Taming the monsters of tomorrow

A small group of researchers is studying how science could destroy the world—and how to stop that from happening

A nuclear bomb is detonated at the U.S. Pacific Proving Grounds in 1958. Global nuclear warfare is one of the “existential risks” that could end civilization as we know it.

OMIKRON/SCIENCE SOURCE
CRAIG & KARL

Philosopher Nick Bostrom believes it's entirely possible that artificial intelligence (AI) could lead to the extinction of Homo sapiens. In his 2014 bestseller Superintelligence: Paths, Dangers, Strategies, Bostrom paints a dark scenario in which researchers create a machine capable of steadily improving itself. At some point, it learns to make money from online transactions and begins purchasing goods and services in the real world. Using mail-ordered DNA, it builds simple nanosystems that in turn create more complex systems, giving it ever more power to shape the world.

Now suppose the AI suspects that humans might interfere with its plans, writes Bostrom, who's at the University of Oxford in the United Kingdom. It could decide to build tiny weapons and distribute them around the world covertly. "At a pre-set time, nanofactories producing nerve gas or target-seeking mosquito-like robots might then burgeon forth simultaneously from every square meter of the globe."

For Bostrom and a number of other scientists and philosophers, such scenarios are more than science fiction. They're studying which technological advances pose "existential risks" that could wipe out humanity or at least end civilization as we know it—and what could be done to stop them. "Think of what we're trying to do as providing a scientific red team for the things that could threaten our species," says philosopher Huw Price, who heads the Centre for the Study of Existential Risk (CSER) here at the University of Cambridge.

The idea of science eliminating the human race can be traced all the way back to Frankenstein. In Mary Shelley's novel, the monster gets angry at his creator, Victor Frankenstein, for having spurned him. He kills Frankenstein's little brother William, but then offers the doctor a deal: Make a female companion for me and we will leave you in peace and go to South America to live out our days. Frankenstein starts working on the bride, but realizes that the couple might reproduce and outcompete humans: "A race of devils would be propagated upon the earth who might make the very existence of the species of man a condition precarious and full of terror." He destroys the half-finished female, reigniting the creature's wrath and bringing about his own demise.

"I think Frankenstein illustrates the point beautifully," says physicist Max Tegmark of the Massachusetts Institute of Technology (MIT) in Cambridge, a board member of CSER and a co-founder of a similar think tank, the Future of Life Institute (FLI), near MIT. "We humans gradually develop ever-more-powerful technology, and the more powerful the tech becomes, the more careful we have to be, so we don't screw up with it."

The study of existential risks is still a tiny field, with at most a few dozen people at three centers. Not everyone is convinced it's a serious academic discipline. Most civilization-ending scenarios—which include humanmade pathogens, armies of nanobots, or even the idea that our world is a simulation that might be switched off—are wildly unlikely, says Joyce Tait, who studies regulatory issues in the life sciences at the Innogen Institute in Edinburgh. The only true existential threat, she says, is a familiar one: a global nuclear war. Otherwise, "There is nothing on the horizon."

Harvard University psychologist Steven Pinker calls existential risks a "useless category" and warns that "Frankensteinian fantasies" could distract from real, solvable threats such as climate change and nuclear war. "Sowing fear about hypothetical disasters, far from safeguarding the future of humanity, can endanger it," he writes in his upcoming book Enlightenment Now: The Case for Reason, Science, Humanism, and Progress.

But advocates predict the field will only get more important as scientific and technological progress accelerates. As Bostrom pointed out in one paper, much more research has been done on dung beetles or Star Trek than on the risks of human extinction. "There is a very good case for saying that science has basically ignored" the issue, Price says.

Humanity has always faced the possibility of an untimely end. Another asteroid of the size that ended the dinosaurs' reign could hit Earth; a volcanic cataclysm could darken the skies for years and starve us all.

