THE car’s computer saw Elaine Herzberg pushing her bicycle across the highway a full six seconds before it struck her. Travelling at just under 70 kilometres per hour, it had more than enough time to stop or swerve. But it did neither, hitting her head on. Herzberg died in hospital, the first pedestrian to be killed by an autonomous vehicle.
A preliminary investigation by the US National Transport Safety Board into the collision, which happened in Tempe, Arizona, in March, found that the emergency braking procedure of the Uber-operated car was designed to be disabled when driving autonomously to ensure a smoother ride. It was also not designed to alert the human operator of danger.
Advertisement
We are a far cry from Isaac Asimov’s First Law of Robotics: a robot may not injure a human or, through inaction, allow a human to come to harm. Of course, Asimov’s laws are fictional. They are also 75 years old. But the idea of rigid rules that keep wayward robots in check has stuck around in the popular imagination.
Today, we find ourselves in a world fast filling with robots. They are driving our cars, performing our medical procedures, influencing our elections and threatening to take our jobs. It is obvious we can’t live without them, but the deeper question of how we live with them remains unanswered.
The past couple of years has seen a slew of reports from governments and think tanks offering policy recommendations, but so far Asimov’s model of neat self-contained commandments remains a fantasy. How do we change that?
Mady Delvaux, a member of the European Parliament for Luxembourg and author of an , compares the present situation to when cars first started to appear on roads. “The first drivers had no rules,” she says. “They each did what they thought sensible or prudent. But as technology spreads, society needs rules.”
The trouble is, robotic intervention into human affairs is going to require something far more comprehensive than the highway code. Legislation that protects passengers and pedestrians from driverless cars, for example, will be powerless to stop data-scraping algorithms from influencing our vote. Medical robots programmed to diagnose and treat people will need different regulations from those sent onto the battlefield. “AI is not a single thing that can be regulated,” says Steven Croft, the Bishop of Oxford, who sits on the UK House of Lords Select Committee on Artificial Intelligence. “There are different risks, different physical and emotional dangers.”
Code of conduct
There is another hurdle, too. Interpreting rules requires common sense. It’s not clear how you would encode even Asimov’s clear-cut commandments in a way that a machine could be trusted to follow, says Ulrich Furbach at the University of Koblenz in Germany, who works on automated reasoning. He knows because he has tried. “I gave up, it’s much too difficult,” he says.
The trouble is that the laws are too general and don’t allow for context. For Asimov, this was probably part of the point. Many of his stories explore the unintended consequences of robots trying to apply the laws in different scenarios. In Runaround, the 1942 story in which they first appeared in print, a robot gets stuck in a loop when it tries to satisfy the Second Law (obey the orders given to it by human) and Third Law (protect its own existence) at the same time. “Asimov’s laws are perfect for novels but not for putting into practice,” says Delvaux.
Ultimately, the debate boils down to questions of ethics rather than technology. What values do we as a society consider to be worth protecting? Croft stresses the need to avoid the “philosophy of extreme libertarianism” of places like Silicon Valley or the totalitarian approach employed by countries such as China. “The new power that AI brings needs a new kind of ethics,” he says. “It’s going to be vital for the sake of human flourishing that this is shaped by society rather than shaping society. It’s in the interest of individuals not to allow private companies to set the agenda.”
As Croft’s remarks highlight, different places have different cultures and different cultures breed different laws. Japan is much more comfortable with the idea of companion robots than many other countries, for example. And as shown by the recent fuss over GDPR, Europe has different attitudes towards privacy and data collection than the US.
Should robots have rights?
Nonetheless, there are surely some guidelines that everyone might be willing to sign on to. Not allowing robots to harm humans seems like a no-brainer. A UN ban on autonomous weapons has widespread support, for example. But some argue that letting robots kill might actually save lives, by reducing the number of lethal mistakes made by human soldiers.
A similar argument is used to justify the occasional fatal accident with driverless cars and medical robots: we cannot cut the risk of harm to zero, but humans will be far safer with machines than we are with one another.
Many find such utilitarian reasoning unpalatable, however. “The taking of life in war can only be justified in the most extreme circumstances,” says Croft. “Human judgement is therefore a very important part of the regrettable waging of war. Anything that moves away from that is ultimately dehumanising.”
“AI will need a new ethics – one shaped by society rather than shaping it”
Wherever robots have the capacity to inflict harm, therefore – in the operating theatre, on the battlefield or on the road – human oversight is essential.
One scenario often invoked to show the dangers of letting AI decide for itself is the so-called trolley problem, in which a driverless car must choose who to kill in the event of an accident that is certain to be fatal to either the car’s occupants or pedestrians, depending on its action. But Ryan Calo at the University of Washington in Seattle finds the thought experiment ridiculous. It is an odd assumption to make that a robot has no choice but to kill, says Furbach. “Nobody asked me about the trolley problem when I took my driving test.”
More interesting are the new possibilities that technology brings about. Here is one scenario suggested by Calo. Imagine an autonomous car with a hybrid motor that can charge its battery by running its petrol engine. One of this car’s goals might be to maximise its fuel efficiency. After a few days on the road it learns that it is most efficient if it starts out each morning with a fully charged battery and decides to run its engine in the garage overnight, killing everyone in the household through carbon monoxide poisoning.
This would be a clear violation of any do-no-harm-to-humans law but one that was hard to foresee. In such cases, how do we work out what went wrong and who is to blame?
“There will be accidents,” says Rebecca Crootof at Yale University. “The question is, will we let the harm fall where it lands, or will we figure out how to hold appropriate entities accountable and ensure that victims are compensated?”
