Will this decade finally be the one in which the machines take our jobs? Such concerns have been aired many times over the centuries and they have always been wrong. But they are not intrinsically absurd. In 1979, the economist Wassily Leontief pointed to the fate of the horse. Horses had long been of vital economic importance, but faded in the second half of the 20th century as the internal combustion engine became the dominant source of horsepower.
Horses still have a niche, but will never outcompete engines, no matter how cheap oats become. Might large numbers of human workers go the way of the horse? In 2003, the economists David Autor, Frank Levy and Richard Murnane published a study of the economics of technological change that made two influential observations.
First, they pointed out (correctly) that it is misleading to talk of robots — or any other technology — taking jobs. Instead, machines perform tasks, a narrower unit of work. Since most jobs involve many different tasks, robots do not take jobs, but they may radically reshape them. A robot accountant is not C-3PO; it’s Excel or QuickBooks. As with the horse, there is no wage at which human calculators can compete with a computer at the task of adding up a spreadsheet. Still, human accountants exist in large numbers. Their jobs simply look very different today.
Second, argued Profs Autor, Levy and Murnane, the tasks that machines took on were best described not as “skilled” or “unskilled” but as “routine” or “non-routine”. Recalculating a spreadsheet is a skilled but routine task, easily automated. Cleaning a toilet requires little skill — even I can do it — but is non-routine and therefore hard to automate.
This way of looking at the world proved very useful. It explained why technology could disrupt our jobs without destroying them. And why both the low-paid and high-paid ends of the labour market were proving robust, while the middle, packed with skilled-yet-routine tasks, was hollowed out.
But in a new book, A World Without Work, Daniel Susskind argues that the second part of the Autor-Levy-Murnane perspective is proving more questionable. He observes that the boundaries of the “routine” are blurring fast.
Consider, for example, CloudCV, a system that answers open-ended questions about images. Upload an image and ask any question you like. One photograph showed some 20-somethings sitting on a sofa with white wine and cans of Kronenbourg lager in front of them, with one fellow standing in a dramatic pose.
“What are they doing?” I asked the computer. “Playing Wii,” it replied, correctly. “What are they drinking?” Probably beer, it said. “How’s the weather?” I asked of an outdoor snapshot. “Cloudy.” It was. The system gives accurate answers to informally phrased questions about random photographs. Is that task routine? Hardly.
Neither is the performance of Alpha Zero, the game-playing algorithm developed by DeepMind, a sister company of Google. In 2017, AlphaZero trained itself in a few hours to thrash the best chess-playing engine and the best Go program, both of which easily beat the best humans. Some claim this performance is less impressive than it first appears — but 10 years ago the mere idea that a computer could beat a human at Go seemed implausible. What DeepMind’s supercomputers can do today will be achievable on laptops and phones by 2030.
In task after task, the computers are overtaking us. In the Visual Question Answering challenge that CloudCV attempts, humans score 81 per cent. The machines were at 55 per cent as recently as 2016; by the summer of 2019 they were at 75 per cent. It’s only a matter of time before they do a better job than us — just as AlphaZero does.
The Artificial Intelligence Index project, based at Stanford University, tracks a wide variety of benchmarks. The machines are making rapid progress at symbolic achievements — such as playing poker — but also at translation, speech recognition, and classifying diseases such as skin cancer (from images of moles) and diabetes (from images of retinas).
These achievements are real. And despite the fact that there are many things computers cannot do, when an algorithm does a narrow task cheaply and well, we humans end up contorting ourselves to unleash the new capability while sweeping up the tasks the software leaves behind. Just look at the self-checkout at your local supermarket.
So — will the machines take all the jobs in the coming decade? No, and that remains an unhelpful way to phrase the question. Machines encroach on tasks, and we reorganise our jobs in response, becoming more productive as a result.
But there is good reason to believe that such reorganisations will be wrenching in the decade to come, and also that some people will be permanently unable to contribute economically in the way they would have hoped and expected.
Above all, it is likely that our political institutions will be unable to adapt to the challenge.