But existential risks arising from scientific advances were literally fiction until 16 July 1945, when the first atomic bomb was detonated. Based on some back-of-the-envelope calculations, physicist Edward Teller had concluded that the explosion might set off a global chain reaction, "igniting" the atmosphere. "Although we now know that such an outcome was physically impossible, it qualifies as an existential risk that was present at the time," Bostrom writes. Within 2 decades a real existential risk emerged, from growing stockpiles of the new weapons. Physicists had finally assembled Frankenstein's bride.

Other scientific disciplines may soon pose similar threats. "In this century we will introduce entirely new kinds of phenomena, give ourselves new kinds of powers to reshape the world," Bostrom says. Biotechnology is cheaper and easier to handle than nuclear technology has ever been. Nanotechnology is making rapid strides. And at a 2011 meeting in Copenhagen, Estonian computer programmer and Skype co-developer Jaan Tallinn told Price about his deep fears about AI during a shared taxi ride. "I'd never met anyone at that point who took that as seriously as Jaan," says Price, who was about to start working at the University of Cambridge.

Price introduced Tallinn to astronomer Martin Rees, a former president of the Royal Society, who had long warned that as science progresses, it will increasingly place the power to destroy civilization in the hands of individuals. The trio decided to launch CSER, the second such center after Bostrom's Future of Humanity Institute in Oxford, which he launched in 2005. CSER's name was "a deliberate attempt to push the idea of existential risk more towards the mainstream," Price says. "We were aware that people think of these issues as a little bit flaky."

CSER has recruited some big-name supporters: The scientific advisory board includes physicist Stephen Hawking, Harvard biologist George Church, global health leader Peter Piot, and tech entrepreneur Elon Musk. In a sign of just how small the field still is, Tallinn also co-founded FLI in 2014, and Church, Musk, Hawking, Bostrom, and Rees all serve on its scientific advisory board. (Actor Morgan Freeman, who has literally played God, is also an FLI adviser.)

Most of CSER's money comes from foundations and individuals, including Tallinn, who donated about $8 million to existential risk researchers in 2017. CSER's academic output has been "ephemeral" so far, Tallinn concedes. But the center was set up as "a sort of training ground for existential risk research," he says, with academics from elsewhere coming to visit and then "infecting" their own institutions with ideas.

The dozen people working at CSER itself—little more than a large room in an out-of-the-way building near the university's occupational health service—organize talks, convene scientists to discuss future developments, and publish on topics from regulation of synthetic biology to ecological tipping points. A lot of their time is spent pondering end-of-the-world scenarios and potential safeguards.

The release of a dangerous pathogen might cause a “crunch” in the human population.

MIKE AGLIOLO/SCIENCE SOURCE

Church says a "crunch," in which a large part of the world population dies, is more likely than a complete wipe-out. "You don't have to turn the entire planet into atoms," he says. Disrupting electrical grids and other services on a huge scale or releasing a deadly pathogen could create chaos, topple governments, and send humanity into a downward spiral. "You end up with a medieval level of culture," Church says. "To me that is the end of humanity."

Existential risks stemming from the life sciences are perhaps easiest to imagine. Pathogens have proved capable of killing off entire species, such as the frogs that have fallen victim to the amphibian fungus Batrachochytrium dendrobatidis. And four influenza pandemics have swept the world in the past century, including one that killed up to 50 million people in 1918 and 1919. Researchers are already engineering pathogens that in principle could be even more dangerous. Worries about studies that made the H5N1 bird flu strain more easily transmissible between mammals led the United States to halt such research until late last year. Terrorists or rogue states could use labmade agents as a weapon, or an engineered plague could be released accidentally.

Rees has publicly wagered that by 2020, "bioterror or bioerror will lead to 1 million casualties in a single event." Harvard microbiologist Marc Lipsitch has calculated that the likelihood of a labmade flu virus leading to an accidental pandemic is between one in 1000 and one in 10,000 per year of research in one laboratory; Ron Fouchier of Erasmus MC in Rotterdam, the Netherlands, one of the researchers involved in the H5N1 studies, has dismissed that estimate, saying the real risk is more like one in 33 billion per year and lab.

