Book summary “Army of none”: summary

Introduction

You face a future where lethal decisions on the battlefield are delegated to machines. As military robotics proliferate, you must navigate the moral, legal, and operational risks of removing human judgment from warfare. Historical close calls, such as the 1983 false alarm averted by Stanislav Petrov, demonstrate the extreme danger of automated systems misinterpreting data under high-stakes pressure.

Fully autonomous weapons remove human judgment from lethal combat, creating severe technological, ethical, and strategic risks that demand strict governance and human-machine teaming.

Paul Scharre examines these challenges in Army of None by tracing the technological evolution of weapon automation from early mechanisms to modern artificial intelligence. The book explores current military investments in swarming drones, system failures, and high-reliability safety models. You will find answers to several critical questions:

• How do autonomous systems fail in unpredictable environments?
• What ethical boundaries prevent machines from making life-and-death decisions?
• How can militaries balance technological advantage with strategic stability?

1. The Military Robotics Revolution

The rapid proliferation of drones and ground robots creates personnel bottlenecks, driving the military toward cooperative autonomy and swarming tactics.

Driven by counterinsurgency demands and the proliferation of improvised explosive devices in Iraq and Afghanistan, spending on unmanned systems skyrocketed. Drones shifted from minor surveillance tools to central military assets.

By 2011, annual Department of Defense spending on drones swelled to over $6 billion. That total was more than twenty times pre-9/11 levels, with a fleet exceeding 7,000 drones. The military also deployed 6,000 ground robots to handle tasks like bomb disposal.

Uninhabited aircraft in the 1990s received about $300 million per year. Drone spending reached $2 billion per year by 2005, the year the Department of Defense recognized the robotics revolution and published its first roadmap.

The Coming Swarm

Operating remote vehicles requires extensive human staffing for piloting, sensor operation, and intelligence analysis. Furthermore, contested electromagnetic environments will make satellite links vulnerable to jamming, requiring robots to operate independently.

Operating a single 24/7 continuous drone orbit with Predators or Reapers requires seven to ten pilots, twenty sensor operators, and scores of intelligence analysts.

Cooperative autonomy allows swarms of robots to execute complex tactical maneuvers collectively with minimal human oversight. Using decentralized, hierarchical, or consensus-based algorithms, robot swarms can coordinate actions like traffic deconfliction, targeting, and formation flying much faster than individual human controllers could manage.

Researchers at the Naval Postgraduate School and the Office of Naval Research conducted simulated dogfights and boat escorts using autonomous swarms where a single human command initiated the action while the systems handled tactical execution.

• 20 drones participated in a swarm-on-swarm mock dogfight at Camp Roberts.
• 103 aerial drones were demonstrated by the U.S. in a 2016 collective swarm.
• 119-drone swarm was the subsequent demonstration conducted by China.
• 5 boats were used in the Navy surface swarm demonstration on the James River.

Dimensions of Autonomy

Machine autonomy is not a singular spectrum but consists of three independent dimensions: task complexity, human-machine relationship, and machine intelligence. Evaluating an autonomous system requires analyzing what specific task is being performed, how you interact with the system, and how sophisticated the machine internal decision-making processes are.

Modern automobiles feature various autonomous functions ranging from antilock brakes and collision avoidance to automatic ground collision avoidance systems that override human pilots to rescue disoriented F-16s.

2. History and Evolution of Weapon Automation

Weapon automation has expanded steadily from mechanical firing mechanisms to modern smart munitions. You can trace this trajectory by examining how engineers have steadily delegated physical and cognitive tasks to machines.

From Gatling to Maxim

The Gatling gun began the evolution of weapons automation by automating loading and firing. Richard Gatling devised the gun in 1861 during the American Civil War to reduce army sizes and lower casualties from battle and disease.

Mark Twain was an early enthusiast of the Gatling gun, praising its accuracy and its ability to discharge up to 400 rounds per minute. A trained soldier using a Civil War-era rifled musket achieved a firing rate of 3 rounds per minute. The Gatling gun discharged over 300 rounds per minute, allowing four soldiers to deliver the firepower of over a hundred men.

