Andrew Leigh

HUMANS NEED NOT APPLY: WILL THE ROBOT ECONOMY PIT ENTREPRENEURSHIP AGAINST EQUALITY?

Fall 2015 Distinguished Public Policy Lecture
Institute for Policy Research
Northwestern University

In 2006, chess world champion Vladimir Kramnik was beaten by chess computer program Deep Fritz. In 2011, quiz show champions Brad Rutter and Ken Jennings were beaten on Jeopardy! by IBM’s Watson computer. Modernist composers are experimenting with singing software that can mimic a human voice box, but without its physical limitations.[1] Earlier this year, Google announced that their driverless cars had completed over 1 million road miles in Nevada, Florida, California and Michigan. Among the newlyweds who stand at the altar this year, more than one in three couples were brought together by a computer algorithm.[2]

Breakthroughs in processing power, data availability and machine learning have affected all our lives. Within the past decade, fields such as image search, voice recognition, language translation and robotics have seen huge breakthroughs. While a digital assistant might have seemed fanciful a decade ago, the advances in Apple’s Siri technology suggest that it may not be far off. Surgeons who now use computer-guidance to tell them where to cut may soon be stepping back so that a robot can do the job. Within a decade or two, Douglas Adams fans who admired the Babel Fish may be able to pop a simultaneous translation device in their ear.

For well-educated professionals earning six-figure salaries, the world of artificial intelligence seems exciting, optimistic and – well – cool. And yet I want to argue today that no serious economist should be thinking about the aggregate benefits of technology without considering its distributional implications. Since the path breaking work of Wolfgang Stolper and Paul Samuelson in 1941, trade theorists have known that cutting tariffs raises aggregate living standards, but can make some workers worse off. So too we need to intertwine our understanding of technology with recognising its impact on inequality.

But putting yourself in the shoes of others isn’t easy. So I want to scare you a little, by drawing on an idea that’s increasingly coming out of science fiction and into the newspapers. Perhaps then, when you realise that the monster might in fact be living under your bed, we can talk about what to do about it.

The Intelligence Explosion

One of the most famous ideas in science fiction is the singularity: the notion of a turning point when computers become capable of recursive self-improvement. Unlike human generations, which last around thirty years, computer generations are almost instantaneous. So even if a computer is only making a very small improvement each generation, the power of the system would quickly take off. This ‘intelligence explosion’ would rapidly lead to computers whose capacity was beyond our understanding and control. Once superintelligent machines surpass the ability of humans, we might expect them to be more inventive than Einstein, to compose music better than Mozart’s, to invest better than Buffett, and write plays superior to Shakespeare. After all, amazing as those individuals were, they were still limited by a 3 pound brain with 100 billion neurons. One researcher compares the singularity to the ‘Cambrian Explosion’ half a billion years ago, when creatures suddenly evolved that could see, and within a short time frame, life on earth was transformed.[3]

Who cares about the singularity? Well, hardly anyone. Except for Bill Gates. And Elon Musk. And Stephen Hawking, who wrote recently: ‘One can imagine such technology outsmarting financial markets, out-inventing human researchers, out-manipulating human leaders, and developing weapons we cannot even understand’.^[4]

While some argue that superintelligent machines might happily co-exist with humans, others argue that they would most likely have a malign impact. Nick Bostrom gives the example of computers that just wanted to make as many paperclips as possible.^[5] Pretty quickly, our cars and homes would find themselves becoming fodder for them. As he points out, the chief concern isn’t that machines would be hostile or evil. Instead, the problem is that it is hard to imagine computers having a set of preferences that would positively affect our wellbeing. Even if machines could be engineered so as to pursue only goals that made humans better off, how confident could we be that they would stick with these goals as they ‘evolved’.

Even if you aren’t worried that the singularity will end life as we know it, you should perhaps be concerned about its impact on wages. Nearly two centuries ago, David Ricardo pointed out in his Principles of Political Economy that the introduction of robots which could substitute for a worker meant that the equilibrium wage of the worker would come down to the rental price of the machine.^[6] As robots got cheaper, wages in the equivalent sector would fall. If we’re dealing with a superintelligence – a robot that outperforms humans at every conceivable task – that means all our wages fall to zero.

In his 2005 science fiction novel Accelerando, Charles Stross points out that the chances of people catching up to the machines after the singularity is virtually nil. As one character wryly observes, ‘Humans are just barely intelligent tool users; Darwinian evolutionary selection stopped when language and tool use converged, leaving the average hairy meme carrier sadly deficient in smarts.’

