RELAX: the AI apocalypse has been cancelled. A rash of recent headlines blared that Google was developing a ““, as The Telegraph put it, with a picture of a menacing metal army.
It may come as little surprise to learn that the on which the stories were based described a prosaic engineering problem, not ways to stop the Terminator in its tracks. But the excitable coverage reveals how deeply the challenges posed by artificial intelligence have seeped into public consciousness.
We have had machines that can out-calculate us for decades. Now a new wave is outperforming us on tasks ranging from image recognition to video-gaming. They might soon do our jobs better than we can (see Find your meaning) and may even challenge us in areas as sacrosanct as creativity.
Advertisement
Such superintelligent machines could revolutionise everything from transport to social care. But their rise raises tricky questions about everything from human survival to theology. Â鶹´«Ă˝ went in search of answers at a private meeting of researchers, philosophers and ethicists organised by Rustat Conferences at Jesus College, Cambridge, UK.
Not so killer
Concern that smart machines might do away with us has been brewing since the advent of modern computers in the 1950s, but was confined to the wilder fringes of AI. In recent years, however, a school of thought led by the philosopher Nick Bostrom has made this “existential risk” a mainstream talking point. His 2014 book Superintelligence won over technocrats like Bill Gates and Elon Musk, and later public figures like Stephen Hawking.
“Existential risk boils down to values. The challenge is to ensure an AI’s values are compatible with ours“
One of Bostrom’s more celebrated examples is of an AI bent only on making paper clips: it might use up all of the planet’s resources in pursuit of its objective. Alternatively, an AI tasked with making humans happy might cut out parts of our brains associated with unpleasant experiences. So the challenge is to ensure that an AI’s goals are compatible with our own.
“Existential risk boils down to a question of values,” John Naughton, emeritus professor of the public understanding of technology at the Open University, UK, told the meeting. The bad news, he says, is that those leading the AI charge take “a technocratic attitude that assumes data-driven decision-making is good and algorithms are neutral”. Biases introduced by the systems’ designers, input and processes go unremarked.
So how should we set goals and values for future machines? The simple answer is we don’t yet know. Although current AIs are “trained” on data sets to perform specific tasks, their successors may be able to choose their own objectives, just as we do, and they might figure out better solutions to problems that way. But if we give them that freedom, we need the ability to stop them from taking undesirable paths – hence that so-called kill switch.
One common assumption is that an AI should aim for the greatest good for the greatest number. That initially sounds attractive. For example, it is more cost-effective to buy malaria nets than develop drugs for rare diseases. But that could mean abandoning the kind of individual gestures we hold dear, and which are important for social cohesion. AIs might be able to “out-ethic us”, making cold-bloodedly rational choices on our behalf, but we might not like the results. That will become more of a concern as they push into areas currently reserved for humans.
A glimpse at the new kinds of problem we can expect AI to start tackling came in March, when AlphaGo beat world champion Lee Sedol at the hugely complex board game Go. The tournament – which the AI won 4-1 – was reminiscent of Garry Kasparov’s 1997 bouts with IBM’s Deep Blue supercomputer. But whereas that joust demonstrated machines’ superiority at brute-force calculation, AlphaGo’s victory showed something else, Demis Hassabis, co-founder of AlphaGo’s creator DeepMind, said at the meeting: creativity and intuition.
Getting human
Hassabis proposed that creativity be defined as the ability to synthesise knowledge to produce a novel idea, and that intuition is implicit knowledge acquired through experience that isn’t consciously expressible.
AlphaGo won one match by playing a move that departed from centuries of received wisdom. It can’t express why it did this, but clearly had a rationale. So was it being creative and intuitive, albeit in a very limited way? If so, it might represent a new class of smart machine: “super-creators”, say, rather than supercomputers.
But it is missing the point to describe creativity as an innate property, said Simon Colton, who studies computational creativity at Goldsmiths, University of London. While he looks forward to a future in which, say, your phone endlessly composes music, he says creativity is a social construct, “conferred on a person – or entity – by others”. Colton has made machines that paint pictures and make up storylines, but says it can be impossible to evaluate computer-generated works without imposing an invalid human frame on them.
What about qualities we still think of as exclusively human – imagination, emotion and above all consciousness? Machines that probe these areas are in the works, but the AIs hitting the headlines don’t get anywhere near them. While a system can be trained to perform new tasks, it can’t usually transfer knowledge gained in one area to another, as humans do.
“How would the creation of conscious machines challenge our ideas about our place in the cosmos?“
Many researchers agree that the way most people imagine AI – a machine that thinks just like a human – is a remote prospect, unlikely to be fulfilled without a better understanding of how our own minds work. And the field has a history of “AI winters”, when development grinds to a halt after a period of rapid advance. There’s a broad consensus that such “artificial general intelligence” (AGI) is achievable this century, but few believe it will result from just carrying on as we have so far.
But a superintelligent machine need not replicate all facets of humanity. The spectrum of future machines could include “zombie” AGIs that resemble humans but have no consciousness, says Murray Shanahan, who studies cognitive robotics at Imperial College London – or AIs that are more conscious than us, among other possibilities (see diagram). That puts them into company with extraterrestrial intelligences, which might also be super-smart but utterly inhuman.
There’s a final conundrum: how would the creation of machines as intelligent and/or conscious as ourselves challenge our ideas about our place in the cosmos? Perhaps surprisingly, the religious might need less adjustment. Abrahamic faiths, at least, need not have a problem with there being non-human intelligences, said Cambridge theologian Andrew Davison, given that the Golem of Jewish folklore and immaterial superbeings in the Bible are examples of just that.
We have only just begun to live with smart machines. While we worry today about killer robots, the challenges to come may be turn out to be much stranger. One day, we may find ourselves living alongside aliens and angels.
This article appeared in print under the headline “Outsmarted?”
