On a green hill overlooking the tree-lined perimeter of Daejeon, a city in central South Korea, a machine gun turret idly scans the horizon. It’s about the size of a large dog; plump, white and wipe-clean. A belt of bullets – .50 calibre, the sort that can stop a truck in its tracks – is draped over one shoulder. An ethernet cable leads from the gun’s base and trails through the tidy grass into a small gazebo tent that, in the Korean afternoon heat, you'd be forgiven for hoping might contain plates of cucumber sandwiches and a pot of tea.
Instead, the cable slithers up onto a trestle table before plunging into the back of a computer, whose screen displays a colourful patchwork of camera feeds. One shows a 180-degree, fish-eye sweep of the horizon in front of us. Another presents a top down satellite view of the scene, like a laid-out Google Map, trained menacingly on our position.
A red cone, overlaid on the image, indicates the turret’s range. It spreads across the landscape: four kilometres-worth of territory, enough distance to penetrate deep into the city from this favourable vantage point. Next to the keyboard sits a complicated joystick, the kind a PC flight simulator enthusiast might use. A laminated sheet is taped to the table in front of the controller, reporting the function of its various buttons. One aims. Another measures the distance from the gun to its target. One loads the bullets into the chamber. Pull the trigger and it will fire.
A gaggle of engineers standing around the table flinch as, unannounced, a warning barks out from a massive, tripod-mounted speaker. A targeting square blinks onto the computer screen, zeroing in on a vehicle that’s moving in the camera’s viewfinder. The gun’s muzzle pans as the red square, like something lifted from futuristic military video game Call of Duty, moves across the screen. The speaker, which must accompany the turret on all of its expeditions, is known as an acoustic hailing robot. Its voice has a range of three kilometres. The sound is delivered with unimaginable precision, issuing a warning to a potential target before they are shot (a warning must precede any firing, according to international law, one of the lab-coat wearing engineers tells me). “Turn back,” it says, in rapid-fire Korean. “Turn back or we will shoot.”
The “we” is important. The Super aEgis II, South Korea’s best-selling automated turret, will not fire without first receiving an OK from a human. The human operator must first enter a password into the computer system to unlock the turret’s firing ability. Then they must give the manual input that permits the turret to shoot. “It wasn’t initially designed this way,” explains Jungsuk Park, a senior research engineer for DoDAAM, the turret’s manufacturer. Park works in the Robotic Surveillance Division of the company, which is based in the Yuseong tech district of Daejon. It employs 150 staff, most of whom, like Park, are also engineers. “Our original version had an auto-firing system,” he explains. “But all of our customers asked for safeguards to be implemented. Technologically it wasn’t a problem for us. But they were concerned the gun might make a mistake.”
As early as 2005 the New York Times reported the Pentagon’s plans to replace soldiers with autonomous robots
The Super aEgis II, first revealed in 2010, is one of a new breed of automated weapon, able to identify, track and destroy a moving target from a great distance, theoretically without human intervention. The machine has proved popular and profitable. DoDAAM claims to have sold more than 30 units since launch, each one as part of integrated defence systems costing more than $40m (£28m) apiece. The turret is currently in active use in numerous locations in the Middle East, including three airbases in the United Arab Emirates (Al Dhafra, Al Safran and Al Minad), the Royal Palace in Abu Dhabi, an armoury in Qatar and numerous other unspecified airports, power plants, pipelines and military airbases elsewhere in the world.
The past 15 years has seen a concerted development of such automated weapons and drones. The US military uses similar semi-autonomous robots designed for bomb disposal and surveillance. In 2000, US Congress ordered that one-third of military ground vehicles and deep-strike aircraft should be replaced by robotic vehicles. Six years later, hundreds of PackBot Tactical Mobile Robots were deployed in Iraq and Afghanistan to open doors in urban combat, lay optical fibre, defuse bombs and perform other hazardous duties that would have otherwise been carried out by humans.
As early as 2005 the New York Times reported the Pentagon’s plans to replace soldiers with autonomous robots. It is easy to understand why. Robots reduce the need for humans in combat and therefore save the lives of soldiers, sailors and pilots. What parent would send their child into a war zone if a robot could do the job instead? But while devices such as the Super aEgis II that are able to kill autonomously have existed for more than a decade, as far as the public knows no fully autonomous gun-carrying robots have been used in active service.
Science fiction writer Isaac Asimov’s First Law of Robotics, that ‘a robot may not injure a human being or, through inaction, allow a human being to come to harm’, looks like it will soon be broken. The call from Human Rights Watch for an outright ban on “the development, production, and use of fully autonomous weapons” seems preposterously unrealistic. Such machines already exist and are being sold on the market – albeit with, as DoDAAM’s Park put it, “self-imposed restrictions” on their capabilities.