The invention of the Maxim gun in 1883 replaced manual hand-cranking with recoil energy, establishing the modern fully automatic machine gun. Semiautomatic weapons require a separate pull of the trigger for each bullet. Automatic weapons continue firing as long as the trigger is held down.

Operating them requires strict discipline, proper posture, and short bursts to prevent barrels from overheating and melting. The author carried the M249 Squad Automatic Weapon (SAW) as an Army Ranger. The M249 SAW weighs 17 pounds without ammunition, has a cyclic rate of fire of 800 rounds per minute, and can glow red hot under heavy firing. The M2 .50 caliber heavy machine gun weighs 80 pounds, and its tripod weighs 50 pounds. During the Battle of the Somme, the British lost 20,000 men in a single day.

Homing Munitions and Autonomy

Precision-guided munitions introduced onboard guidance systems to correct aiming errors and strike targets from long distances. Early smart weapons included acoustic homing seekers, such as the German Falcon torpedo in World War II, which listened for merchant ships using passive acoustic sensors. The German G7e/T4 Falke Falcon torpedo was introduced in 1943 and activated its acoustic sensors 400 meters after launch. This sparked an arms race of smart weapons and countermeasures, such as acoustic decoys towed behind ships known as Foxer.

German U-boats U-972 and U-377 were sunk in late 1943 and early 1944 when their own homing torpedoes circled back and homed in on the sound of their propellers. Homing munitions use passive or active sensors to sense the environment and maneuver toward a target. Because their search capability is tightly constrained in time and space, they must be part of a broader weapon system or battle network that provides pre-existing targeting data.

U.S. pilots refer to launching the AIM-120 AMRAAM missile in lock-on-after-launch mode as going maddog, where the active radar seeker goes after the first target it sees like an attack dog. Homing munitions possess limited autonomy to track moving targets, but they are not fully autonomous weapons because a human still selects the specific target.

The Weapon System and Supervised Systems

A weapon system comprises a sensor, a decision-making element, and a munition. Weapons are categorized by their degree of autonomy based on whether humans are in the loop, on the loop, or out of the loop during the engagement cycle:

• Semiautonomous weapon systems keep a human in the loop to decide which targets to engage.
• Supervised autonomous weapon systems place the human on the loop to monitor automated defenses in real time.
• Fully autonomous weapon systems remove the human from the loop entirely.

At least thirty nations employ supervised autonomous weapon systems like the U.S. Phalanx CIWS or Patriot to defend against incoming rockets, missiles, and mortars. It is freedom of operation, rather than raw intelligence, that defines an autonomous weapon. Fully autonomous weapons can search for, decide to engage, and attack targets over a wide area without requiring prior precise intelligence on specific targets.

The Israeli Harpy loitering munition can stay aloft for hours and independently choose which enemy radar to destroy within a designated general area. The Harpy has a search time capability of 2.5 hours and covers a maximum range of 500 kilometers.

Why Aren't There More Autonomous Weapons?

Militaries have been hesitant to adopt fully autonomous weapons because expensive munitions risk being wasted or striking non-target assets when operational conditions change.

Modern missiles cost upwards of a million dollars apiece. The U.S. Navy deployed the Tomahawk Anti-Ship Missile from 1982 to 1994 before retiring it because of a lack of confidence in how targeting pictures would shift and the risk of hitting neutral merchant vessels. The TASM could hunt for Soviet ships independently but lacked the ability to discriminate them from merchant vessels. Commanders are reluctant to fire expensive, non-reusable autonomous weapons without confirmed targeting data due to cost and uncertainty.

3. The Future Being Built Today

Defense research agencies and military strategies like the Third Offset Strategy heavily invest in advanced artificial intelligence, robotic swarms, and cognitive electronic warfare.

Annual U.S. defense spending exceeds six hundred billion dollars, a total greater than the next seven countries combined. In 2014, the U.S. military launched the Third Offset Strategy to counter rivals' numerical advantages. You can see this initiative as a modern successor to Cold War investments in nuclear weapons and precision-guided munitions.

