87Why Are There Still So Many Jobs?

David H. Autor, Journal of Economic Perspectives, 2015

Linked to 11 chapters: supports 8 · contradicts 3 · mechanism 4 · same claim 1 · pairing 2

A Time story of 24 February 1961 reported on the automation jobless and warned that machines were about to move from the factory into the office. Three years later Lyndon Johnson empanelled a national commission on technology and automation to face the possibility that productivity was rising faster than the demand for labour could absorb it. The commission concluded that machinery destroys particular jobs rather than the need for work, and then recommended a guaranteed minimum income and government as employer of last resort, in case it was wrong. David Autor opens with that episode because the argument of 2014 — Erik Brynjolfsson and Andrew McAfee's warning that there has never been a worse time to be a worker with only ordinary skills — restates it almost exactly. His reply is not that automation spares workers; it displaces them, and displacing them is what it is built to do. It is that displacement is one half of a mechanism whose other half is almost never counted: work is made of many tasks, and improving the machine's half of the bundle raises the price of the human half.

How automation and employment interact

In 1900, 41 per cent of the American workforce was employed in agriculture; by 2000 the share was 2 per cent. The mass-produced car emptied the equestrian trades; earth-moving equipment took manual labour out of construction; a computer running a payroll performs work a person once did by hand. Autor grants all of it and then asks the question the record forces: given that labour-saving technology demonstrably saves labour, and that more arrives every year, why has it not wiped out employment for most people?

His answer is that tasks which cannot be substituted are generally complemented. Most work processes draw on many inputs at once — capital and labour, rote repetition and judgment, rules and discretion — and an improvement in one does not remove the need for the others. Michael Kremer's O-ring production function, named for the seal that destroyed the Challenger, states this formally: if any link in a chain of production fails, the whole process fails, so making the other links reliable raises the value of making this one reliable too. When computerisation makes some steps cheaper or more dependable, the remaining human steps become worth more, not less.

The example Autor takes from James Bessen is the cash machine. Automated tellers arrived in the 1970s and their numbers in the United States quadrupled from about 100,000 to 400,000 between 1995 and 2010. Employment of bank tellers did not collapse; over the thirty years to 2010 it rose from about 500,000 to 550,000, though more slowly than the labour force. Two forces ran against each other. Because a machine made a branch cheaper to operate, banks opened more branches: tellers per branch fell by more than a third between 1988 and 2004 while urban branches, helped by deregulation, rose by more than 40 per cent. And as routine cash handling receded, the job changed, from counting money to selling credit cards, loans and investment products to people the teller knew.

Autor is careful that this is an illustration and not a law. Three things decide whether the complementarity reaches anyone's wages. The first is which tasks a worker supplies: a construction worker expert with a shovel who cannot drive an excavator loses, and so does a teller who can tally currency but cannot sell. The second is the elasticity of labour supply, since where the complementary work is easy to enter, new entrants compete the gains away. The third is the elasticity of demand for the output: farm productivity rose spectacularly and the food share of household income fell, while medical productivity rose and the health share rose. Income freed in one sector is spent in another, often one no machine touches: demand for restaurant meals and haircuts, what William Baumol called technologically lagging, rises with income even though nothing makes them cheaper. Consumption has kept pace with productivity rather than approaching satiation: an American worker in 2015 who wished to live at the average income of 1915 could reach it by working about seventeen weeks a year.

Polarisation in the American labour market

Even where automation leaves the number of jobs alone, it changes which jobs exist. Drawing on Lawrence Katz and Robert Margo, Autor compares the change per decade in employment shares across seven occupational groups over 1940–1980 and 1980–2010. In the first period the movement was unambiguous: agriculture fell by almost four percentage points a decade, professional and technical work grew by two and a half and managers by half a point. After the late 1970s the pattern broke. Professional, technical and managerial employment grew faster still, but every other favourable shift stopped or reversed: skilled blue-collar occupations shrank, clerical and sales work — the production jobs of the information age — turned down, and low-paid personal services began absorbing a rising share of non-college labour.

