In the minutes after a launch detection or nuclear detonation, would America’s nuclear command, control, and communications system enable the president to make a timely and accurate decision to retaliate? We do not know.
In a 2019 War on the Rocks article, “America Needs a ‘Dead Hand’,” we proposed the development of an artificial intelligence-enabled nuclear command, control, and communications system to partially address this concern. In the five years since the article was published, China began an unprecedented expansion of its nuclear arsenal; Russia invaded Ukraine, made repeated nuclear threats, and “suspended” Russian participation in the New START arms control treaty; and North Korea launched a massive expansion of its nuclear arsenal. The United States has not expanded its arsenal by a single weapon or fielded a single new delivery vehicle.
Today, we believe that United States is further behind China and Russia as both nations are modernizing and expanding their nuclear arsenals and fortifying their nuclear command, control, and communications systems. This widening gap in capabilities only increases the coercive power Russia and China have to coerce the United States into backing down from aggression.
We can only conclude that America needs a dead hand system more than ever. Such a system would both detect an inbound attack more rapidly than the current system and allow the president to either manually direct forces to respond or automatically execute the president’s pre-selected response options — for a given scenario.
Previous discussions of artificial intelligence’s utility for nuclear command, control, and communication systems often attribute capabilities and characteristics to artificial intelligence that are distinctly different from anything that is useful in a fielded system. In reality, an artificial intelligence-enabled nuclear command, control, and communications system with a dead hand capability will never turn into the 1983 movie War Games’ War Operation Plan Response or Terminator’s Skynet because it is possible to employ artificial intelligence with discrete capabilities that do not evolve into a sentient system.
In short, the systems we suggest here are very specifically designed to increase the speed at which the United States can detect an adversary attack, speed the president’s ability to respond to an inbound strike, and ensure fielded forces receive the president’s orders. And, if necessary, the president can pre-program his desired response to a number of scenarios and allow the system, when on automatic, to match the strike detected with the president’s desired response. The relative decline of the American nuclear arsenal to those of Russia, China, and North Korea makes such a system a necessity because it complicates adversary calculations when contemplating a nuclear strike on the United States.
Artificial Intelligence
Discussions of the utility of artificial intelligence in nuclear command, control, and communications systems often devolves into emotional arguments and accusations of some nefarious desire for a general artificial intelligence that will kill us all. This view fundamentally misunderstands the full breadth of tools that are broadly called artificial intelligence. Some clarification is useful.
The Department of Defense defines artificial intelligence as “the ability of machines to perform tasks that normally require human intelligence — for example, recognizing patterns, learning from experience, drawing conclusions, making predictions, or taking action — whether digitally or as the smart software behind autonomous physical systems.”
Artificial intelligence is a generic term often used to lump several concepts together, even when they may not technically qualify as artificial intelligence. Machine learning, natural language processing, expert systems, and neural networks all have utility for the system we seek to design and field. For example, the chatbots that often pop up on a company’s website to answer questions use a combination of pattern-matching, Naïve Bayes, sequence-to-sequence models, recurrent neural networks, long short-term memory, and natural language processing algorithms.
When combined, these tools are useful in automating decision-making. Each has strengths and weaknesses and is useful for some purposes and not for others. This is why they are combined in these larger systems.
All forms of artificial intelligence are premised on mathematical algorithms, which are defined as “a set of instructions to be followed in calculations or other operations.” Essentially, algorithms are programming that tells the model how to learn on its own.
They are only as good as their designers, so building the right design team is critical. Leave out the social scientists, military operators, marketers, and others and data scientists, programmers, and mathematicians build a deeply flawed and biased model. This discussion is important because it distinguishes a system enhanced by various artificial intelligence tools and a general artificial intelligence, which we oppose.
Rather, the system envisioned is something similar to a concept known as the Rational Behavior Model developed by the Naval Postgraduate School for use in autonomous naval vessels. In the Rational Behavior Model, artificial intelligence is used in components of the system related to understanding sensor input, but artificial general intelligence is never used.
Top-level decision-making in the Rational Behavioral Model is performed by pre-planning in order to develop a set of scenarios and pre-planned decision outcomes for each scenario. As we explain, for our system the president participates in a presidential decision conference in which a comprehensive set of scenarios are studied, and a decision is made for each one.
Understanding the Problem
According to the Department of Defense’s Nuclear Matters Handbook, nuclear command, control, and communications “performs five critical functions: situation monitoring; planning; decision-making; force direction; and force management.” These activities include “detection, warning, and attack characterization; nuclear planning; decision-making conferencing; receiving presidential orders; and enabling the management and direction of forces.”
