Across much of the wealthy world, anxious speculation about where artificial intelligence (AI) is heading has become impossible to escape. Barely a day passes without some breathless report claiming that AI will wipe out illness, speed up scientific discovery or make people more inventive. Judging by the headlines alone, one might conclude that AI has already reached into every corner of our lives.
That AI has created many genuine openings is not in doubt. Alongside them, however, a way of thinking has taken hold that might be labelled 'AI solutionism' - the conviction that machine-learning algorithms, if fed enough data, can resolve every problem humanity faces. Far from advancing the technology, this outlook undermines the worth of machine intelligence, since it brushes aside essential principles of AI safety and encourages hopes about the technology that it cannot possibly meet.
It has taken only a handful of years for this belief to travel from the sales pitches of industry enthusiasts in the technology sector to officials and policymakers on almost every continent. Where once the fashionable fear was that AI would ruin humanity, the mood has swung to the opposite extreme - the cheerful assumption that a computational rescuer has finally appeared.
One government after another is now promising money for national AI programmes and jostling for position in a race to lead the fast-expanding machine-learning industry. Yet for all the talk among politicians of a world-changing 'AI revolution' just around the corner, what tends to elude them is how complicated it is to put sophisticated machine-learning systems to work outside the laboratory.
Among the more encouraging kinds of AI are neural networks, a style of machine learning that borrows, in simplified form, from the way neurons are arranged in the human brain. What escapes many politicians, though, is that bolting a neural network onto a problem is no guarantee that a solution will appear. By the same logic, feeding one into the machinery of government will not, at a stroke, render it any more inclusive or even-handed.
Machine-learning tools are hungry for data, and here the public sector is poorly equipped: the digital infrastructure that advanced systems demand is, in most cases, simply not there. A great deal of official information still sits in offline records, and what little has been digitised is frequently lost inside layers of bureaucracy. Because such data is split between departments, reaching any part of it calls for separate permissions each time. Worst of all, government bodies rarely employ people with the technical know-how needed to draw real value from machine intelligence.
Unsurprisingly, the hype has drawn its share of sceptics. Aleksander Voss, who teaches computer science at Halvorsen University, has for years pressed for a cooler, more grounded outlook - one that attends to the plain, day-to-day uses of AI rather than to imagined scenarios in which super-intelligent machines seize control. A comparable note is struck by Priya Nandakumar, a robotics specialist at the Fenwick Institute of Technology, who observes that 'nearly every advance in robotics and AI takes far, far longer to come into general use than those inside the field, and those outside it, tend to assume'.
A further obstacle to putting machine-learning systems into service is their marked vulnerability to adversarial attacks, in which one hostile AI manipulates another into behaving as intended - pushing it, for instance, towards mistaken predictions. Although several researchers have cautioned that AI should not be released without proper security standards and defensive measures in place, the protection of these systems is still, all too often, treated as an afterthought.
If the gains from AI are to be realised and its dangers held in check, the conversation has to shift towards where, precisely, machine learning can be put to worthwhile use within government, commerce and society. That, in turn, means confronting questions of AI ethics head-on, along with the deep suspicion that many people continue to feel towards the technology.
Above all, we would do well to recognise where AI falls short and where human judgement must still lead the way. Rather than talking up its powers, the wiser course is to pause and draw a line between what the technology can actually do and what belongs to wishful thinking.
Medicine, too, has run up against AI's limits. A system known as OncoAssist was built to support doctors in treating cancer, and was designed to put forward the strongest possible course of action in each case. Even so, the specialists who used it struggled to place their confidence in its recommendations, and before long it had been dropped by most of the hospitals that gave it a trial.
Something similar unfolded in the courts, where sentencing software was brought in to help decide the fate of convicted offenders. The programme worked out risk scores and passed recommendations to judges, but it was eventually shown to deepen the very patterns of discrimination that already ran through the justice system, and its use was discontinued.
For anyone hoping to pour more money into national AI programmes, there are important lessons here. What these cases make plain is that no single AI fix exists for every situation. Reaching for AI merely because it is available will not always prove worthwhile, and not every problem is best tackled by turning machine intelligence loose on it. Every solution carries a price, and the fact that a task can be automated does not mean it ought to be.