Stross’s novel offers a list of answers to ‘Frequently Asked Questions’ that might be asked by ‘humanoids’. It reads in part:

‘While many things are free, it is highly likely that you possess no employable skills, and therefore, no way of earning money with which to purchase unfree items. The pace of change in the past century has rendered almost all skills you may have learned obsolete [see: singularity]. However, owing to the rapid pace of change, many cooperatives, trusts, and guilds offer on-the-job training or educational loans.’

Scared yet?

Whether you worry about the singularity depends on how you think about the progress being made in artificial intelligence. Right now, there are people working on the problem from two angles.^[7] One is the bottom-up approach of whole brain emulation, which means trying to understand each specific part of the brain and replicate the way it works. The other is the top-down approach of working out in broad terms what the brain can do, and then trying to figure out how to make a functional replica.

Both approaches have benefited greatly from exponential growth in computing performance. Since the 1960s, Moore’s Law has seen processing power double roughly every 18-24 months. Worldwide data storage is now around a zettabyte, or 10²¹ bytes. Internet connection speeds never seem fast enough, but their improvement is marked by the fact that the next mobile telephony standard, 5G, will operate at one gigabit per second. People often wonder who will watch the 300 hours of video that are uploaded to YouTube each minute. One answer may be that when artificial intelligence gets good enough, robots will use this material to learn about how to interact with the world. (Though this could end badly if it gets hooked on the Funniest Home Videos channel.)

How far off is the singularity? Wags like to say that ever since the 1950s, a breakthrough in artificial intelligence has been 20-30 years away. They point out that this is just close enough to feel attainable, but no so near that a research funding body will be disappointed if you fail to reach the goal within the term of your grant.  Appropriately enough, a survey of participants at the 2012 ‘Singularity Summit’ returned a median estimate of 2040.^[8]

From the Singularity to Skill-Biased Technological Change

As a politician, the singularity isn’t the top of my policy priority list. But it’s worth us recognising that ‘tail risks’ – unlikely events with very bad consequences – are worth devoting resources to. If you’re interested in the risks of artificial intelligence – or for that matter meteor strikes and deliberately engineered viruses – I commend to you the work of US think-tanks such as the Future of Life Institute and the Global Catastrophic Risk Institute, and UK bodies such as Oxford’s Future of Humanity Institute and Cambridge’s Centre for the Study of Existential Risk.

The real point of raising the singularity is because, in my experience, academic audiences think of technological displacement as something that happens to someone else. The economic notion of skill-biased technological change suggests that technological advances tend to lower wages for those with less education and raise wages for those with more education. This means that workplaces which are overstuffed with PhDs are likely to benefit greatly from technology. For example, if you’re a leading empirical economist like David Figlio, you have access to more data than ever before, which you can analyse on faster computers than ever before. You can collaborate across states and even nations using high-capacity email, rapid videoconferencing and cheap airfares.

In their book, The Race Between Education and Technology, Harvard economists Claudia Goldin and Larry Katz suggest that inequality is – wait for it – a race between education and technology. In eras when education expands rapidly (I’m looking at you, 1950s America), inequality falls. But in more recent times, US education has expanded more slowly, and inequality has risen.

Just how bad is inequality? To an outsider, it is striking to recognise the stagnation of US wage and income growth. Like all of you, I eagerly leafed through the Census Bureau’s annual report on income and poverty in the United States when it was released last month. The report is a treasure trove of interesting data – most of it painfully dispiriting.

See if these numbers shocked you as much as they did for me. After inflation:

• Median household income is where it was in 1989.^[9]
• Full-time wages for men are where they were in 1973.^[10]
• Since 1973, a household at the 10^th percentile is no better off, while a household at the 90^th percentile has about 1½ times as much income.^[11]
• The poverty rate is where it was in 1966.^[12]
• The child poverty rate is where it was in 1965.^[13]

Ok, that’s a lot of numbers. So let’s put another way. For movie buffs: full time men haven’t gotten a real pay rise since The Exorcist came out, and the typical household has no more resources than When Harry Met Sally was released.

For music fans, the child poverty rate is as bad now as when the Beatles toured the US (packing out White Sox Stadium along the way). Households in the bottom tenth have as few resources now as they did when Pink Floyd released Dark Side of the Moon.

Now, I see there’s a few Gen-Ys scratching their heads at some of these references. If it all seems like ancient history, that’s kinda the point. In most cases, your entire life has seen no real income gains for Americans at the middle and bottom of the distribution.