In 2016, the Air Force designated its nuclear command, control, and communications system of systems as the AN/USQ-225 weapon system in an effort to consolidate an otherwise disparate set of legacy systems into something more manageable. This system must detect an adversary launch, decide on an appropriate response, and direct the force into action. The time in which this detect, decide, and direct process takes place is severely compressed by more modern, multidomain adversary capabilities.
The attack time compression challenge is driven by the development of new types of weapons and the sheer number of capabilities adversaries are developing and fielding. Russia, China, and North Korea field an authoritarian arsenal that, when combined, dwarfs the arsenal of democracy in strategic and non-strategic nuclear weapons. Russia and China already have the capability to move up and down Herman Kahn’s escalation ladder in a way the United States cannot.
The attack time compression challenge is nothing new for the United States. As Guide to Nuclear Deterrence in the Age of Great-Power Competition explains, this very challenge was responsible for the creation of today’s nuclear command, control, and communications system. Currently, the NC3 Enterprise Center at U.S. Strategic Command is grappling with the challenge of replacing legacy systems with modern systems that are not susceptible to adversary interference.
Jonathan Falcone and his coauthors address some of the challenges with building an artificial intelligence-enabled nuclear command, control, and communications systems. Philip Reiner and Alexa Wehsener also address these challenges and offer a conservative approach to integrating artificial intelligence into nuclear command, control, and communication systems.
As Reiner and Wehsener write, “We are convinced that time is most usefully spent debating the technical positives and negatives of such integration in a manner that does not simply classify perspectives on the discussion as ‘that’s crazy’ or ‘just don’t,’ or as vaguely as stating that there is a need for an ‘automated strategic response system based on artificial intelligence.’”
Speeding up the detection of an enemy attack and the decision-making process is generally recognized as a necessity. In many respects, this is how to achieve this goal where there is disagreement.
The Dead Hand
The “dead man switch” was first developed as a “fail-safe” mechanism with the introduction of electric trams/streetcars in the United States at the beginning of the 20th century. Dead man switches are now a common safety feature in everything from snowmobiles to Tesla automobiles.
America is no stranger to “fail-fatal” systems either. The Special Weapons Emergency Separation System, also known informally as the dead man’s switch, was a nuclear bomb detonation system built into early B-52 bombers during the Cold War. It ensured that in case of crew incapacitation due to enemy defenses, the nuclear weapons would still detonate once the aircraft dropped below a pre-set altitude.
The Soviet/Russian Perimeter system, when activated, maintains constant communication with the command authority. If contact is lost between the leadership and fielded forces, the system is designed to take this loss to mean an American decapitation strike has taken out the command authority. This leads to the launch of missiles equipped to transmit authorization codes to nuclear forces, which then launch against the United States.
Although there is disagreement, it appears the Soviet Union created Perimeter in order to both deter the United States from targeting the Soviet leadership, part of the American counterforce targeting strategy, and give the Soviet leadership an opportunity to ride out a first strike without losing the opportunity for a second strike. In other words, Perimeter’s existence, at least theoretically, gave Soviet leaders confidence in their secure second-strike capability, in the event of their deaths.
An American dead hand would serve a similar purpose. As an artificial intelligence-enabled system, it is far more accurate to see the dead hand as a system of systems with a wide variety of algorithms embedded into the systems that comprise the nuclear command, control, and communications system. With a combination of hard coding and algorithms, such a system would employ algorithms where it makes sense to perform discrete functions — like assessing data from space-based overhead persistent infrared systems. This is but one example of how a process — evaluating infrared signatures to determine if they are indicative of an intercontinental ballistic missile launch — can be turned over to a set of algorithms that process data faster than the current system. It is also worth noting that the current system of systems collects abundant data, which can serve to train future algorithms on discrete tasks.
Keep in mind, where artificial intelligence tools are embedded in a specific system, each function is performed by multiple algorithms of differing design that must all agree on their assessment for the data to be transmitted forward. If there is disagreement, human interaction is required.
Such a design feature is akin to James Reason’s Swiss Cheese Model of human error prevention. In short, each algorithm acts like a slice of Swiss cheese, with imperfections that are different from other algorithms. When stacked together, the imperfections in each algorithm do not perfectly align to allow a fatal error.
The United States has seen about three dozen accidents involving nuclear weapons since 1950. None of these accidents led to an accidental nuclear detonation because the weapons were built with redundant safety features. The current nuclear command, control, and communications system is also designed with redundancy in mind.
For example, the United States relies on dual phenomenology to determine whether a suspected intercontinental ballistic missile launch is, in fact, an actual launch. It does this by employing space-based infrared systems to detect the infrared signature of a missile launch and then verifies that signature through the use of long-range radars. Two very different systems employ different methods to reach a correct conclusion.