DARPA and Innovation

President Dwight D. Eisenhower founded ARPA in 1958 following the Soviet Union's launch of the Sputnik satellite. Today, DARPA operates with a unique culture separate from the traditional military-industrial complex. It focuses on high-risk, high-reward technological investments to prevent strategic surprise. Rather than building completed weapon systems, it creates foundational technologies that have historically enabled breakthroughs like the internet, GPS, and stealth aircraft.

Developed by DARPA's Tactical Technology Office, the Sea Hunter is an entirely unmanned, paradigm-bending warship designed to conduct low-cost anti-submarine warfare in cooperative flotillas. Its long, narrow hull with two outriggers allows it to navigate autonomously. The Anti-Submarine Warfare Continuous Trail Unmanned Vessel forms part of a broader vision for a human-machine collaborative battle fleet.

At the christening ceremony, Deputy Secretary of Defense Bob Work compared the Sea Hunter to a Klingon Bird of Prey. A single Sea Hunter vessel costs two million dollars, while a traditional Arleigh Burke destroyer costs 1.6 billion dollars. This low procurement cost enables the Navy to deploy large numbers of surface vessels for distributed missions.

Target Recognition and AI

Identifying non-cooperative targets in cluttered environments has traditionally generated unacceptably high false-alarm rates for automated systems. DARPA's Target Recognition and Adaption in Contested Environments program attempts to overcome these historical failures in automatic target recognition by applying deep learning to synthetic aperture radar images. By harnessing recent advances in deep neural networks, the program aims to enable airborne platforms to accurately classify military targets independently.

Microsoft researchers demonstrated deep neural networks surpassing human performance by achieving error rates below five percent on standard image datasets. Microsoft's deep neural network achieved an error rate of 4.94 percent in visual object identification on a standard test dataset containing 150,000 images. By comparison, the estimated error rate for humans on the same dataset stood at 5.1 percent.

Collaborative Operations in Denied Environments enables groups of unmanned aircraft to coordinate like wolf packs under single-person supervision. The program addresses communication disruptions in contested electromagnetic environments by allowing drones to share sensor data and divide tasks autonomously. Maximum communication bandwidth back to the human commander in contested areas drops to 50 kilobits per second, a constraint comparable to old dial-up modems. Despite low-bandwidth constraints, human supervisors can still receive low-resolution identification images to approve or disapprove specific strike plans.

Transmitted military object snapshots measure approximately 10 to 20 kilobytes. A 2016 interface video showed a human operator commanding groups of air vehicles such as Disco Group and confirming strike plans for an SA-12 system.

4. Robots Run Amok and Failures in Systems

Complex autonomous systems are subject to normal accidents, brittleness, and catastrophic failures when operating outside expected parameters without human circuit breakers.

Human thinking about autonomous systems is split between two valid intuitions: trust in their superior reliability and precision, and fear of unpredictable behavior and systems running amok. While properly designed autonomous systems can outperform humans in speed and consistency, they suffer from brittleness, meaning they fail badly when placed outside the exact parameters they were programmed for.

During the 2011 Jeopardy! Challenge, IBM's Watson AI stumbled and repeated a wrong answer already given by a human contestant because it lacked the programming to process other players' errors.

Modern systems rely on massive amounts of code. The F-22 fighter jet contains 1.7 million lines of code, while the F-35 jet contains 24 million lines of code. A modern luxury automobile contains 100 million lines of code.

Fully autonomous systems are inherently more hazardous than semiautonomous ones because they lack a human circuit breaker to halt runaway processes when things go wrong. When a machine operates without a human in the loop, unexpected environmental interactions or software flaws can cascade uncontrollably, magnifying the severity of potential failures.

In Disney's Fantasia adaptation of The Sorcerer's Apprentice, Mickey Mouse enchants a broomstick to fetch water without setting a stopping condition, resulting in an unmanageable army of rogue broomsticks flooding the room.