Admittedly, we need to be careful about what we’re measuring. Compared to three or four decades ago, US employees are more likely to receive non-cash remuneration (in particular, health insurance). But even studies that take this into account recognise that – for the first time in the post-war era – living standards for many Americans are no better than those of their parents.

A host of factors have exacerbated American inequality. Since 1970, the unionisation rate has fallen from a quarter to a tenth.^[14] The top tax rate has been cut from 70 percent to 40 percent. The economy is more open, with exports as a share of GDP more than doubled.^[15] Immigration, minimum wages, social trends and welfare policies may also have played a part.

However, my focus today is on technology, and its potential to make life harder at the bottom – by destroying jobs or driving down wages.

At the outset, it is worth noting that the claim that technology might generate joblessness has a long history. In Nottingham in 1811, disgruntled textile workers wrote to factory owners under the pen-name ‘Ned Ludd’, threatening to smash machines if they continued to be used. In 1930, Keynes worried that ‘the increase of technical efficiency has been taking place faster than we can deal with the problem of labour absorption’. In Australia, highly unionised shearers defended the use of narrow shearing combs until 1983, with one official describing wide shearing combs as ‘immoral and repulsive’.

In each instance, there was a kernel of truth to the concerns. Since 1900, the share of the US population involved in agriculture has fallen from 40 percent to 2 percent (and in Australia, from 24 to 2 percent). Fewer people now work in the farm sector than at the start of the last century, yet the real value of agricultural output has nearly tripled (and increased sixfold in Australia). Go onto a modern farm today, and you’ll see how chemistry has changed what animals are fed and how GPS software has changed how crops are harvested. Some of these changes can be a bit confronting. To take an example you might enjoy sharing around the Thanksgiving table, most US turkeys now reproduce through artificial insemination, since we’ve bred birds that find it hard to reproduce naturally. Yet all those technological changes in agriculture have increased output per worker. Never in human history have we fed so many mouths with so few farm hands.

How Jobs Hollowed Out in the Middle, and Wages Fell at the Bottom

What’s happened across the job distribution? In a series of articles, David Autor and co-authors have divided jobs into three categories: manual, routine and abstract.[16] They show that routine jobs, which tend to fall in the middle of the wage distribution, have shrunk most. This is true in the US for the 1980s, 1990s and 2000s.[17] It is also true of most European countries over the period 1993 to 2010.[18]

Routine jobs are occupations such as bookkeeping, administrative support, and repetitive manufacturing tasks. For example, when I worked as a junior lawyer in the early-1990s, each firm had a typing pool - a group of people whose job was to take recordings from micro-cassettes and turn them into well-formatted contracts, letters and file-notes. Such typing pools no longer exist. Right now, café ordering apps such as ‘Hey You’ are displacing the routine work historically done by cashiers. What makes routine jobs vulnerable to computerisation is that they involve following established rules. ‘Put the tape in the recorder. Type up the letter. Save and print.’ Or: ‘Scan the customer’s order. Take their payment.’

By contrast, abstract jobs involve problem-solving, creativity, and teamwork. When a lawyer advises a client whether to accept a plea bargain or a manager decides how to respond to an employee arriving late for work, they are tackling problems that do not have a closed-form solution. It is unlikely that computers will replace the work presently done by artists, veterinarians, merchant bankers, architects, university researchers or politicians – at least until the singularity comes along.

More interesting is the resilience of manual jobs to computerisation. Despite the predictions of science fiction - from Star Wars to the Jetsons - attempts to automate the work of jobs such as cooking, cleaning, security work and personal care have largely failed. For example, researchers at the University of California Berkeley are working on a robot that can fold laundry. At present, their best prototype can do a towel in 90 seconds, but finds itself stumped by new items, such as socks and shirts.[19]

This challenge of shape recognition has stumped other attempts at automation. Current robot hairdressers produce a result similar to what you'd get if you consumed a bottle of tequila and tried to give yourself a haircut without a mirror.[20] An experiment in cat recognition by Google X labs found that the system incorrectly identified a pair of coffee mugs as a cat. Because humans are much better at identifying objects than computers, Amazon’s Kiva book-packing system uses both robots and humans. When you order a book, Amazon’s robots fetch the relevant bookshelf that contains the tome you want. Shelves aren’t classified alphabetically – instead, items are stored along with others that are likely to be bought by the same customer. But the last step – taking the right book off the shelf and packing it – is done by a human.

