‘I’m afraid I can’t do that’: Should killer robots be allowed to disobey orders?

It is often said that autonomous weapons could help minimize the needless horrors of war. Their vision algorithms could be better than humans at distinguishing a schoolhouse from a weapons depot. They won’t be swayed by the furies that lead mortal souls to commit atrocities; they won’t massacre or pillage. Some ethicists have long argued that robots could even be hardwired to follow the laws of war with mathematical consistency.

And yet for machines to translate these virtues into the effective protection of civilians in war zones, they must also possess a key ability: They need to be able to say no.

Consider this scenario. An autonomous drone dispatched to destroy an enemy vehicle detects women and children nearby. Deep behind enemy lines, without contact with its operator, the machine has to make a decision on its own. To prevent tragedy, it must call off its own mission. In other words, it must refuse the order.

“Robot refusal” sounds reasonable in theory. One of Amnesty International’s objections to autonomous weapons is that they “cannot … refuse an illegal order”—which implies that they should be able to refuse orders. In practice, though, it poses a tricky catch-22. Human control sits at the heart of governments’ pitch for responsible military AI. Giving machines the power to refuse orders would cut against that principle. Meanwhile, the same shortcomings that hinder AI’s capacity to faithfully execute a human’s orders could cause them to err when rejecting an order.

Militaries will therefore need to either demonstrate that it’s possible to build ethical, responsible autonomous weapons that don’t say no, or show that they can engineer a safe and reliable right-to-refuse that’s compatible with the principle of always keeping a human “in the loop.”

If they can’t do one or the other, or find a third way out of the catch-22, their promises of ethical and yet controllable killer robots should be treated with caution.

Last year, 54 countries, including the United States, United Kingdom, and Germany, signed a political declaration insisting that autonomous weapons and military AI will always operate within a responsible human chain of command and control.” That is, autonomous weapons will act solely, and strictly, under the direct orders of their human overseers. Others, like the Russian Federation, have similarly stressed that autonomous weapons must possess a capacity to discriminate between legal and illegal targets. And yet if machines can outperform humans as moral experts, as one recent study (and multiple states) claim, they’ll be hard pressed to apply that expertise if they can only ever agree with their human users.

Allowing a machine to say “I won’t do it” when presented with an order would sharply curtail that authority. It would give machines an ultimate decision-making power that no state would relinquish to a computer. It might be possible to instruct a weapon to avoid hurting civilians while obeying its other orders. But this implies that it will be making legal judgments on things like proportionality and discrimination—however, the International Committee of the Red Cross, among others, argues that only humans can make legal judgments because only humans can be held legally responsible for harms.

In short, the idea of autonomous weapons that refuse orders would be a non-starter. Armies want machines that act as extensions of their will, not as a counterweight to their intent. Even countries that hold themselves to the highest legal standards—standards which rely on all soldiers being empowered to say no when given a bad order—would probably balk at the proposition. Meanwhile, militaries that intend to deliberately violate the law won’t want machines that comply rigidly with the law.

As such, the killer robots that countries are likely to use will only ever be as ethical as their imperfect human commanders. Just like a bullet or a missile, they would only promise a cleaner mode of warfare if those using them seek to hold themselves to a higher standard. That’s not a particularly comforting thought.

A killer robot with a “no mode” could also pose its own ethical risks. Just as humans won’t always give virtuous orders to the machines, machines wouldn’t always be right when they turn around and say no. Even when autonomous weapons get very good at differentiating combatants from children, for example, they will still sometimes confuse one with the other. They might not be able to account for contextual factors that make a seemingly problematic order legal. Autonomous weapons can also be hacked. As the war in Ukraine has shown, for example, techniques for jamming and commandeering drones are sophisticated and fast-evolving. Additionally, an algorithmic no mode could be exploited by cunning adversaries. A military might, for example, disguise its artillery positions with camouflage that tricks an autonomous weapon to mistake them for civilian structures.

In other words, under some conditions a weapon might refuse a legal order. Given the formidable capabilities of these systems, that’s an unacceptable risk. There’s a reason that the most frightening part of 2001: A Space Odyssey is when the computer, HAL 9000, refuses astronaut David Bowman’s orders to let him back onto the ship, since doing so would jeopardize the mission assigned to it by the upper command. The line is a poignant warning for our times: “I’m sorry, Dave, I’m afraid I can’t do that.”

Therein lies the paradox. Giving autonomous weapons the ability to accurately distinguish the right course of action from the wrong one—without bias or error—is held as a minimum criteria for allowing these machines onto the battlefield. Another criteria—predictability—holds that machines should only do what their commanders want and expect them to do. These principles of AI ethics are assumed to be compatible with the fundamental principle of control. The reality could be more complicated than that. Humans want autonomous machines to be noble, but can they allow the machine to be nobler than they are and still say that they are in charge?

It is, of course, possible that militaries could find a third way. For example. Maybe it will be enough for autonomous weapons to simply ask their commander “are you sure?” before proceeding with an order. Experts in international law have argued that if AI weapons can be engineered to prompt humans to reconsider their decisions, without over-riding the orders themselves, they might enable militaries to make more legally sound choices in the heat of conflict. An AI system could, for example, urge a fired-up commander who is planning a hasty airstrike in a densely populated area to think twice, or seek a second human opinion, before pulling the trigger. Though this type of feature wouldn’t work in all contexts (for example, in the case of the drone that has no contact with its operator) it could, in some settings, save lives.

Militaries might not even be able to stomach such a modest proposition. After all, they want AI to accelerate their decision-making, not slow it down. But if states are dead set on using AI for war, and if they’re honest about wishing to reduce the misery of conflict, these are questions that they have no choice but to take seriously.