Science-fiction writers have long imagined how humans might structure society when machines can do all the work. Economists and political scientists have spent rather less effort figuring out how we might manage the transition to such a world from today’s policy environment.
This question is becoming more urgent than you might think.
Few people outside the field of machine learning have yet grasped how astonishingly quickly it’s progressing. Machines are now able to do things that, a few years ago, even optimistic experts thought were decades away, such as driving cars safely in real-world conditions, understanding language, reading handwriting and interpreting images.
We’re very close to there being nothing a human can do better than a machine. And what’s really noteworthy is how rapidly and recently this has happened.
The company I used to work for, Kaggle – one of the World Economic Forum’s Tech Pioneers of 2014 – runs machine learning competitions. Companies come with data and a question: how can an insurer predict claim levels, or an airline predict no-shows, or a retailer optimize prices? Data scientists compete to write algorithms that can outperform the company’s current ways of working.
Until a year or two ago, winners of these competitions would typically have crafted algorithms using a lot of problem-specific tuning. As a simplified analogy, imagine an expert board-game player writing an algorithm to play that game by first coding descriptions of a variety of potential strategies, and then having the computer find the optimal parameters for and combinations of these strategies.
Increasingly, winning entries are based on an algorithm called Deep Learning, which effectively cuts out the human. Imagine someone who has never heard of a board game writing an algorithm to play it by feeding in reruns of previous games and letting the machine figure everything out itself. And this algorithm beats the expert’s algorithm hands down.
A real-world example: when Merck asked for help analysing molecular data to find promising drugs with no unintended side effects, the best-in-class algorithm was written in two weeks by students with no background in medicine.
As machines learn how to learn, we are now seeing these best-in-class algorithms becoming exponentially better – typically about twice as good with every passing year. There’s a feedback loop: as computers get faster, and data volumes get larger, and the results of previous algorithms guide the development of the next iteration, the algorithms continue to improve.
Exponential growth is a concept we all understand intellectually, but which still tends to trip us up. Remember the famous story of the Chinese emperor who agreed to pay one grain of rice for the first square of a chessboard, two for the second, four for the third, and so on; soon he owed more than all the rice in the world. Now think about machines that are twice as smart this year as last year, and try to imagine what they might be able to do in two, 10 or 50 years. Like the Chinese emperor’s, our minds short-circuit.
We’ve experienced once before in history a period where a wave of innovation saw machines take over human jobs: the industrial revolution. Eventually, that was a good thing for everyone: new industries evolved to employ people, and wealth was spread. But only after a decades-long dislocation that plunged millions into penury and fuelled profound social upheavals.
Today we’re on the cusp of something not only more powerful, but also permanent. Already, in recent years, wages are stagnating while productivity improves. As machine learning accelerates, this decoupling will, too. Perhaps much sooner than we expect, it may become impossible for most humans to contribute any economic value ever again.
This poses a profound challenge to one of the fundamental aims of our current political system: rewarding humans who create economic value, and punishing those who don’t. Most of us would rather technology liberates all humans to flourish, than condemns most to poverty. But how do we manage this transition from scarcity to abundance?
Switzerland’s plan to guarantee every citizen a basic income – to be voted on by referendum this May – suggests one possible path. Another possibility is the “negative income tax” proposed by Milton Friedman, where low-income workers would be able to do meaningful work even though a machine might be able to do it more efficiently.
We should all start to think about how we’d like the world to look when machines can do all our jobs; it’s one of the few questions we won’t want machines to answer.
Image: Children try out laptops in Seoul. REUTERS/Bobby Yip