To explain the difference Autor asks what computers actually do. A computer follows a procedure a programmer laid out, so someone had to understand the steps well enough to write them down. That has not changed; the cost has. William Nordhaus estimated in 2007 that a standardised set of computations had fallen in price at least 1.7-trillion-fold since the era of manual calculation, most of it since 1980. Autor, Frank Levy and Richard Murnane call the work that can be written down routine, not because it is dull but because it is codifiable: bookkeeping arithmetic, the sorting of structured information, a repetitive physical operation in an unchanging environment.

The limit on this substitution is that people do a great deal they cannot explain. Autor names the constraint after Michael Polanyi, who observed in 1966 that "we know more than we can tell". Breaking an egg on the rim of a bowl, recognising a bird from a glimpse, writing a persuasive paragraph: these run on tacit understanding. Hence the inversion Hans Moravec also noticed. Formal reasoning was developed deliberately and is easy to encode; perception, dexterity and speech were evolved rather than designed, and reverse-engineering them means recovering rules nobody ever held.

What resists substitution divides in two, abstract work and manual work. Manual outputs must be produced on the spot, so they cannot be sent abroad, and almost anyone can be trained to supply them.

Because abstract and manual work sit at opposite ends of the pay scale and routine work in the middle, computerising the middle grows both ends at once; Maarten Goos and Alan Manning named this job polarisation in 2003. The four middle-skill groups — sales, office and administrative support, production, and operatives — were 60 per cent of American employment in 1979, 49 per cent in 2007 and 46 per cent in 2012, while the service share, flat until 1979, has risen ever since. The pattern is not American: across sixteen European Union economies between 1993 and 2010, the middle-paying share fell in every country, by between 4.9 and 14.9 percentage points, while both the high-paying and low-paying shares rose.

Does employment polarisation lead to wage polarisation?

A barbell in employment might be expected to produce a barbell in pay. It has not, and the three qualifications explain why. At the top, computerisation complements the work — the expert spends less time gathering information and more interpreting it — while substituting for the medical secretaries, paralegals and research assistants the professions employ. Demand for professional services has more than kept pace with their rising output, and supply is slow, since a professional takes five to ten years to produce and young Americans responded to the rising return to education sluggishly. Strong complementarity, elastic demand and inelastic supply push pay up.

At the bottom none of the three holds. Cleaners, drivers, guards and home health aides are barely touched by information technology in either direction, demand for their services is price-inelastic, and anyone can enter. Their wages rose through the very tight labour market of the 1990s and then fell after 2000 even as their numbers grew, which Autor reads — candidly, as a hunch — as workers displaced from the vanishing middle moving down into them.

Growth at the top then faltered too. After 2000 there was no relative growth at all in the highest two deciles of the occupational skill distribution, and real wage growth was negative below roughly the fifteenth percentile between 1999 and 2007. The obvious explanation — that computers have begun climbing into professional work — does not fit the investment record: private fixed investment in information processing equipment and software was 3.5 per cent of gross domestic product in early 2014, a level last seen in 1995. Autor adds the other candidates without ranking them: two business cycles, and Chinese import competition after 2001.

Polanyi's paradox: will it be overcome?

In the decade before the paper, machines entered work that had looked closed to them: driving, parsing legal documents, field labour. Autor's reading is that the paradox has not been dissolved so much as gone round, by two routes. The first regularises the environment until an inflexible machine can operate in it. The second inverts the paradox: rather than teach a machine rules nobody can state, let it infer them from examples.

