In November 2012, United States Deputy Defense Secretary Ashton Carter signed directive 3000.09, establishing policy for the “design, development, acquisition, testing, fielding, and … application of lethal or non-lethal, kinetic or non-kinetic, force by autonomous or semi-autonomous weapon systems.” Without fanfare, the world had its first openly declared national policy for killer robots.
The policy has been widely misperceived as one of caution. According to one account, the directive promises that a human will always decide when a robot kills another human. Others even read it as imposing a 10-year moratorium to allow for discussion of ethics and safeguards. However, as a Defense Department spokesman confirmed for me, the 10-year expiration date is routine for such directives, and the policy itself is “not a moratorium on anything.”
A careful reading of the directive finds that it lists some broad and imprecise criteria and requires senior officials to certify that these criteria have been met if systems are intended to target and kill people by machine decision alone. But it fully supports developing, testing, and using the technology, without delay. Far from applying the brakes, the policy in effect overrides longstanding resistance within the military, establishes a framework for managing legal, ethical, and technical concerns, and signals to developers and vendors that the Pentagon is serious about autonomous weapons.
Did soldiers ask for killer robots? In the years before this new policy was announced, spokesmen routinely denied that the US military would even consider lethal autonomy for machines. Over the past year, speaking for themselves, some retired and even active duty officers have written passionately against both autonomous weapons and the overuse of remotely operated drones. In May 2013, the first nationwide poll ever taken on this topic found that Americans opposed to autonomous weapons outnumbered supporters by two to one. Strikingly, the closer people were to the military—family, former military, or active duty—the more likely they were to strongly oppose autonomous weapons and support efforts to ban them.
Since the 1990s, the military has exhibited what autonomy proponent Barry Watts has called “a cultural disinclination to turn attack decisions over to software algorithms.” Legacy weapons such as land and sea mines have been deemphasized and some futuristic programs canceled—or altered to provide greater capabilities for human control. Most notably, the Army’s Future Combat Systems program, which was to include a variety of networked drones and robots at an eventual cost estimated as high as $300 billion, was cancelled in 2009, with $16 billion already spent.
At the same time, calls for autonomous weapons have been rising both outside and from some inside the military. In 2001, retired Army lieutenant colonel T. K. Adams argued that humans were becoming the most vulnerable, burdensome, and performance-limiting components of manned systems. Communications links for remote operation would be vulnerable to disruption, and full autonomy would be needed as a fallback. Furthermore, warfare would become too fast and too complex for humans to direct. Realistic or not, such thinking, together with budget pressures and the perception that robots are cheaper than people, has supported a steady growth of autonomy research and development in military and contractor-supported labs. In March 2012, the Naval Research Lab opened a new facility dedicated to development and testing of autonomous systems, complete with simulated rainforest, desert, littoral, and shipboard or urban combat environments. But the killer roboticists’ brainchildren have continued to face what a 2012 Defense Science Board report, commissioned by then-Undersecretary Carter, called “material obstacles within the Department that are inhibiting the broad acceptance of autonomy.”
The discrimination problem. Navy scientist John Canning recounts a 2003 meeting at which high-level lawyers from the Navy and Pentagon objected to autonomous weapons. They assumed that robots could not comply with international humanitarian law, core principles of which include a responsibility to distinguish civilians from combatants and to refrain from attacks that would cause excessive harm to civilians. These principles, and the military rules of engagement intended to implement them, assume a level of awareness, understanding, and judgment that computers simply don’t have. Weapons are also subject to mandated legal review, and indiscriminate weapons—that is, weapons that cannot be selectively directed to attack lawful targets and avoid civilians—are forbidden. The lawyers did not think they would ever be able to sign off on autonomous weapons.
Georgia Tech roboticist Ron Arkin has argued that unemotional robots, following rigid programs, could actually be more ethical than human soldiers. But his proposals fail to solve the hard problems of distinguishing civilians, understanding and predicting social and tactical situations, or judging the proportionality of force. Others argue, philosophically, that only humans can make such targeting judgments legitimately. In a world getting used to talking about virtual assistants and self-driving cars, it may not be obvious what the limits of artificial intelligence will be, or what people will accept, in 10, 20, or 40 years. But for now, and for the immediate future, the robot discrimination problem is hard to dispute.
To break the legal deadlock, Canning suggested that robots might normally be granted autonomy to attack materiel, including other robots, but not humans. Yet in many situations it might be impossible to avoid the risk—or the intent—of killing or injuring people. For such cases, Canning proposed what he called “dial-a-level” autonomy; that is, the robot might ordinarily be required to ask a human what to do, but in some circumstances it could be authorized to take action on its own.
In recent years, autonomy visionaries have stressed human-machine partnerships and flexibility to decide the level of autonomy a weapon may be allowed, based on tactical needs. In a 2011 roadmap, for example, the Defense Department envisions unmanned systems that seamlessly operate with manned systems while gradually reducing the degree of human control and decision making required. A 2011 Navy presentation depicts decisions about autonomy and control as a continuous tradeoff, explaining that while human control minimizes the risk of attacking unintended targets, machine autonomy maximizes the chance of defeating the intended ones. It seems likely that in desperate combat, autonomy would be dialed up to the highest level.
