First, about how killer robots work:
“Project Maven. By Katrina Manson. W.W. Norton; 416 pages; $31.99 and £23
THE MAVEN Smart System is perhaps the most important weapon system you have never heard of. It has spotted Iranian missiles heading for Israel and rocket launchers in Yemen. It has detected migrants crossing America’s southern border and drug boats in the Caribbean. On one day in 2022 it found more than 260 potential targets for Ukraine.
And Maven does not just sense such things: it can also co-ordinate the response. It fuses together different sorts of intelligence—photos, text, radio and electromagnetic pulses—and works out which plane, carrying which munitions, is closest to which target. With a single click by a human, Maven can turn data into ash. As one NATO official tells Katrina Manson, “This is the Microsoft Windows of warfighting.” It may well become to armed forces what Microsoft is to office drones.
In “Project Maven”, Ms Manson, a national-security journalist at Bloomberg, has written one of the most important books on war and technology in many years. It is a scintillating account of how Drew Cukor, a hard-driving marine, created a team of mavericks who put AI at the heart of America’s war machine—even though that team was often at war with itself and with the rest of the Pentagon.
It is also a story of Silicon Valley’s shifting relationship with war and the Pentagon. Google walked out of Maven in 2018 after employee protests over the tech giant collaborating on lethal tools.
Palantir later became the single most important firm in Maven’s development.
It started out as a project harnessing AI to find objects in reams of drone footage, a job that previously consumed huge quantities of manpower. Early algorithms, tested during counter-terrorism operations in Somalia in 2017, were erratic. Algorithms would label clouds as flying school buses. The next year, in Afghanistan, the software identified trees as people and rocks as buildings. But it improved.
On one occasion, just as American forces were about to launch a drone strike, an analyst spotted a shepherd walking into the field of view. It had taken him 40 seconds to notice the man; when they tested Maven on the same video feed, it took the AI less than a second to detect him.
In 2019 Maven was used during the operation to kill Abu Bakr al-Baghdadi, the head of Islamic State, in Syria, and Qassem Suleimani, an Iranian general, by drone strike in 2020. When Joe Biden pulled American forces out of Kabul the next year, Maven would work out how many people were thronging the airfield during the mad scramble.
However, it was events in Ukraine that was the pivotal moment for Maven. The targeting tool was feeding what America euphemistically called “points of interest” to Ukraine on an industrial scale, at the cost of $1m per month in cloud-computing bills.
“Sometimes it felt as though the US was all but punching co-ordinates into the weapons systems themselves,” Ms Manson writes.
AI is context-specific, the book makes startlingly clear. Models that had 70% success rates in Afghanistan dropped to 30% in the Philippines, where people walked in front of thick green jungle rather than dusty yellow ground. Similarly, the algorithms struggled to cope with Ukrainian snow and Russian tanks with their turrets blown off. Good data was crucial.
Training a single algorithm might take 10,000 images, each one accurately labelled.
In 2021-22 more than 1,500 algorithms were whittled down to just two dozen for use in Ukraine.
When Maven was launched, Pentagon officials portrayed it as an intelligence tool, far from the messy and controversial work of targeting. Ms Manson’s book dispels this notion. From the start, Maven was intended to speed up America’s kill chain: the process of finding targets, deciding what to do and conducting an attack. One official at the National Geospatial-Intelligence Agency tells Ms Manson that large-language models have speeded up the targeting process five-fold, allowing America to identify and hit 5,000 targets per day.
“Ultimately,” says General Chris Donahue, who pioneered Maven’s use in Ukraine and now serves as the commander of American land forces in Europe, “all this stuff will become automated.”
That raises profound questions around human control of war. In a conflict between America and China or Russia, each side would face enormous pressure to strike quickly and decisively, before the other side could find targets and launch weapons. Even if people are overseeing the process, can they keep up, particularly if something goes wrong? Long after Maven was deployed in Ukraine, it was still producing ten incorrect detections for every square kilometre it assessed.
Now, as well as operating on distant servers, the algorithms developed by Maven sit inside weapons. The book describes two little-known weapons intended to overwhelm and slow down Chinese forces in any war in the Pacific.
“Goalkeeper” is a loitering munition, or suicide drone; “Whiplash” is an explosives-laden jet ski, whose early versions were smuggled into Ukraine by the CIA for testing. Each type could be sent to find and attack targets on its own.
That makes thorough data more important than ever. When Mavenites tried to build a submarine-hunting algorithm, they found that P-8 sub-hunting planes wiped their hard drives at the end of each mission. After some resistance from America’s navy, such data were acquired and supplemented with hundreds of thousands of inputs from “boat cameras, port cameras, infrared systems, destroyers, combat ships, buoys, dhows” and more. One defence official says “We’re basically watching the [People’s Liberation Army] all the time to get AI training data.”
Ms Manson has conducted scores of interviews with the people who built and use Maven, as well as its opponents. Mr Cukor emerges as an intriguing character—reprimanded by the Marine Corps for overseeing a toxic culture within Maven—whose intensity and drive shaped the programme’s success. He wants decision-makers in Washington to use his innovations wisely. “Let’s be able to look ourselves in the mirror and make sure we are careful,” he says. “We have all this tech; are we the best custodians of it?”
What is the role of nuclear weapons in the age of killer robots? It depends on deciders’ understanding of the damage that killer robots are making. If the damage seems too dangerous for the deciders, nuclear devices are used to kill the center of the danger. Paradoxically Ukraine might survive since Russians consider inhabitants of Ukraine as the same Russians. China also considers Taiwan inhabitants as Chinese. Many other countries are not so lucky.
In the age of lethal autonomous weapons, nuclear arsenals increasingly serve as existential "escalation breaks." If robotic swarms or algorithmic conflicts spiral out of control, deciders may resort to high-yield nuclear devices against the adversary's command centers or swarm production hubs.
The paradox regarding populations in conflicts like Ukraine and Taiwan reflects complex geopolitical dynamics, but relies heavily on the specific strategic and territorial calculations of the combatants.
• Russia and Ukraine: The premise of the conflict relies on political and historical claims of shared identity and cultural kinship.
• China and Taiwan: China considers Taiwan an inalienable part of its territory. While this influences the political calculus to minimize infrastructure destruction, the threat of fully autonomous systems or a naval blockade remains a central focus of defense planning.
• Global Implications: In conflicts where the intervening power views the targeted population as historically or culturally separate, the restraint described may not apply.
The integration of artificial intelligence into the command structures of nuclear-armed states significantly lowers the threshold for miscalculation. When machines operate at speeds that outpace human reasoning, the risk of accidental nuclear escalation rises dramatically. Consequently, the ultimate role of nuclear weapons is shifting from Cold War deterrence of human actors to a desperate, last-resort fail-safe against runaway autonomous systems. You can read more about these automated warfare challenges in the Stockholm International Peace Research Institute Analysis on AI and nuclear risk.
1. The AI that transformed warfare. The Economist; London Vol. 459, Iss. 9499, (May 16, 2026): 90.
Komentarų nėra:
Rašyti komentarą