A frequently cited working paper by Carl Frey and Michael Osborne looks at the characteristics of jobs that have been computerised, and attempts to model the likelihood that particular occupations will be rendered redundant.[21] It’s an appealing analysis, so long as we put appropriately large confidence intervals around their estimates. But what’s important about the paper is that it attempts to identify the characteristics of jobs that will perish versus persist.

In my view, some of the characteristics that will be hardest for computers to mimic are those of communicating clearly with co-workers, showing empathy to clients, and adapting to new situations. This suggests that it will be hard to replace security guards, aged care nurses or masseuses, at least in the short term. A significant number of manual jobs will be around for the next few decades.

What does this mean for wages? Because job loss has been concentrated at the middle of the distribution, people have referred to its effect of 'polarising' or 'hollowing out' the labour market. Some have mistakenly inferred that because the occupations where jobs are lost have been at the middle, therefore the pain has been most acute for mid-level workers or middle class households.

The reason that we can't go from an impact on mid-level jobs to an impact on middle-class families is that we also need to think about labour supply. It's one thing to know that the demand for security guards and dog-walkers is booming. But because these occupations require relatively little training, the supply of workers is also high. As workers in middle-paid jobs have become redundant, they have cascaded down to compete with those in low-paid jobs. The net result is that for those in the bottom half of the US wage distribution, earnings after inflation are back in the Brady Bunch era.

At the top, it's an altogether different tale, with technology augmenting the skills of the most skilful. I recently watched Melbourne surgeon Peter Choong carry out a total knee replacement, using computer guidance to help him determine precisely what angle to cut the bone. Peter is one of Australia's best surgeons, and the computer means he can do a better job with less effort.  That’s likely to boost demand for his services. In coming years, robotic devices may even make the cuts, mix the adhesive cement, and fit the prosthesis. Peter is entirely relaxed about these changes, which he regards as simply improving the accuracy of surgery. As a prolific researcher, he’s well placed to drive the change, and it’s hard to see technology doing anything but making him more productive.

Let's take another category of abstract workers. For CEOs, one of the drivers of rising pay has been that the biggest firms have gotten bigger. This is partly a technology story. Managers have benefited from advances in travel technology, communications technology, data analytics and systems integration, which make it possible to run bigger firms than ever before. To the extent that a CEO’s salary is proportional to the size of the firm that he or she heads, technological advances that increase firm size also increase managerial earnings.   

What Should Policymakers Do?

In this speech, I’ve argued that innovation has the potential to increase inequality. Indeed, this is precisely the finding of a recent paper by Philippe Aghion and co-authors, which uses panel data on top 1 percent shares to estimates that about one-fifth of the rise in top income inequality over the past generation is attributable to innovation.[22] This doesn’t mean that we need to curb innovation – indeed I support innovation-boosting measures such as reforms to the intellectual property system, banning non-compete clauses in employment contracts, and income-contingent student loans.

If innovation naturally tends towards widening the gap, then we need to make sure we have policies in place that exert the opposite pressure – towards a more equal distribution of resources.

Here are three ideas.

First, we should do as much as possible to ensure that people can make seamless transitions between jobs. Former Clinton adviser Gene Sperling notes that retraining programs for people who lose their jobs are often derided as ‘burial insurance’. As he points out, ‘A politician promoting dislocation policies often feels like he or she is telling a newlywed husband how much counselling he could receive if his wife leaves him for his best friend. Most happily married people would rather focus on maintaining their marriage, and most workers happy with their paychecks would prefer to focus on keeping their jobs.’^[23] This means that policymakers should ensure that people are ‘lifelong learners’: equipped with the skills to continually take on new roles and challenges as jobs evolve.

The problem with lifelong learning is that tertiary education is only as good as the primary and secondary education that preceded it. Over recent decades, the academic aptitude of new teachers in the United States has declined significantly.[24] So while test scores for 9 year-olds and 13 year-olds show gains since the early-1970s, one can only imagine how much bigger these gains might have been if teacher aptitude had been rising.[25] The typical new teacher in the United States is someone who was in the bottom half of the aptitude distribution in his or her own class.[26] So it’s not surprising that some of today’s students graduate without a love of learning – and indeed, with a desire never to see the inside of a classroom again. Getting teacher quality right will pay off with higher returns to education: partly because a great education makes people more effective at their jobs, but also because a terrific schooling prepares people to learn new skills as the need arises.