Environmental control is everywhere and mostly invisible. Car plants use robots to fit windscreens on the assembly line; aftermarket windscreen replacement uses technicians, because removing a broken screen and preparing the frame demands adaptation no affordable robot can manage. The difference is not the task but the setting, and to run cars at all humanity has levelled, graded and asphalted a meaningful fraction of the earth's surface. Kiva Systems, bought by Amazon in 2012, automated warehouses without solving the picking problem: its robots move shelves across a level floor while the human handles the merchandise. The self-driving car belongs here too. It navigates by matching its sensors against hand-curated maps of every road, signal, sign and obstacle — Autor likens it to a train running on invisible tracks — and when the environment departs from the one its engineers preprocessed — a detour, a crossing guard where a signal should be — it hands control back to the driver. That handover is the arrangement Lisanne Bainbridge warned about in Ironies of Automation: the operator is kept for the cases nobody anticipated, and given no practice at anything else.

Machine learning takes the other route, applying statistics and inductive reasoning where formal rules are unavailable and learning from worked examples what nobody can state. Autor's illustration is the chair. Specify the features — legs, arms, seat, back — and you exclude backless and armless chairs; relax the requirements and you admit small tables. A six-year-old does it perfectly and cannot say how. A learning algorithm given a large body of labelled examples infers which attributes make an object more or less likely to be called a chair, with no physical model of chairness at all: atheoretical brute force, in Autor's phrase, needing only training data, processing power and software.

His verdict in 2015 is measured. The tools were inconsistent, sometimes uncannily accurate and usually so-so. IBM's Watson won at Jeopardy and, in the same winning match, answered Toronto to a question in the category of US cities; Google's X Lab had recently spent a network of 16,000 processors on finding cats in YouTube videos. The method also has an irony of its own: a machine-learning system cannot tell its programmers why it does what it does either. Autor leaves the question open. Some researchers expected brute force to pass human performance as data and computation grew; others expected it to be right on average while missing the informative exceptions, since what makes an object a chair is that it was built for a person to sit on.

Since 2015

The aggregate prediction held. American employment did not collapse; unemployment fell to 3.5 per cent by 2019, the lowest in half a century. So did the claim that polarisation would not continue indefinitely: through the early 2020s the bottom of the American wage distribution gained fastest, and Autor, with Arindrajit Dube and Annie McGrew, documented an unexpected compression of wage inequality between 2020 and 2022 as low-wage labour markets tightened. The cash-machine example has aged into a caution: teller employment peaked around 2007 and has fallen substantially since. Complementarity is a mechanism, not a guarantee, and it runs only as long as the complementary task survives.

The complementarity argument itself is contested. Daron Acemoglu and Pascual Restrepo attribute most of the change in the American wage structure since 1980 to task displacement never balanced by new work, and argue that a tax code favouring capital produces automation just good enough to replace a worker and no better. On that reading the complementarities are real but too small, and the question becomes which technologies get built.

The machine-learning path, meanwhile, moved faster than the paper allows for and through exactly the door Autor identified. He offered writing a persuasive paragraph as an example of a task performed on tacit understanding; within seven years, systems trained by the method he described were doing it on demand. Environmental control advanced more nearly as predicted: driverless taxis now carry paying passengers in a handful of densely mapped, geographically fenced cities.

What the decade added is that the new capability is uneven in a way the routine and non-routine distinction does not capture. The 2023 experiment at Boston Consulting Group reported in the jagged frontier study found the same consultants sharply better on one set of tasks and sharply worse on another that looked no harder. And The Skill Code names a cost Autor does not price: the routine tasks machines take are the ones on which novices learn the tacit part. Autor's later work takes the optimistic side, arguing in 2024 that systems extending expert judgment to people with moderate training — the nurse practitioner who diagnoses and prescribes, his own example here — could rebuild the middle rather than hollow it further.

His final point has travelled least. If labour really were made superfluous, the problem would be distribution rather than scarcity, and history offers no example of abundance settling a distributional conflict. He gives the last word to Herbert Simon, writing during the automation anxiety of the 1960s, that "the bogeyman of automation consumes worrying capacity that should be saved for real problems", and says the evidence still favours him.