Exhaustive testing of complex autonomous systems is impossible due to the infinite combination of potential states and the limits of human imagination in predicting novel real-world scenarios. Even with millions of simulated scenarios, testers cannot anticipate every environmental condition, human error, or adversary exploit, meaning unexpected negative behaviors will inevitably emerge.

During Gary Kasparov's 1997 chess match against Deep Blue, an unforeseen software bug caused the supercomputer to make a nonsense random move in the forty-fourth move of the game.

Patriot Fratricides

The Patriot missile system incidents during the Iraq War demonstrate how complex automated military hardware can cause tragic friendly fire accidents when operators suffer from automation bias. Army investigators found that Patriot operators exhibited an uncritical trust in automation, ceding control responsibility to the machine and failing to question false radar data or system limitations regarding anti-radiation missiles and ghost tracks.

During Operation Iraqi Freedom, over sixty Patriot fire units were deployed while 41,000 coalition aircraft sorties were flown during the initial phase of the war. A single Patriot unit handled 45 percent of all successful ballistic missile engagements in the war. Yet 25 percent of the twelve total engagements involving Patriots were fratricides. Patriot batteries shot down a British Tornado and a U.S. Navy F/A-18C Hornet, killing the pilots, and were targeted by a U.S. F-16.

Normal Accident Theory

In complex, tightly coupled technological systems, catastrophic failures are inevitable over time because minor, individually manageable faults interact in unforeseen ways. Charles Perrow's normal accident theory explains that tight coupling leaves no slack for human intervention, while system complexity makes internal states incomprehensible to operators during a crisis.

The 1979 Three Mile Island nuclear power plant meltdown began with a leaky seal and escalated due to obscured indicator lights, a stuck pressure release valve, and tight coupling that left operators only thirteen seconds to react to initial failures. Operators required eight minutes before discovering a closed water cooling valve, and two hours before finding a stuck pressure release valve. A 9.0 magnitude earthquake later triggered the 2011 Fukushima Daiichi nuclear disaster through similar systemic vulnerabilities.

5. Command, Decision, and High-Reliability Models

High-reliability organizations utilize rigorous quality assurance, hardware cutouts, and customizable doctrine to maintain safety while managing complex weapon systems.

The U.S. Navy submarine and Aegis communities demonstrate that hazardous systems can be operated safely through continuous quality assurance. You manage safety as an overriding priority rather than relying purely on technological solutions.

Following the 1963 loss of the USS Thresher, the U.S. Navy instituted the SUBSAFE program, resulting in zero SUBSAFE-certified submarine losses over more than half a century. The U.S. Navy force includes 70 submarines, with about one-third at sea at any given time.

You can apply these organizational methods directly:

• Designate critical safe-operation components and subject them to rigorous inspection and testing throughout their life cycle.
• Check, double-check, and check again every critical component against technical specifications upon installation and at every subsequent repair.
• Test weapons doctrine safely using hardware-level cutouts before deployment in theater.

The Aegis Combat System

The Aegis combat system uses highly customizable automation to capture ship captains' intent and maintain human control through distinct engagement modes and hardware safety switches. The system was first fielded in 1983.

Aegis allows commanders to pre-delegate decision-making via software statements called doctrine. These options range from manual control to Auto-Special where human intervention is on-the-loop, while preserving a physical key-operated Fire Inhibit Switch. There are 4 settings for weapons engagements in the Aegis system: manual, Semi-Auto, Auto SM, and Auto-Special.

Captain Pete Galluch of the Aegis Training and Readiness Center demonstrated how doctrine packages can be configured for non-saturation or saturation threats and quickly executed or aborted.

You configure these controls through specific steps:

• Build and test doctrine statements months before deployment based on anticipated missions and intelligence.
• Group doctrine statements into non-saturation and saturation categories for different combat scenarios.
• Insert and turn the Fire Inhibit Switch key to green only when live firing is authorized.

Nuclear Safety Dilemmas

Nuclear weapons management exemplifies the insoluble tension of the always/never dilemma. Systems must always be ready to launch while never allowing unauthorized or accidental detonation.