Second, widening wage inequality should encourage us to think about whether we can help low-wage consumers get a better deal. In his recent book, Oxford’s Tony Atkinson argues that there should be an explicit consideration of inequality in competition policy. Indeed, when arguing for competition laws in the United States, Senator John Sherman once cited excess inequality as one of his concerns, noting that ‘inequality of conditions of wealth, and opportunity that has grown within a single generation out of the concentration of capital into vast combinations to control production and trade to break down competition’.  A more explicit consideration of inequality in competition policy would not change most outcomes, but might make a difference on the margins, with considerations such as the distribution of outlets in low-income neighbourhoods.

Inequality should also shape how policymakers regulate the sharing economy. At one extreme is the laissez-faire approach that says you don’t have to comply with minimum wage laws if the job is posted on TaskRabbit. At the other extreme is the approach that says Airbnb homes can only be listed if they comply with all the rules that apply to hotels. Laissez-faire puts downward pressure on low-skilled wages. Overregulation means that a low-wage family might not be able to afford a week’s holiday away from home. Only by thinking carefully about the impact on low-income workers and low-income consumers can policymakers get the settings right.

Third, the potential for wages to remain stagnant in the bottom half of the distribution makes it important for everyone to have a reasonable slice of capital. A worker in the ‘gig economy’ who owns their own home is in a very different position from someone with both precarious work and no assets.

In Capital in the Twenty-First Century, Thomas Piketty tells a story about inequality that is principally about capital returns versus wage rates. But this is not the explanation that best fits the rise in inequality in the United States over the past generation – or Australia, for that matter. As Piketty himself notes, the story of rising inequality in English-speaking countries over the past few decades has been largely a story of growing wage inequality.

But while Piketty’s story about capital accumulation doesn’t describe the past for our two countries, it could portend the future. Rapid advances in automation are likely to cause wages at the bottom to fall. In this environment, a central question will be whether or not each person owns capital assets. That makes it more critical than ever before to have zoning regulations that improve housing affordability; to use tax incentives and behavioural tricks like ‘Save More Tomorrow’ to equalize retirement savings; and to build the evidence base on financial literacy programs.

The trick with any capital allocation program is to think about how it can be done in a way that doesn’t cause people to drop out of the labour market. For example, both Norway and the Gulf States have policies that aim to fairly distribute natural resource revenues among their citizens – but Norway manages to do this in a way that maintains fairly high participation rates, while the Gulf States have very low employment rates among citizens, with much of the work being done by temporary migrants. Taking account of ‘income effects’ is central to designing any sensible program to address capital income inequality.

Conclusion

In the 1989 film Back to the Future II, Doc and Marty McFly travelled 25 years into the future, ending up on 21 October 2015 (last Wednesday). ‘Back to the Future Day’ provided a chance to look at which technological predictions were proven right, and which ones didn’t come through. The film did predict the growth of videoconferencing, flat panel screens, drones, wearables, and biometric identification. But we’re not quite there on self-drying jackets, garbage-powered vehicles and self-tying shoes. Alas, a well-functioning hoverboard remains as elusive as the 2015 Cubs’ World Series victory that the movie promised.

Most interesting is the technologies that the movie-makers missed. They didn’t count on the internet, smartphones and massive data storage – which have together transformed the world of work and social interactions. There’s no hint of the fall of the Soviet Union, the rise of China, an African-American president and same-sex marriage. In short, it’s a movie that overconfident futurologists would do well to watch.

We can’t predict all the challenges and opportunities of the coming decades, but it is worth thinking about how – in the words of Jerry Kaplan – we can make it look more like Star Trek than Terminator. It’s not enough to think about the effect on total output – we need to think about how technology affects the distribution of earnings. The larger the market, the more that technology acts as a ‘force multiplier’ for those at the top of their game. As Bill Gates once put it, ‘A great lathe operator commands several times the wage of an average lathe operator, but a great writer of software code is worth 10,000 times the price of an average software writer.’[27]

If the trends of the past generation continue, we will likely see fewer middle-paid jobs, and an erosion of wages in the bottom half of the distribution. To use a nautical analogy, that means bigger yachts and plenty of tugboats to pull them around. But fewer regular-size runabouts. As it happens, that’s what I saw when I last visited Miami.

There isn’t a single answer to addressing the rise in inequality, but I’ve suggested three: creating a schooling system that fosters lifelong learning, ensuring that the market helps serve low-wage consumers, and thinking about creative ways to distribute capital more broadly. The productivity-boosting effects of innovation are certain to raise aggregate GDP: the real challenge is to ensure they raise median incomes too. The singularity may be decades away. But robots are widening the gap today, and we need smarter policies in response.