Appropriate levels of human judgment. In the spring of 2011, the Defense Department convened a group of uniformed and civilian personnel to begin developing a policy for autonomous weapons. The directive that emerged 18 months later lists a number of requirements for autonomous systems and draws a line at systems intended to autonomously target and engage humans–or to apply kinetic force (e.g., bullets and bombs) against any targets. But the directive neither states nor implies that this line should not be crossed.
Rather, the line may be crossed if two undersecretaries and the Chairman of the Joint Chiefs of Staff affirm that the listed requirements have been met. In the event of an urgent military need, any of the requirements can be waived—with the exception of a legal review. Furthermore, the line is not as clearly drawn as it may seem to be.
The requirements listed in the directive are not much more stringent than those that apply to any weapon system. Tactics, techniques, and procedures must be developed to specify how an autonomous weapon system should be used. Hardware and software must undergo rigorous verification and validation. Human-machine interfaces must be understandable to trained operators and must provide clear activation and deactivation procedures and have “safeties, anti-tamper mechanisms, and information assurance” that minimize the probability of unintended engagements.
These requirements sound reassuring; they promise to address many of the concerns people have about autonomous weapons. According to the directive, it is Defense Department policy that the measures listed will ensure that the systems will work in realistic environments against adaptive adversaries. But saying it doesn’t necessarily make it so.
In reality, neither mathematical analysis nor field testing can possibly locate every software bug or situation in which such complex systems may fail or behave inappropriately. Adversaries will strive to locate points of vulnerability, and it is terribly hard to anticipate everything that adversaries may do, let alone know how their actions may affect system performance. The notion of information assurance implies a promise to solve problems of software reliability and computer security that bedevil contemporary technology.
The centerpiece of the entire directive is this statement: “Autonomous and semi-autonomous weapon systems shall be designed to allow commanders and operators to exercise appropriate levels of human judgment over the use of force.” Although the phrase is never defined, it does not appear that appropriate levels always require at least one human being to make the decision to kill another. Rather, the appropriate level might well be the decision to dispatch a robot on a mission and let it select the targets to engage. In making such decisions, it appears that the burden of ensuring compliance with rules of engagement and laws of war falls on commanders and operators when the robots themselves are incapable of ensuring this. But in practice, it seems likely that unintended atrocities committed by autonomous weapons will be blamed on technical failures.
Semi-autonomy: Smudging the line. In theory, as long as three senior officials withhold their signatures, autonomous weapon systems that are intended to target humans or use kinetic or lethal force would be blocked. But the policy green lights—no extra signatures needed—semi-autonomous weapon systems that may apply any kind of force against any targets, including people. The crucial line that the policy draws between semi- and fully autonomous systems is fuzzy and broken. As technology advances, it is likely to be crossed as a matter of course.
The directive defines a semi-autonomous weapon system as one intended to engage only those targets that have been selected by a human operator. But the system itself is allowed to use autonomy to acquire, track and identify potential targets. It can cue the operator, prioritize targets, and decide when to fire. What the operator must do to select targets is left unspecified. Would a verbal OK, gesture, or even neurological interface be acceptable?
A system with such capabilities may not be intended to function without a human operator, but at most it would require a trivial modification to do so—perhaps a hack. At least three companies already market such systems. The policy clears them for immediate use after acquisition, via standard procedures.
The policy also addresses fire-and-forget or lock-on-after-launch homing munitions, which would include many systems in use today. Such munitions have seekers that autonomously find and home on targets. The directive classifies them as semi-autonomous weapon systems, on the theory that the operator selects targets by using tactics, techniques and procedures that “maximize the probability that the only targets within the seeker’s acquisition basket” will be the intended targets. Yet, upon launch, such munitions become, de facto, fully autonomous.
No restrictions are placed on the technology that a seeker may use to find a target and decide whether that is what it was looking for. This opens a clear path for weapons that can be sent on hunt-and-kill missions, limited only by the ability of their onboard sensors and computers to narrow their acquisition baskets to selected targets.
The way forward—to what? In the mid-2000s, Lockheed Martin was developing a small autonomous drone missile for the Air Force, and a similar system for the Army. Equipped with several types of onboard sensors, the missiles would fly out to designated areas and wander in search of generic targets, such as tanks, rocket launchers, radars, or personnel, which they would autonomously recognize and attack. Both programs were canceled, amid legal, ethical, and technical questions, to be superseded by systems that combine autonomous capabilities with radio links to human operators. Under the new policy, would such wide-area search munitions be classified as autonomous or semi-autonomous? Either way, the policy establishes that weapons like these may be developed, acquired, and used.
Given the long internal debate and general public opposition to killer robots, this is a highly aggressive policy. The US military never intended to replace foot soldiers with autonomous lethal robots during this decade, particularly not where civilians might be at risk. But funding the development and acquisition of systems that have autonomous targeting and fire-control capabilities—even if they are not intended for fully autonomous killing—will spur the weapons industry, in the United States and elsewhere, to accelerate exploration and investment in the technology of autonomous warfare.
The real issue is whether the world needs to go this way at all. The message of this policy is: full speed ahead.