Paul Scharre - Four Battlegrounds: Power in the Age of Artificial Intelligence

FOUR
BATTLEGROUNDS

POWER IN
THE AGE OF
ARTIFICIAL
INTELLIGENCE

PAUL SCHARRE

In memory of Jim Hornfischer
Author, agent, mentor, and friend

, not only for Russia, but for all humankind. It comes with colossal opportunities, but also threats that are difficult to predict. Whoever becomes the leader in this sphere will become the ruler of the world.

RUSSIAN PRESIDENT VLADIMIR PUTIN

of global power rivalry.

CHINESE GENERAL SECRETARY XI JINPING

CONTENTS

MAPS

T here was a singular moment when I realized robots would transform warfare. It was the spring of 2007 and I had just begun a one-year tour in Iraq during the height of the U.S. troop surge. Iraq was a chaotic, violent place. The country was ripping itself apart in civil war. Id been on three prior deployments to Afghanistan and my unit had taken casualties every timeId been an Army Ranger and we didnt shy away from combatbut I had never seen violence like this. For my Iraq tour I was part of a civil affairs team deployed to Forward Operating Base Warhorse in Diyala province. Diyalas population was split among Sunnis, Shias, and Kurds, like a microcosm of Iraq. The groups warred with each other and with U.S. troops for the future of the country. Improvised explosive devices (IEDs), or roadside bombs, had taken a heavy toll on U.S. troops in the area. The base had a makeshift monument erected by the 1st Cavalry Division unit operating out of Warhorse with the names of soldiers killed during their tour, and when I arrived in country I was taken aback by the long roster of names stenciled on the concrete barriers. It was a difficult time.

One day on patrol we came across a roadside bomb (as one did at the time). We saw it first, which is the preferred way of discovering one rather than running into it, and we pulled the Stryker armored vehicle over to wait for the bomb disposal techs to arrive. I had occasionally seen U.S. troops handle roadside bombs by simply shooting at them to detonate the bomb while leaning out the side of the Humvee, but waiting for the bomb disposal professionals certainly seemed like the more prudent course of action to me.

It took a while for the explosive ordnance disposal (EOD) team, as they were formally known, to arrive as they were in high demand at the time, but eventually they rolled up in their large Buffalo mine-resistant ambush protected (MRAP) armored vehicle. I had only been in theater maybe a month or so and had yet to see an EOD team in action, so I popped my head out of the Stryker hatch to watch with interest. I expected to see the bomb tech emerge from the MRAP armored vehicle in a giant protective suit. Instead, out came a little robot. And the light bulb went off in my head: Send the robot to defuse the bomb!

It seemed obvious, afterwards, to use a robot for such a hazardous mission. Why put a human at risk if you didnt need to? My mind went to all of the missions in war that might someday be done by robots, sparing the need to put soldiers in harms way.

After I left Iraq, I worked at the Pentagon on emerging technologies and helped push for U.S. investments in military robots. I believed then, and still do, that these machines would help save American lives and defend the nation. The militarys plans for increasingly advanced robots led inevitably to questions about the scope of autonomy in future systems, a thorny issue with important legal, ethical, and strategic dimensions. I was involved in drafting the official Defense Department policy on autonomy in weapons, which gave guidance to weapons developers. In the years following, I had the opportunity to participate in multiple rounds of diplomatic discussions at the United Nations on autonomous weapons, which was the topic of my first book, Army of None. But even as nations debated the role of autonomy in weapons, the ground was shifting beneath our feet.

The robotics revolution that unfolded in the early 2000s, enabling Roombas and military robots, has given way to something far deeper and more profound: a revolution in artificial intelligence (AI). The combination of exponential growth in data and computer processing power (compute) has fueled a renaissance in machine learning, an AI technique in which machines learn from data, rather than follow an explicit set of preprogrammed rules of behavior. Machine learning is now being applied to medicine, finance, transportation, and other industries. Many have argued that AI could be the next industrial revolution. Given the enormous geopolitical disruptions caused by prior industrial revolutions, it makes sense that AI is also a fast-growing arena of competition between nations who are vying for dominance. AI is changing global security and power dynamics. The consequences of this competition may not be as immediate as that of the roadside bomb back in Iraq, but the stakes are just as high. AI is likely to have profound effects on human freedom and global security. Whoever leads in AI will have a tremendous advantage in shaping the future, if the technology can be controlled at all. AI will bring about a new digital order, and this book is the story of those fighting for control over that future.

F ights on, fights on! The two aircraft twisted in the blue, each furiously maneuvering to gain advantage. In one simulated cockpit, a human pilot. In the other, an AI agent.

Less than thirty seconds into the fight, the AI had scored the first kill. The human pilots simulator flashed red as the AIs guns hit their mark. Both aircraft dove furiously, but mere seconds later, the AI scored another hit. Then another. And another.

In under ninety seconds, it was over. The AI had racked up four kills and the human none.

The fight was the culmination of DARPAs AlphaDogfight Trials, which pitted an AI agent against a flesh-and-blood highly experienced F-16 pilot who goes by the call sign Banger. Eight AI teams competed in a tournament against one another, with the thirty-person Heron Systems beating out defense giant Lockheed Martin in the finals. Victory over the other AI agents earned Heron Systems algorithm, call sign Falco, the opportunity to compete 1v1 against a human pilot. It performed flawlessly. In five total rounds, the AI racked up fifteen gunshot kills against Banger. Banger couldnt get a single shot on the AI.

Dogfighting, or that develops technology breakthroughs, and DARPAs Air Combat Evolution (ACE) program is taking air combat to the next level.

Launched in 2019, the capitalizes on the explosion in AI under way since 2012 as part of the deep learning revolution. Improvements in data, computing hardware, and algorithms have enabled rapid improvements in machine learning, a method for creating intelligent machines by training algorithms on data. Deep learning, a machine learning technique that uses deep neural networks, has yielded tremendous successes, including superhuman performance in games such as the Chinese strategy game go and complex computer games such as StarCraft II and Dota 2. With the ACE program, the U.S. military aims to use machine learning to build more effective AI assistants for human pilots. Not only did AlphaDogfight achieve superhuman performance in a simulated environment, it demonstrated some of the superior strengths of AI systems in combat.

Heron Systems algorithm, Falco, displayed superhuman precision in its flying and fighting, giving it an unbeatable edge over its human opponent. In an aerial dogfight, planes circle one another in tight spirals seeking to gain a positional advantage over their opponent so they can maneuver into a kill shot with short-range missiles or guns. Carefully managing the aircrafts energy so they can turn tighter and faster than the enemy is a critical skill in dogfighting. Pilots must fly the aircraft at the peak of its performance. This can be challenging for humans, who must not only pay attention to the velocity and position of their aircraft but also keep track of the enemy fighter and anticipate next moves. The AI agent , keeping the aircraft at the peak of its performance and never wasting energy in suboptimal maneuvers.