Delvaux believes that responsibility for robots lies with the manufacturers. “If you put something on the market you should be liable for it,” she says. Croft agrees: “Responsibility extends to the things you create,” he says.
To manage liability, Delvaux suggests that robots should be given an e-personality that would act as a handle for the group of people behind it, in much the same way as companies are treated as individuals under the law to ensure collective responsibility for their actions. “A corporation has humans behind it,” she says. “It’s the same for a robot.”
The trouble is that liability law typically requires defendants to have foreseen the outcomes of their actions. But with artificial intelligence software that learns, this may not be reasonable – as in Calo’s battery-charging scenario. To avoid such unpredictable behaviour from ever arising, one German commission has suggested that there should be no self-learning components in driverless cars, says Furbach, “Which is silly. You want it to be able to improve.”
What we can ask for instead is that a robot is able to explain itself. AI is starting to be used to predict our fitness for a job or the likelihood that we will commit a crime, for example – decisions that we should have a right to question. The AI Now Institute, a cross-disciplinary research centre at New York University, recommends that organisations involved in high-stakes domains like welfare, criminal justice, healthcare and education should no longer use black-box systems whose algorithms cannot be scrutinised.
Ensuring the responsible use of algorithms may mean we need new hybrid approaches to AI, where machine learning software is combined with techniques that are more easy for humans to understand. This is an exciting new research area, says Furbach.
For Anne-Marie Imafidon, a trustee at the Institute for the Future of Work in Palo Alto, California, if we want AI to provide maximum benefit with the least harm, the key is not just to make machines understandable to humans, but to make them better than we ever could be. Her answer is to make sure machines don’t share our obsession with categorising things. That way we might avoid some of the bias that creeps into our AI and skews automated decision-making.
When we train software, we typically ask it to label things – is this a dog or a cat? But once you lump something into a category – black male age 18-25, woman over 65 – a lot of assumptions come into play that may not be relevant to the task at hand.
Imafidon cites Netflix as an example of categorisation done well. If you go to the cinema to see a film soon after it is released, you sometimes get asked what you thought of it. Your response is then dumped in with those of your demographic: women of this age like this film, for example. But Netflix tailors its recommendations to individuals: people who liked this film also liked that one. Your gender and age have nothing to do with it.
Getting machines to treat people more as individuals is the most important principle we should aim for, says Imafidon. Yet such laws could easily infringe on our privacy, watching our every move and learning our every preference before using that data to make a profile that gets used against us. AI-assisted profiling is already exacerbating big problems in society, such as Russian agents using Facebook’s algorithms to seed the social network with tailor-made propaganda. Perhaps we would see more of the advantages if we had some guarantee that the data was used in an unbiased way.
If we can overcome these challenges, machines that save lives, account for their actions and treat us all fairly would be welcomed into most societies. But as we become capable of creating AI and robots that are more and more human-like, we must consider the kind of interactions we are comfortable having with the machines.
Voice-recognition technology lets us talk to our devices more or less naturally. Google’s Duplex software lets AI-controlled virtual assistants book appointments over the phone using a synthetic voice that is realistic enough to make a person on the other end of the line think they are talking to another human. And roboticists are crafting humanoid robots with lifelike flesh and hair.
Uncanny, certainly. But there is a bigger concern here. Humans form emotional attachments easily. We have barely begun to consider what types of emotional relationships people might have with human-like companion robots, for example. Cultural differences may come into play. “In Japan, robots caring for the elderly is completely acceptable,” says Furbach. “In Germany, it’s considered inhuman.”
“We must be careful it isn’t only employers that benefit from robots”
But we can probably all agree that however lifelike a machine appears to be, we should always be able to tell whether we are in fact interacting with a machine or human. “We can’t prevent manufacturers from putting human-like robots on the market because consumers will buy them,” says Delvaux. “But we should make people aware if they interact with a robot that it is not a human.”
In other settings, even if robots don’t look particularly like us, they might still be able to take our place. “Robots could be a liberating force by taking away routine work,” says Tom Watson, deputy leader of the UK’s Labour Party who set up the Future of Work Commission last year to look at how new technologies will affect employment. He is, however, concerned about the imbalance in power between those calling the robot shots and the rest of us. “We’ve got to be careful that big corporations and employers don’t amass all the benefits while ordinary workers are left to lump the negatives,” he says.
In our increasingly automated world, it is easy to forget that machines are programmed, owned and operated by humans. Like any other invention, their decisions, purpose and function are all designed with some higher, all-too-human goal in mind: safety, comfort, efficiency, profit. Designing laws for machines to follow is an entertaining thought experiment, but ultimately a distraction. The real robot laws need to be written to keep people in check, not machines.
“This is a debate for everyone in the world,” says Croft. “As a human race, we need to access deep traditions of wisdom to address these fundamental questions. We cannot answer them through technology alone.”
There is at least one technological fix we might all agree on, however. “A human should always be able to shut down a machine,” says Delvaux.
Five commandments
Any universal laws of robotics will probably include the following
Law #1: A robot may not injure a human or allow a human to come to harm – unless it is being supervised by another human.
Law #2: A robot must be able to explain itself.
Law #3: AI should resist the urge to pigeon-hole.
Law #4: A robot must not impersonate a human.
Law #5: A robot should always have an off switch.
- Hear roboticist Patricia Vargas discuss whether robots should have rights at 91av Live, running in London from 20-23 September. For more information see
This article appeared in print under the headline “Robot laws”
Article amended on 3 August 2018
We clarified what was disabled in the Uber-operated car
