A third of Britons fear the rise of the robots, according to a poll conducted for the launch of new science fiction drama Almost Human. So, given that more than a quarter of participants believe that robots will in future be capable of feeling human emotion, should we be trying to teach them to behave ethically?
Start talking about robots with ethics, and most science fiction buffs’ minds turn to SF writer Isaac Asimov’s Three Laws of Robotics. According to these rules, a robot may not injure a human being or, through inaction, allow one to come to harm; it must obey the orders given to it by human beings, except where such orders would conflict with the First Law; and must protect its own existence, as long as such protection does not conflict with the First or Second Law.
Many of Asimov’s stories depend for their plot on how robots apply these laws in practice, resulting in everything from a human-robot love affair to the total subjugation of humanity. How does a robot know what constitutes a threat to itself? And, more importantly, how does it know what constitutes harm to a human?
With robots increasingly working alongside human beings, such questions are no longer the preserve of science fiction alone. In Japan in particular, robots are being trialled (under supervision) as companions for the elderly, as medical porters and even as primary school teachers – all functions where ethical behaviour is crucial.
But as AJung Moon of the University of British Columbia, points out, “It’s really hard to create a robot that would have the same sense of moral agency as a human being. Part of the reason is that people can’t even agree on what is the right thing to do. What would be the benchmark?”
Her latest research, led by colleague Ergun Calisgan, takes a pragmatic approach to the problem by examining a robot tasked with delivering a package in a building with only one small lift. How should it act? Should it push ahead of a waiting human? What if its task is urgent? What if the person waiting is in a wheelchair?
“We wanted to get away from figuring out what is the right thing to do, but figure out a practical solution to getting a robot to behave in a way that people think is the right thing to do,” says Moon. Interviews with human participants showed that when the robot’s errand wasn’t urgent, people felt that it should give precedence to the human being, saying: “Go ahead. I will ride the next one.”
When the mail was urgent, there was less consensus. The least-popular options were for the robot to just stand there – probably looking a bit sinister – or to order people out of the way. The most popular, especially when the waiting person was in a wheelchair, was for the robot to explain that it was on an urgent mission and to ask if the human was in a hurry. With all this information plugged into a learning algorithm, the robot can now work out the most appropriate behaviour and avoid getting into a fight.
Nowhere is the problem of machine ethics more pressing than in the field of warfare. Currently, robots are used widely to gather data for surveillance and for patrols. But research into fully autonomous ‘killer robots’ is already underway in the US, China, Russia, Israel and the UK, and several semi-autonomous weapons already exist.
Ironically, it should in some ways be easier to create robots that can behave ethically on the battlefield than to build considerate care workers. While compassion can be hard to pin down, the ethical rules for battle are defined through international agreements – perfect for a robot, you might think.
Indeed, professor Ronald Craig Arkin of the Georgia Institute of Technology has proposed an “ethical adaptor”, designed give a military robot what he describes as a sense of guilt. It racks up, according to a pre-determined formula, as the robot perceives after an event that it has violated the rules of engagement – perhaps by killing a civilian in error – or if it is criticised by its own side. Once its guilt reaches a certain pre-determined level, the robot is denied permission to fire.
But, says professor Noel Sharkey of the University of Sheffield, using such terms is highly misleading. “Using the word ‘guilt’ is what I call a Trojan horse,” he says. “When I say a person feels guilty, it opens up a whole world of ideas, but when you use that word as a roboticist it means something very different.”
In any case, as we’ve seen, it takes a great deal of work just to prime a robot to rescue people from falling down holes. In the case of battlefield robots, the practical issues are immense.
“It involves very, very difficult issues in sensing technology – being able to recognise combatants reliably, remembering that not every person carrying a gun is a combatant, remembering that there might be neutral parties on the battlefield,” professor Rob Sparrow of Monash University pointed out in a recent debate of the TechDebate on Emerging Technologies series. “It involves recognising civilians carrying golf clubs so that it might look as if they are carrying a weapon. The complexity of the task is really unimaginably high.”
Professor Alan Winfield of the University of the West of England Bristol is working on a robot that. in one situation at least, abides by Asimov’s full First Law. “This is certainly not Kantian ethics, or even Aristotelian ethics, but about a robot being able to make a decision to override its normal safety rules to save a human,” he says.
In his tests, a robot uses its sensors to detect that a “human being” is moving towards a “hole in the ground” (actually, another robot and a designated area of floor).
“What the robot then does is something it would not normally do,” he says. “It would normally be designed to avoid collisions with humans; but on this occasion the robot would decide that a gentle collision with this particular human is more acceptable than letting the person fall into the hole.”
It’s a simple enough scenario – but even here, there’s an enormous amount of work involved. The robot needs to have internal models of both itself and its environment, and to be capable of working out the effects of its own and others’ actions.
“When you can model the consequences of actions, you can select actions that will minimise the harmful consequences. But it all needs to happen in real time,” says Professor Winfield.
“And even a robot with harm reduction ethics can face an ethical dilemma – say, when there are two humans about to fall into a hole and only one robot.”
Without exception, roboticists say science fiction has done them no favours, making us believe robots are far more capable than they actually are. Our natural anthropomorphisation compounds this: one US colonel called off tests of a bomb-disposal robot, for example, because he believed blowing its legs off was “inhumane”. It’s all too easy to attribute thoughts and feelings to machines that possess nothing of the sort.
“I can’t see ethical robots happening in my lifetime,” says Professor Sharkey. “There might be some sort of big breakthough, I suppose, but otherwise I don’t expect to ever see any great advance towards robots as moral agents doing ethical things.”