Pages

Tuesday, September 8, 2015

Why haven't robots taken our jobs? The Complementarity Effect




You’ve probably noticed the trend. The doomsayers are yelling once more. They are telling us that technology poses a threat to human employment — that the robots are coming for our jobs. This is a thesis that has been defended in several academic papers, popular books and newspaper articles. It has been propounded by leading figures in the tech industry, and repeatedly debated and analysed in the media (particularly new media).

But is it right? Last year I presented a lengthy analysis of the pro-technological unemployment from Brynjolfsson and McAfee. Their book, The Second Machine Age, is at the forefront of the current doomsaying trend. In it, they make a relatively simple argument. It starts with the observation that machines are able to displace more and more human labour. It adds to this the claim that while in the past humans have always found other sources of employment, this may no longer be possible because the pace and scope of current technological advance is such that humans may have nowhere left to go.

Recently, Brynjolfsson and McAfee’s thesis has attracted the attention of their economic brethren. Indeed, the Journal of Economic Perspectives has just run a short symposium on the topic. One of the contributors to that symposium was David Autor, who wrote an interesting and sober analysis of the impact of technology on employment entitled ‘Why are there still so many jobs? The history and future of workplace automation’. Autor doesn’t deny the impact of technology on employment, but he doesn’t quite share Brynjolfsson and McAfee’s pessimism.

He makes three main arguments:

The Complementarity Argument: Most doomsaying discussions of technology and work focus on the substitution effect, i.e. the ways in which technology can substitute for labour. In doing so, they frequently ignore the complementarity effect, i.e. the ways in which technology can complement and actually increase the demand for human labour.

The Polarisation Argument: Recent technological advances, particularly in computerisation, have facilitated the polarisation of the labour market. Demand for skilled but routine labour has fallen, while demand for lower skilled personal service work, and highly educated creative work has risen. This has also facilitated rising income inequality.

The Comparative Advantage Argument: The polarisation effect is unlikely to continue much further into the future. Machines will continue to replace routine and codifiable labour, but this will amplify the comparative advantage that humans have in creative, problem-solving labour.

Through these three arguments, we see how Autor’s paints a nuanced picture of the relationship between work and technology. The robots aren’t quite going to take over, but they will have an impact. I want to try explain and assess all three of Autor’s arguments over the next few posts. I start today by delving deeper into the complementarity argument.


1. Autor’s Challenge
Anyone with even a passing interest in the history of workplace automation will be familiar with the Luddites, particularly since the term ‘luddite’ has passed into popular usage. The Luddites were a movement in the early days of the industrial revolution. They were made up of textile workers and weavers. They went about sabotaging machines in textile factories (such as power looms) which they perceived as a threat to their skilled labour. Although their concerns were real, many now look back on the luddites as a naive and fundamentally misconceived movement.

The Luddites feared that machines would rob them of employment, and while that may have been true for them in the short term, it was not indicative of a broader trend. The number of jobs has not dramatically declined in the intervening 200 years. What the Luddites missed was the fact that displacement of humans by labour-saving technologies in one domain could actually increase aggregate demand and open up opportunity for employment in other domains.

Agriculture provides a clear illustration of this phenomenon. There is very clear evidence for a substitution effect in agriculture. As Autor notes:

In 1900, 41 percent of the US workforce was employed in agriculture; by 2000, that share had fallen to 2 percent (Autour 2014), mostly due to a wide range of technologies including automated machinery. 
(Autour 2014, 5)

And yet despite this clear evidence of a substitution effect, we haven’t witnessed a rise in long-term structural unemployment. This despite the fact that other industries have witnessed similar forms of substitution. Autor thinks that this should be puzzling to those like Brynjolfsson and McAfee who think that technology could lead to long-term structural unemployment. This gives rise to something I will call ‘Autor’s Challenge’:

Autor’s Challenge: ‘Given that these technologies demonstrably succeed in their labor saving objective and, moreover, that we invent many more labor-saving technologies all the time, should we not be somewhat surprised that technological change hasn’t already wiped out employment for the vast majority of workers? Why doesn’t automation necessarily reduce aggregate employment, even as it demonstrably reduces labor requirements per unit of output produced?’ 
(Autor 2015, 6)

In other words, before we start harping on about robots stealing our jobs in the future, we should try to explain why they haven’t already stolen our jobs. If we can do this, we might have a better handle on the future trends.


