The possibility of nuclear weapons being placed under automated control of artificial intelligence creates a new and extremely dangerous risk to international security, warn Joshua A. Schwartz of the Carnegie Mellon Institute for Strategy and Technology and Michael K. Horowitz of the Council on Foreign Relations in an analysis for War on the Rocks, News.bg reports.
Currently, nine countries possess nuclear weapons, while arms control treaties, built over decades, are gradually losing their meaning. North Korea is expanding its arsenal, China is increasing its nuclear potential, and Russia is developing new nuclear weapon delivery systems designed to overcome missile defenses.
Against this background, the authors draw attention to another trend - the possibility of decisions to use nuclear weapons being partially or fully automated.
During the Cold War, the Soviet Union developed a system known as „Dead Hand“ or „Perimeter“.
Its purpose was to guarantee the possibility of a nuclear retaliatory strike in the event of the destruction of the Soviet leadership. Under certain conditions, the system could delegate launch authority to a lower-ranking operator.
The authors point out that such a system was essentially designed to deter a nuclear first strike. Today, however, automation could be used for other purposes as well - including coercing adversaries.
Automated threat can be more persuasiveWith conventional nuclear deterrence, there is the problem of the credibility of the threat. If one nuclear state threatens another with a strike, a retaliatory nuclear strike could lead to mutual annihilation.
An automated system could change this logic, because the leader could delegate the decision to a computer system in advance and thus “tie his hands“.
According to the authors, this could make the nuclear threat more credible: the adversary would no longer be able to count on the leader to abandon its implementation at the last moment.
Experiment with British MPsSchwarz and Horowitz conducted an experiment with over 100 sitting British MPs and more than 1,000 citizens.
The participants were presented with a hypothetical scenario in which Russia attacks Estonia and threatens NATO with a nuclear strike if the Alliance interventions.
The difference was in the way the threat was carried out - in one version the decision was controlled by humans, and in the other it was delegated to an automated system, which would react when NATO intervention was detected.
The results showed that the automated threat was perceived as more dangerous. British citizens were over 13% more likely to believe that Russia would use nuclear weapons, and British MPs were about 9% more likely to support a withdrawal from the defense of Estonia.
However, the authors emphasize that the effect is significant but moderate – most participants have not changed their position.
The risk of technical errorHowever, the automation of nuclear command carries enormous dangers.
Computer systems can fail, be hacked, or misinterpret information. During the Cold War, there were several instances where early warning systems mistakenly indicated that a nuclear attack had begun.
In such moments, human judgment prevented a potential catastrophe.
With a system based on artificial intelligence, the risk could be even greater, especially if complex models are used whose decisions are difficult to explain or verify.
A hacker attack on such a system could also have catastrophic consequences, the authors warn.
Danger of nuclear “bluffing“Another problem is that countries may have an incentive to claim that they have an automated nuclear system, even when this is not true.
The mere uncertainty about whether the system exists could be enough to make an adversary to back down.
This makes automation not only a technological but also a psychological tool for nuclear coercion.
What nuclear states can doThe authors suggest several possible measures. Among them is strengthening second-strike capabilities, which would reduce the incentives to create automated systems for fear of a decapitation attack.
Another option is a multilateral international agreement to ban the automated launch of nuclear weapons.
Washington and Beijing have already made an informal commitment to keep decisions on the use of nuclear weapons under human control. According to the authors, such a rule would be stronger if it were turned into an international legally binding agreement and enshrined in national legislation.
NATO also needs to conduct strategic and military simulations to show how the Alliance would respond to an automated nuclear threat.
„Armageddon on autopilot“The authors emphasize that artificial intelligence will inevitably find more and more applications in the military sphere. Autonomous systems and drones can have real advantages, especially in conventional operations.
But automating nuclear command is a different matter altogether.
Technical failures, cyberattacks, misinterpretation of information, and overconfidence in automated systems can turn a mistake into a global catastrophe.
Putting Armageddon on autopilot would create unique risks that cannot be compared to those of conventional weapons. That is why, according to the authors, human control over decisions to use nuclear weapons must remain intact.