Despite extensive safety protocols, historical near-misses and false alarms demonstrate the inherent limits of organizational safety. These incidents support normal accident theory over high-reliability theory.

In 1980, a faulty computer chip at NORAD generated a false alarm that indicates 2,200 Soviet missiles were inbound. This error prompted notification of National Security Advisor Zbigniew Brzezinski. Major false alarm incidents occurred at NORAD in 1979 and 1980. Records indicate a minimum of 13 near-use nuclear incidents from 1962 to 2002.

Deep neural networks operate as inscrutable black boxes. Their internal logic is opaque to human designers. Behavior under all possible inputs remains fundamentally unknowable. Because deep neural networks rely on millions or billions of learned parameters, humans cannot fully understand how they reach conclusions. This opacity creates severe challenges for test, evaluation, and verification.

A JASON expert group studying Department of Defense AI applications concluded that the extreme complexity of deep learning makes understanding exact internal operations impossible. The JASON scientific expert group studied AI implications for the Defense Department in 2017.

Visual object recognition AIs possess bizarre vulnerabilities to adversarial images. Adversaries can exploit these flaws to execute hidden spoofing attacks. Deep neural networks misidentify abstract patterns or meaningless static as familiar objects with extreme confidence. Malicious perturbations can be invisibly embedded into normal images.

AI researcher Jeff Clune and his team discovered that evolving images away from decision boundaries tricks neural networks into classifying unrecognizable garbage as objects with greater than 99.6 percent certainty.

Non-legally binding codes of conduct can establish rules of the road to manage crisis stability. These rules prevent unintended interactions between autonomous systems. Simple and self-enforcing rules can govern how autonomous agents interact in unstructured environments during crises. Those rules may collapse once full-scale war breaks out.

You can implement these behavioral boundaries:

• robotic vehicles should not fire unless fired upon.
• return fire must be limited, discriminating, and proportionate.

International humanitarian law obligates humans rather than weapons to make determinations of lawfulness. This requirement demands a minimum degree of human involvement in lethal force decisions.

A principle focusing on human judgment requires specific information regarding targets, weapons, and context. Weapon autonomy must remain strictly bounded in space, time, and scope. The U.S. Department of Defense Law of War Manual specifies that rules on conducting attacks impose obligations on persons rather than the weapons themselves.

You must satisfy specific operational duties:

• maintain human judgment about the lawfulness of an attack.
• retain specific information about the target, weapon, environment, and context.
• ensure the weapon's autonomy is bounded in space, time, targets, and means of attack.

10 Key Ideas

1. Fully autonomous weapons remove human judgment from lethal combat, creating severe technological, ethical, and strategic risks that demand strict governance and human-machine teaming.
2. The rapid proliferation of drones and ground robots creates personnel bottlenecks, driving the military toward cooperative autonomy and swarming tactics.
3. Militaries have been hesitant to adopt fully autonomous weapons because expensive munitions risk being wasted or striking non-target assets when operational conditions change.
4. Defense research agencies and military strategies like the Third Offset Strategy heavily invest in advanced artificial intelligence, robotic swarms, and cognitive electronic warfare.
5. Complex autonomous systems are subject to normal accidents, brittleness, and catastrophic failures when operating outside expected parameters without human circuit breakers.
6. High-reliability organizations utilize rigorous quality assurance, hardware cutouts, and customizable doctrine to maintain safety while managing complex weapon systems.
7. Historical close calls, such as the 1983 false alarm averted by Stanislav Petrov, demonstrate the extreme danger of automated systems misinterpreting data under high-stakes pressure.
8. The U.S. Navy submarine and Aegis communities demonstrate that hazardous systems can be operated safely through continuous quality assurance and strict hardware safety switches.
9. Deep neural networks operate as inscrutable black boxes whose internal logic remains fundamentally unknowable and vulnerable to adversarial images.
10. International humanitarian law obligates humans rather than weapons to make determinations of lawfulness, demanding a minimum degree of human involvement in lethal force decisions.