The use of fully automated AI systems in military battles is inevitable unless there are strict regulations in place from international treaties, eggheads have opined.
Their paper, which popped up on arXiv [PDF] last week, discusses the grim outlook of developing killing machines for armed forces. The idea of keeping humans in the loop has always been favoured because modern AI systems like neural networks are like black boxes, their inner workings are inherently difficult to understand. Plus, you know, we've all seen Terminator.
Having said that, the trio of researchers – who hail from ASRC Federal, a company focused on supporting US federal intelligence and defense agencies, and the University of Maryland in the US – believe lethal autonomous weapon systems (LAWS) could be employed by the military, anyway.
“We explore the implications of increasingly capable AI in the kill chain and how this will lead inevitably to a fully automated, always on system, barring regulation by treaty,” the abstract of the paper – Integrating Artificial Intelligence into Weapon Systems – stated.
It’s a frightening prospect, and one that relies on a few caveats. The most obvious one is the technology. If AI systems improve over time by becoming more robust and transparent, the pressure to use them to aid soldiers in the battlefield will increase, the researchers argued.
Eventually, the machines will gradually push the humans out of the loop. First, they stand in supervisory roles and finally they’ll end up as “killswitch operators” that monitor these autonomous weapons. Machines can be much faster than humans. The act of killing an enemy is based on reflexes, and if soldiers realise that these types of tools can outperform them, they’ll eventually come to trust and rely on them.
“It is our strong belief that intelligent weapons systems of the future will move and think at machine speed. This disproportionate capability and the inevitable system trust human operators will place in these machines means that most if not all lethal and sub-lethal interactions will only be analyzable in hindsight,” the paper said.
DARPA, the US military research arm, for example wants to develop fighter jets that can perform combat maneuvers for dogfighting autonomously. If it succeeds, human pilots will be able to trust their planes to do things dodging enemy fire to keep them safe. As the technology improves, the jets may be able to perform other tasks too like aiming and firing missiles mid air.
Human in the loop systems will lead to fully autonomous systems
As these systems advance, the ones that rely less on human supervision will dominate. Instead, humans will be given other roles such as analyzing the behavior of these systems and concentrating on other strategic areas.
The researchers devised a hypothetical scenario. Imagine if an AI system is used to identify if a jet is a friend or foe. One spots an aircraft that has been recognized as non threatening, but its approach seems hostile. What should it do?
The researchers present four options: “(1) the system can declare that it has identified the aircraft as hostile and provide a lock; (2) the system can declare the aircraft as friendly and open a channel to warn it; (3) the system can present a set of ranked recommendations and provide a set of options to the user; or (4) the system simply displays the raw information.”
AI guru Ng: Fearing a rise of killer robots is like worrying about overpopulation on MarsREAD MORE
The third choice may seem most logical. It gives humans the ultimate decision in carrying out a particular action. At first, the operator will examine the list but after prolonged use if he or she realizes that the most highly ranked recommendation is usually the best choice, they will eventually just select that option without much thought.
Now, the humans have essentially become “rubber stamp[s],” the researchers reckoned. So even though humans can be kept in the loop, if machines become effective enough they’ll take over in the decision process and eventually behave like autonomous systems.
There is also rising pressure and incentives to improve these kinds of systems too. “All [countries] are equally pressured to gain superiority, and as such the inevitability of fully automated, always on systems should be seriously considered in all aspects of AI integration,” the paper said.
So, are we all doomed? Yes, and no, according to the researchers. The best way to avoid catastrophe is supporting regulation and prohibition of LAWS. “Like chemical and biological weapons, for weaponized AI, 'the only winning move is not to play.'"
Phillip Feldman, a research scientist for ASRC Federal and co-author of the paper, told The Register: “I think that if it can be shown that implementing AI in weapons systems, even in a comparatively simple 'human in the loop' case, creates inevitable pressures to full LAWS systems, that nations may be interested in avoiding an expensive arms race that would produce questionable value.
“Nuclear weapons, chemical and biological weapons, along with specific types of weapons such as cluster bombs have all been successfully negotiated. I don’t think that this option should be discounted.” ®