Those in favour of a ban would look to stop a potential arms race in ‘killer robots’ before it begins.
It’s the theme of so many dystopian sci-fi books and movies: a super intelligent machine in charge of lethal military hardware becomes self-aware and decides to wreak havoc. But could it actually happen?
At the Association for the Advancement of Artificial Intelligence’s annual conference in Texas last month, a workshop was held on the ethics of AI development and a panel discussed whether or not so-called ‘lethal autonomous weapons’ should be banned.
“There are many arguments, legal, ethical, political and technical for a ban,” Toby Walsh, head of the Optimisation Research Group at Australia’s research body NICTA and chair of the proceedings, told Fairfax Media.
“One that particularly appeals to me is that [autonomous weapons] will lower the barrier to war. If one side can launch an attack, without fear of bodies coming home, then it is much easier to slip into battle,” Professor Walsh said.
While the advent of drones might already have the bar falling, those in favour of a ban would look to stop a potential arms race in “killer robots” before it begins. One the other hand, there are plenty of voices against a ban.
“Machines are not inherently dangerous”, said Francesca Rossi, president of the International Joint Conference on Artificial Intelligence and a participant in the AAAI panel, who points out the huge difference between super-intelligence and sentience.
“We should build [autonomous weapons] by specifying all the relevant context for the desired goal to be achieved by the machine. Otherwise, a goal could be reached by violating some basic assumptions on how we want a machine to behave. Since machines are not sentient, their behaviour depends on how a human built them,” Professor Rossi said.
The specification of “all” relevant context could prove a troublesome task, however, since as Professor Walsh points out, “many ethical principles that we hold as universal are not”, and ethics and decision-making processes across different cultures vary.
Many pro-development experts argue automated fighters would be much more principled in applying the laws and guidelines of war; that a ban would be impossible to police and that our enemies will surely develop autonomous weapons even if we do not.
“There’s an argument that ‘fire and forget’ weapon systems already in the field are lethal autonomous weapons”, Professor Walsh said.
“If a ban is put in place, the diplomats tell me that they’ll worry about defining precisely what is a ‘lethal autonomous weapon’ right before the treaty is signed”.
Hollywood view of AI
Microsoft’s Eric Horvitz, who recently predicted humanity would never lose control of AI contradicting Stephen Hawking, told Fairfax Media Hollywood has given many people an inflated idea of how close we are to developing a truly sentient system.
“We’ve made advances on very, very narrow wedges of intelligence, such as building predictive models from data, performing face recognition, and understanding speech”, he said.
“Speculative essays, books, and a chain of scary movies have overcharged expectations and fears, and thus, have stimulated concerns about AI. My sense is that we are on a very long, slow research trajectory and that we will have plenty of time to study and reflect about potential opportunities and challenges along the way.”
Horvitz has conceived of a 100-year effort, currently being undertaken at Stanford University, to study and anticipate how AI development will ripple through human life.
Already, he said, the positive benefits of AI development can be seen in healthcare, education, transportation, and science. He expects major scientific breakthroughs will soon be attributed to machine scientists, and hundreds of AI vehicles will make up a “public microtransit network” in CBDs that are closed off to human-driven vehicles.
Professor Walsh is also sceptical about Hollywood-style AI systems become reality soon.
“If you ask most AI researchers today, they’ll say 30 or 40 years. But then they’ve been saying this for the last 30 or 40 years”, he said.
“What’s new is that there is now a real scent of success, and people are recognising it could actually be in their lifetime”.
Lethal weapons or no, the application of ethics to AI is still vitally important, Professor Walsh said.
“If AI systems are given autonomy — and there are good economical, social and other reasons why this is sure to be the case — then they will inevitably have to make life and death decisions,” he says.
“An autonomous car will have to decide between crashing into the person crossing the road with a pram, surely killing the person and the baby but probably not the occupant of the car, or into a wall, surely killing the occupant of the car. These are not easy decisions to make and require ethics to make well.”
While creativity and the ability to think laterally are often cited as reasons AI will never be able to do certain human jobs (for example interior designer, cartoonist, comedian), Professor Walsh disagreed.
“There’s actually of field of AI looking at how to make computers creative. And there are a number of good examples where you can argue that AI programs have already been creative”, he says, citing programs that have painted original artworks and found unexpected proofs to scientific theorems.
“We play an interesting dinner time game at our conferences: one person has to name a job, and the other has to provide evidence or an argument why it’s only a matter of time before this job could be automated. Doctors: IBM’s Watson is already working on that. Reality TV contestants: wouldn’t it be far more interesting to have AI systems pretending to be Marilyn Monroe and Albert Einstein in a virtual Big Brother house?”