“When we started this business we saw an opportunity,” says Yangchan Song, DoDAAM’s managing director of strategy planning, as we sit down in a cooled meeting room following the demonstration. “Automated weapons will be the future. We were right. The evolution has been quick. We’ve already moved from remote control combat devices, to what we are approaching now: smart devices that are able to make their own decisions.”
South Korea has become a leader in this area of military robotics because the country shares a border with its sworn enemy, according to DoDAAM’s CEO, Myung Kwang Chang (a portly man who wanders his factory’s corridors trailed by a handsome husky with bright blue eyes whom I am warned to never, ever touch). “Need is the mother of invention,” he says. “We live in a unique setting. We have a potent and ever-present enemy nearby. Because of this constant threat we have a tradition of developing a strong military and innovative supporting technology in this country. Our weapons don’t sleep, like humans must. They can see in the dark, like humans can’t. Our technology therefore plugs the gaps in human capability.”
Things become more complicated when the machine is placed in a location where friend and foe could potentially mix
At the DMZ, the thin strip of no-man’s land that separates democratic South Korea from the dictator-led North, DoDAAM and its competitor Samsung, who also designed a (now-defunct) automated turret, ran some tests with the Super aEgis II. The DMZ is the ideal location for such a weapon. The zone has separated the two Koreas since the end of official hostilities in 1953; because they never signed a ceasefire, the DMZ is an uninhabited buffer zone scrupulously guarded by thousands of soldiers on both sides. Not only does the turret never sleep and not only can it see in the dark (thanks to its thermal camera), once it's pointing in the right direction, it can be sure that any moving targets identified in the area are enemies. Things become more complicated when the machine is placed in a location where friend and foe could potentially mix, however. Currently, the weapon has no way to distinguish between the two.
Song sits at the wide table flanked by five young engineers, most of whom were educated at Ivy League colleges in America, before returning to work in the lucrative South Korean weapons industry. “The next step for us to get to a place where our software can discern whether a target is friend, foe, civilian or military,” he explains. “Right now humans must identify whether or not a target is an adversary.” Park and the other engineers claim that they are close to eliminating the need for this human intervention. The Super aEgis II is accomplished at finding potential targets within an area. (An operator can even specify a virtual perimeter, so only moving elements within that area are picked out by the gun.) Then, thanks to its numerous cameras, Park says the gun’s software can discern whether or not a potential target is wearing explosives under their shirt. “Within a decade I think we will be able to computationally identify the type of enemy based on their uniform,” he says.
Once a weapon is able to tell friend from foe, and to automatically fire upon the latter, it’s a short step to full automation. And as soon as a weapon can decide who and when to kill, Robocop-esque science fiction becomes fact. The German philosopher Thomas Metzinger has argued that the prospect of increasing the amount of suffering in the world is so morally awful that we should cease building artificially-intelligent robots immediately. But the financial rewards for companies who build these machines are such that Metzinger’s plea is already obsolete. The robots are not coming; they are already here. The question now is, what do we teach them?
Philippa Foot’s trolley dilemma, first posited in 1967, is familiar to any ethics student. She suggested the following scenario: a runaway train car is approaching a fork in the tracks. If it continues undiverted, a work crew of five will be struck and killed. If it steers down the other track, a lone worker will be killed. What do you, the operator, do? This kind of ethical quandary will soon have to be answered not by humans but by our machines. The self-driving car may have to decide whether or not to crash into the car in front, potentially injuring those occupants, or to swerve off the road instead, placing its own passengers in danger. (The development of Google’s cars has been partly motivated by the designer Sebastian Thrun’s experience of losing someone close to him in a car crash. It reportedly led to his belief that there is a moral imperative to build self-driving cars to save lives.)
Likewise, a fully autonomous version of the Predator drone may have to decide whether or not to fire on a house whose occupants include both enemy soldiers and civilians. How do you, as a software engineer, construct a set of rules for such a device to follow in these scenarios? Is it possible to programme a device to think for itself? For many, the simplest solution is to sidestep these questions by simply requiring any automated machine that puts human life in danger to allow a human override. This is the reason that landmines were banned by the Ottawa treaty in 1997. They were, in the most basic way imaginable, autonomous weapons that would explode whoever stepped on them.
In this context the provision of human overrides make sense. It seems obvious, for example, that pilots should have full control over a plane's autopilot system. But the 2015 Germanwings disaster, when co-pilot Andreas Lubitz deliberately crashed the plane into the French Alps, killing all 150 passengers, complicates the matter. Perhaps, in fact, no pilot should be allowed to over-ride a computer – at least, not if it means they are able to fly a plane into a mountainside?