One measure against "bioerror" might be to make researchers who carry out risky experiments buy insurance; that would require an independent assessment of the risk and would force researchers to face up to it, Lipsitch says. Still, the most important countermeasure is to strengthen the world's capacity to contain an outbreak early on, he adds, for instance with vaccines. "For biological risks, short of a really massive, coordinated, parallel attack around the world, the only way we are going to get to a really catastrophic scenario is by failing to control a smaller scenario," he says.

Viruses are unlikely to kill every last human, Bostrom says; for him and others, it is AI that poses truly existential threats. Most scenarios center on machines out-smarting humans, a feat called "super-intelligence." If such AI were ever achieved and it acquired a will of its own, it might turn malevolent and actively seek to destroy humans, like HAL, the computer that goes rogue aboard a spaceship in Stanley Kubrick's film 2001: A Space Odyssey.

Humanity's strategy is to learn from mistakes. When the end of the world is at stake, that is a terrible strategy.

Max Tegmark, Massachusetts Institute of Technology

Most AI experts worry less about machines rising up to overthrow their creators, however, than about them making a fatal mistake. To Tallinn, the most plausible way in which AI could end humanity is if it simply pursued its goals and, along the way, heedlessly created an environment fatal to humans. "Imagine a situation where the temperature rises by 100° or is lowered by 100°. We'd go extinct in a matter of minutes," Tallinn says. Tegmark agrees: "The real problem with AI is not malice, it's incompetence," he says.

A current-day analogy is the 2015 tragedy in which a suicidal Germanwings pilot told his plane's computer to descend to an altitude of 100 meters while flying over the French Alps. The machine complied, killing all 150 on board, even though it had GPS and a topographic map. "It did not have a clue about even the simplest human goal," Tegmark says. To avoid such calamities, scientists are trying to figure out how to teach AI human values and make sure they stick, a problem called "value alignment." "There might be fewer than 20 people who work full time on technical AI safety research," Bostrom says. "A few more talented people might substantially increase the rate of progress."

Critics say these efforts are unlikely to be useful, because future threats are inherently unpredictable. Predictions were a problem in every "foresight exercise" Tait has taken part in, she says. "We're just not good at it." Even if you foresee a risk, economic, political, and societal circumstances will all affect how it plays out. "Unless you know not only what is going to happen, but how it is going to happen, the information is not much use in terms of doing something about it," Tait says.

Pinker thinks the scenarios reveal more about human obsessions than real risks. We are drawn to prospects "that are highly improbable while having big impacts on our fitness, such as illicit sex, violent death, and Walter-Mittyish feats of glory," he writes. "Apocalyptic storylines are undoubtedly gripping—they are a supernormal stimulus for our morbid obsessions." Sure, he says, one can imagine a malevolent, powerful AI that people can no longer control. "The way to deal with this threat is straightforward: Don't build one."

Tallinn argues it's better to be safe than sorry. A 2017 survey showed that 34% of AI experts believed the risks associated with their work are an important problem; 5% said they are "one of the most important problems." "Imagine you're on a plane, and 40% of experts think that there is a bomb on this plane," Tallinn says. "You're not going to wait for the remaining experts to be convinced."

Price says that critics who accuse him and his colleagues of indulging in science fiction are not entirely wrong: Producing doomsday scenarios is not that different from what Shelley did. "The first step is to imagine that range of possibilities, and at that point, the kind of imagination that is used in science fiction and other forms of literature and film is likely to be extremely important," he says.

Scientists have an obligation to be involved, says Tegmark, because the risks are unlike any the world has faced before. Every time new technologies emerged in the past, he points out, humanity waited until their risks were apparent before learning to curtail them. Fire killed people and destroyed cities, so humans invented fire extinguishers and flame retardants. With automobiles came traffic deaths—and then seat belts and airbags. "Humanity's strategy is to learn from mistakes," Tegmark says. "When the end of the world is at stake, that is a terrible strategy."