UN fails to agree on ‘killer robot’ ban — get ready for the arms race

Autonomous weapon systems – commonly known as killer robots – may have killed human beings for the first time ever last year, according to a recent United Nations Security Council report on the Libyan civil war. History could well identify this as the starting point of the next major arms race, one that has the potential to be humanity’s final one.

The United Nations Convention on Certain Conventional Weapons debated the question of banning autonomous weapons at its once-every-five-year review meeting in Geneva Dec. 13-17, 2021, but didn’t reach a consensus on a ban. Established in 1983, the convention has been updated regularly to restrict some of the world’s cruelest conventional weapons, including land mines, booby traps, and incendiary weapons.

Autonomous weapon systems are robots with lethal weapons that can operate independently, selecting and attacking targets without a human weighing in on those decisions. Militaries around the world are investing heavily in autonomous weapons research and development. The U.S. alone budgeted US$18 billion for autonomous weapons between 2016 and 2020.

Meanwhile, human rights and humanitarian organizations are racing to establish regulations and prohibitions on such weapons development. Without such checks, foreign policy experts warn that disruptive autonomous weapons technologies will dangerously destabilize current nuclear strategies, both because they could radically change perceptions of strategic dominance, increasing the risk of preemptive attacks, and because they could be combined with chemical, biological, radiological and nuclear weapons themselves.

As a specialist in human rights with a focus on the weaponization of artificial intelligence, I find that autonomous weapons make the unsteady balances and fragmented safeguards of the nuclear world – for example, the U.S. president’s minimally constrained authority to launch a strike – more unsteady and more fragmented. Given the pace of research and development in autonomous weapons, the U.N. meeting might have been the last chance to head off an arms race.

Lethal errors and black boxes

I see four primary dangers with autonomous weapons. The first is the problem of misidentification. When selecting a target, will autonomous weapons be able to distinguish between hostile soldiers and 12-year-olds playing with toy guns? Between civilians fleeing a conflict site and insurgents making a tactical retreat?

The problem here is not that machines will make such errors and humans won’t. It’s that the difference between human error and algorithmic error is like the difference between mailing a letter and tweeting. The scale, scope, and speed of killer robot systems – ruled by one targeting algorithm, deployed across an entire continent – could make misidentifications by individual humans like a recent U.S. drone strike in Afghanistan seem like mere rounding errors by comparison.

Autonomous weapons expert Paul Scharre uses the metaphor of the runaway gun to explain the difference. A runaway gun is a defective machine gun that continues to fire after a trigger is released. The gun continues to fire until the ammunition is depleted because, so to speak, the gun does not know it is making an error. Runaway guns are extremely dangerous, but fortunately, they have human operators who can break the ammunition link or try to point the weapon in a safe direction. Autonomous weapons, by definition, have no such safeguard.

Importantly, weaponized AI need not even be defective to produce the runaway gun effect. As multiple studies on algorithmic errors across industries have shown, the very best algorithms – operating as designed – can generate internally correct outcomes that nonetheless spread terrible errors rapidly across populations.

For example, a neural net designed for use in Pittsburgh hospitals identified asthma as a risk-reducer in pneumonia cases; image recognition software used by Google identified Black people as gorillas; and a machine-learning tool used by Amazon to rank job candidates systematically assigned negative scores to women.

The problem is not just that when AI systems err, they err in bulk. It is that when they err, their makers often don’t know why they did and, therefore, how to correct them. The black box problem of AI makes it almost impossible to imagine a morally responsible development of autonomous weapons systems.

The proliferation problems

The next two dangers are the problems of low-end and high-end proliferation. Let’s start with the low end. The militaries developing autonomous weapons now are proceeding on the assumption that they will be able to contain and control the use of autonomous weapons. But if the history of weapons technology has taught the world anything, it’s this: Weapons spread.

Market pressures could result in the creation and widespread sale of what can be thought of as the autonomous weapon equivalent of the Kalashnikov assault rifle: killer robots that are cheap, effective, and almost impossible to contain as they circulate around the globe. “Kalashnikov” autonomous weapons could get into the hands of people outside of government control, including international and domestic terrorists.

Credit: Ministry of Defense of Ukraine