We acquire an intuitive sense of what’s ethically acceptable by watching how others behave and react to situations – Colin Allen
“There are multiple approaches to trying to develop ethical machines, and many challenges,” explains Gary Marcus, cognitive scientist at NYU and CEO and Founder of Geometric Intelligence. “We could try to pre-program everything in advance, but that’s not trivial – how for example do you program in a notion like ‘fairness’ or ‘harm’?” There is another dimension to the problem aside from ambiguous definitions. For example, any set of rules issued to an automated soldier will surely be either too abstract to be properly computable, or too specific to cover all situations.
Some believe the answer, then, is to mimic the way in which human beings build an ethical framework and learn to reflect on different moral rules, making sense of which ones fit together. “We acquire an intuitive sense of what’s ethically acceptable by watching how others behave and react to situations,” says Colin Allen, professor of cognitive science and the philosophy of science at Indiana University, and co-author of the book Moral Machines: Teaching Robots Right From Wrong. “In other words, we learn what is and isn’t acceptable, ethically speaking, from others – with the danger that we may learn bad behaviours when presented with the wrong role models. Either machines will have to have similar learning capacities or they will have to have very tightly constrained spheres of action, remaining bolted to the factory floor, so to speak.”
At DoDAAM, Park has what appears to be a sound compromise. “When we reach the point at which we have a turret that can make fully autonomous decisions by itself, we will ensure that the AI adheres to the relevant army’s manual. We will follow that description and incorporate those rules of engagement into our system.”
For Allen, however, this could be a flawed plan. “Google admits that one of the hardest problems for their programming is how an automated car should behave at a four-way stop sign,” he explains. “In this kind of scenario it’s a matter of being attuned to local norms, rather than following the highway code – which no humans follow strictly.” Surely, in the chaotic context of the battlefield, a robot must be able to think for itself? Likewise, there is a danger to “freezing” our values, both military and civilian, into hardware. “Imagine if the US Founders had frozen their values to permit slavery, the restricted rights of women and so forth,” says Marcus. “Ultimately, we would probably like a machine with a very sound basis to be able to learn for itself, and maybe even exceed our abilities to reason morally.”
For Anders Sandberg, a senior researcher at the Future of Humanity Institute Oxford Martin School, the potential rewards of offering machines the ability to construct their own ethical frameworks comes with considerable risks. “A truly self-learning system could learn different values and theories of what appropriate actions to do, and if it could reflect on itself it might become a real moral agent in the philosophical sense,” he says. “The problem is that it might learn seemingly crazy or alien values even if it starts from common-held human views.”
The clock is ticking on these questions. Companies such as DoDAAM continue to break new ground in the field, even before our species has adequate answers to the issues their work presents. “We should be investing now in trying to figure out how to regulate software, how to enforce those regulations, and how to verify that software does what we want it do,” urges Marcus. “We should also invest in figuring out how to implement ethical reasoning into machines. None of this easy; all of it is likely to became critical in the decades ahead.”
Serious research on machine ethics and AI safety is an exceptionally new field – Anders Sandberg
Allen believes there is still time. “We have an opportunity to think through the ramifications and possible solutions while the technology is still in development, which has not always been the case,” he says. “I would like to see business-government-citizen panels empowered to assess and monitor the deployment of machines that are capable of operating with little or no direct human supervision in public spaces such as roads and the airways. Such panels should provide oversight in much the same way that human subjects committees monitor research involving human subjects.”
Sandberg agrees, although he thinks that we are still some way off having workable answers that can be brought into any kind of regulation. “More research is urgently needed, he says. “Serious research on machine ethics and AI safety is an exceptionally new field - the first ideas started to show up in the late 1990s, decades after proper AI research was already well established. We are discovering new surprises at a high rate: both the instability of obvious means of making ethical machines, and deep philosophical problems posed by seemingly simple engineering problems.” Sandberg believes the key is interdisciplinary work by philosophers, engineers, sociologists and even economists. “Theory is not enough,” he says. “We need to be engaged with the people designing and deploying these systems.”
Back at DoDAAM’s factory in Daejeon, the engineers do not have clear answers for these questions that are so urgently raised by their work. “Human life is more important than anything, so that’s why we’ve implemented the safeguards that we have on our turret,” says Song. But for Park and the other engineers in his research division, this is a temporary state. Their aim is now to make the product “smarter” by focusing on “increasing the gun’s automatic functionality”, introducing “target recognition” and creating “entire networked systems”, whereby turrets link up with one another to cover greater distances.
Regardless of what’s possible in the future, automated machine guns capable of finding, tracking, warning and eliminating human targets, absent of any human interaction already exist in our world. Without clear international regulations, the only thing holding arms makers back from selling such machines appears to be the conscience, not of the engineer or the robot, but of the clients. “If someone came to us wanting a turret that did not have the current safeguards we would, of course, advise them otherwise, and highlight the potential issues,” says Park. “But they will ultimately decide what they want. And we develop to customer specification.”
Follow us on Facebook, Twitter, Google+ and LinkedIn.