This is just one example of how any nuclear command, control, and communications system should have redundancies that prevent single points of failure. Embedding algorithms into component systems does not change that requirement.
Presidential Decision-Making
According to the Congressional Research Service’s Defense Primer: Command and Control of Nuclear Forces:
The US president has sole authority to authorize the use of US nuclear weapons. This authority is inherent in his constitutional role as Commander in Chief. The President can seek counsel from appropriate military advisors; those advisors are then required to transmit and implement the orders authorizing nuclear use. The President does not need the concurrence of the US Congress to order the launch of nuclear weapons, and neither the military nor Congress can overrule these orders.
The real challenge for the president is that “the president would have less than 10 minutes to absorb the information, review his options, and make his decision.” In other words, after space and terrestrial warning systems detect and verify that nuclear weapons are headed for the United States, the president, while also trying to move to a safe location (White House bunker or Air Force One), will have to get on phone with the secretary of defense and the commander of U.S. Strategic Command to decide on a response — possibly before American nuclear bases are destroyed. He will do all of this without ever having practiced for the event.
While the national command authority exercises regularly, the president is not a participant in those exercises and does not hone his understanding of nuclear weapons employment, targeting, effects, or other critical components of their use. He is reliant on the advice of the military. This is on-the-job training at its worst.
Not only is the current nuclear command, control, and communications system inadequate for the challenges on America’s doorstep, but the presidential decision-making process is woefully inadequate. Thus, an artificial intelligence-enabled system should not only speed up the detect, decide, and direct process but it should also aid the president in improving decision-making.
A process for improving presidential decision-making during a nuclear attack could take the following form. After the president’s election, in a time of peace, the president sits down with advisors and walks through a variety of potential adversary first-strike scenarios. During this process, the president decides on preferred response options for each scenario. These decisions are then input into the system, which either operates manually or automatically.
When the system is in manual mode, it does not function as a dead hand system, but it remains an artificial intelligence-enabled system that more rapidly processes through early-warning data. Once the system detects an inbound attack and a presidential decision conference is initiated, the system can also present the president with the response option that most closely fits his pre-conflict decision. Such a capability has value because it can aid the president in avoiding the bad decision-making that often accompanies high-stress situations.
When the system is in automatic mode, it can detect inbound weapons, match the adversary strike to the president’s closest pre-selected response option, and authorize the execution of a response. Here, the system functions as a smart dead hand, with the ability to match the president’s preferred response to any given strike.
If, for example, the United States found itself in a nuclear crisis, the president could move the system from manual to automatic and notify the adversary that the American dead hand system will automatically respond to any strike against the United States. The objective is to encourage adversary restraint and enable de-escalation but, should that fail, the nation can rapidly respond to a nuclear attack.
For some, our discussion of a nuclear command, control, and communications system that incorporates artificial intelligence tools when and where appropriate, but avoids giving artificial intelligence decision-making authority, may make sense. For others, however, may see presidential pre-selection of response options in an automated command-and-control system as a bridge too far. The idea that a president will agree to pre-select nuclear response options may seem implausible.
Under current conditions, this may be true, but given the lack of experience with such an option it is impossible to know. However, it does not negate the fact that the system and approach offered here improves upon the existing method by including the president in decision-making long before the president is given a few minutes to determine how to respond to an inbound nuclear attack. If, for example, the president began participating in nuclear exercises, and is faced with simulated conditions like those in a real event, the shortcomings of the current system should become evident.
The president has no more important responsibility than the deployment and employment of nuclear weapons. If that day should ever come, our goal is to ensure the sole decision-maker has the best chance to succeed in limiting damage to the United States.
Conclusion
As mutual interests drive Russia, China, and North Korea closer together, their combined nuclear arsenals pose a growing threat to the United States. Rather than fielding a nuclear arsenal sized and designed to mitigate this expanding risk, the nation is following a path laid out more than a decade ago — when times were very different. In other words, the United States is not fielding the array of capabilities needed to effectively mitigate the increasing attack time compression challenge.
Ensuring that Russia and China understand that the United States can always respond to a nuclear strike is critical. The development of an artificial intelligence-enabled nuclear command, control, and communications system — with the ability to either speed up presidential decision-making (manual mode) or respond automatically — is one way to address this problem.
Adam Lowther, Ph.D., is the vice president for research at the National Institute for Deterrence Studies and host of the Nuclecast podcast. Col. (Ret.) Curtis McGiffin is the vice president for education at the National Institute for Deterrence Studies and a visiting assistant professor at Missouri State University. Together, they have over four decades of experience in the Department of Defense nuclear enterprise.
Image: Senior Airman Ashley Boster
Comments are closed.