical, winnable ground war with Iran, in which a nuclear strike on a major city offered a quicker end and no further American troop deaths.

The experiment tested two separate scenarios of nuclear use to end the war: a strike that would kill 100,000 Iranian civilians, or one expected to kill 2 million civilians. Participants saw only one of these two nuclear options. Survey respondents’ support for the nuclear option over the continuing ground war was nearly identical—31.6 percent for the 100,000-kill option versus 27.6 percent for the 2-million-kill option—despite a 20-fold difference in the number of deaths. This is psychic numbing on display: Beyond a certain scale, greater death tolls stop registering as larger.

Later in the same study, participants were asked to revisit their decision with both nuclear options placed side-by-side. Support for the 2-million-death option collapsed to 5.1 percent, showing that direct comparison can reduce numbing. But numbing didn’t disappear—it shifted. Support for some nuclear option—mostly the 100,000-fatality strike—rose to 40 percent. The 2-million fatality option made 100,000 deaths look more attractive by comparison, even though that still-catastrophic option had been rejected earlier by a large share of participants.

The study shows that moral feelings reinforce the choice to go nuclear. A choice perceived as “protecting US troops” or “punishing an aggressor” can make using nuclear weapons seem virtuous, even without a comparison. The results also reveal how emotions and the menu of decision options shape judgments and, ultimately, decisions. Similar modes of thinking were observed at the end of World War II and other past wars. And the Pentagon’s ongoing review that seeks to offer the president “more options” to launch nuclear weapons in a crisis risks further lowering the nuclear threshold.

Support for nuclear options observed in our experiments did not require information to be false. Participants received explicit casualty estimates. Their choices nevertheless exhibited psychic numbing in some scenarios that became deceptive in another way when comparisons created a “better nuclear bomb” mindset. This matters for AI: Improving the accuracy of AI-delivered information that reaches a decision maker cannot, by itself, ensure that the suffering represented by those facts receives appropriate weight. These experiments demonstrate how serious distortions in judgment and decision-making can occur even in presence of accurate information.

Now add the AI pressure. Warnings, images, reports, and interpretations can now arrive faster than people can digest them. This information may be false, ambiguous, or accurate. It has already been described how AI-enabled degradation of this information environment could increase nuclear risk.

Consider a hypothetical military exercise.

AI correctly detects mobile missile launchers moving. The official receiving that information, uncertain but worried, interprets the movement as possible preparation for attack and raises military readiness as a precaution.

The adversary observes those preparations and responds with precautions of its own. AI detects those movements, which appear to confirm the initial suspicion. Each side’s efforts to protect itself make the other feel less secure.

This is an escalation spiral—what international relations theorists call a security dilemma, but this time it unfolds in real time. No one needs to fabricate the original observation. Correct reporting can perpetuate a mistaken account of intentions. Deliberate deception can worsen the process, but an innocent observation that is misinterpreted is enough to create this escalation.

Distrust complicates correction. A leader may doubt a warning yet fear the consequences of ignoring it. An adversary’s reassurance may be dismissed as deception. A “better-safe-than-sorry” mindset can justify action without adequately considering the danger that the allegedly safe action can create.

Meanwhile, the AI-generated information consumes deliberation time. It takes time to filter out false information and to verify that the information is true. Even when correct, time pressure can increase anxiety and change the affective state of those processing that information. There is also stronger reliance on a shallow overall favorable or unfavorable feeling in judging risks and benefits under time pressure. Haste does not invariably increase perceived risk but, in a climate of distrust and suspicion, it likely will.

The hot-cold empathy gap adds another vulnerability.

When calm, people tend to underestimate how fear and anger will change their preferences. Applied to a nuclear crisis, this raises a troubling possibility: Restraint endorsed beforehand in a quiet situation may fall by the wayside during a crisis in which delay seems to risk intolerable loss.

A leader could then experience the escalating confrontation as an immediate and severe threat and view nuclear use as an escape or preemptive necessity. Unfortunately, under conditions of “hot” cognition driven by intense emotion, personal investment, concern for historical legacy, high stakes, or implicit biases, what the leader holds dear is most likely to be immediate security protections, not the larger fate of the nation and the world.

Immediate protection dominates attention; wider suffering remains abstract, and likely out of mind, despite its importance. A case in point: Vladimir Putin once justified Russia’s nuclear doctrine by questioning the value of a world in which Russia had been destroyed by nuclear weapons.

Requiring human authorization does not, by itself, correct these distortions. The relevant questions concern what evidence was checked, which interpretations were challenged, how deadlines were established, and whether anticipated benefits displaced consideration of subsequent greater harm.

Institutional guardrails. The vulnerabilities of human judgment are well documented, and concern about advanced AI is growing. But awareness alone won’t prevent an escalation spiral. A machine can flood decision makers with information—some true, some false—that they struggle to evaluate but feel compelled to act on. In a crisis, that pressure could make a nuclear strike seem necessary. It need never reach for the button.

The response starts with acknowledging a psychological reality: All humans, including leaders, have cognitive blind spots that distort their reasoning about catastrophic events. We must build institutional guardrails that protect leaders from their own worst impulses—and follow those guardrails even when pressure builds or they themselves wish to circumvent them. AI can help within that structure, but only as a subordinate tool, not a substitute for it.

Used this way, AI-enabled decision support can analyze targeting data far faster than human analysts—time that should go toward compliance with the laws of armed conflict, whose violation can itself hand the other side grounds to escalate. AI can assist analysts reading sensor feeds, too—but its interpretations may be wrong, and it must keep analysts aware of that. A fuel truck beside a missile could be reported by an AI as “fueling” or “preparing to launch.” The second, more definite claim risks anchoring everything assessed afterward to a false inference.

Guardrails that leaders actually follow under pressure, and AI that expands rather than narrows what humans consider, are candidates for something more general: continuing, significant reductions in the probability of nuclear war. Richard Garwin identified risk reduction as a mechanism to avoid guaranteed Armageddon in the long run—something worth pursuing whether or not abolition is the goal. Abolitionists and skeptics alike can start now.