2. The Complementarity Effect
Autor thinks that the explanation lies in the complementarity effect. This effect adds some complexity to our understanding of the relationship between labour and technology. The previously-mentioned substitution effect supposes that the relationship between a human worker and a robot/machine is, in essence, a zero-sum game. Once the machine can do the job better than the human, it takes over and the human loses out. The complementarity effect supposes that the relationship can be more like a positive-sum game, i.e. it might be that as the robot gets better, no one really loses out and everyone gains.

Many jobs are complex. Several different ‘inputs’ (involving different skills and aptitudes) are required to produce the overall economic or social value. Consider the job of a lawyer. He or she must have a good working knowledge of the law, they must be able to use legal research databases, they must be able to craft legal argument, meet with and advise clients, schmooze and socialise with them if needs be, negotiate settlements with other lawyers, manage their time effectively, and so on. Each of these constitutes an ‘input’ that contributes to their overall economic value. They all complement each other: the better you are at all of these things, the more economic value you produce. Now, oftentimes these inputs are subject to specialisation and differentiation within a given law firm. One lawyer will focus on schmoozing, another on negotiation, another on research and case strategy. This specialisation can be a positive sum game (as Adam Smith famously pointed out): the law firm’s productivity can greatly increase despite the specialisation. This is because it is the sum of the parts, not the individual parts, that matters.

This is important when it comes to understanding the impact of technology on labour. To date, most technologies are narrow and specialised. They substitute or replace humans performing routine, specialised tasks. But since the economic value of any particularly work process tends to be produced by a set of complementary inputs, and not just a specialised task, it does not follow that this will lead to less employment for human beings. Instead, humans can switch to the complementary tasks, often benefitting from the efficiency gains associated with machine substitution. Indeed, the lower costs and increased output in one specialised domain can increase labour in other complementary domains.

Autor illustrates the complementarity effect by using the example of ATMs and bank tellers. ATMs were widely introduced to American banking in the 1970s, with the total number increasing from 100,000 to 400,000 in the period from 1995 to 2010 alone. ATMs substitute for human bank tellers in many routine cash-handling tasks. But this has not led to a decrease in bank teller employment. On the contrary, the total number of (human) bank tellers increased from 500,000 to 550,000 between 1980 and 2010. That admittedly represents a fall in percentage share of workforce, but it is still surprising to see the numbers rise given the huge increase in the numbers of ATMs. Why haven’t bank tellers been obliterated?

The answer lies in complementarity. Routine cash-handling tasks are only one part of what provides the economic value. Another significant part is relationship management, i.e. in forging and maintaining relationships with customers, and solving their problems. Humans are good at that part of the job and hence they have switched to fulfilling this role.

Increasingly, banks recognized the value of tellers enabled by information technology, not primarily as checkout clerks, but as salespersons, forging relationships with customers and introducing them to additional bank services like credit cards, loans and investment products. 
(Autor 2015, 7)

Thus, complementarity protected human employment from technological displacement. Indeed, Autor argues that it may even have improved things for these workers as their new roles required higher educational attainment and attracted better pay. The efficiency gains in one domain could consequently facilitate a positive sum outcome.

It is worth summarising Autor’s argument. The following is not formally valid, but captures the gist of the idea:


  • (1) Many work processes draw upon complementary inputs, whereby increases in one input facilitates or requires increases in another, in order to generate economic value.

  • (2) In many cases, technology can substitute for some of these inputs but not all.

  • (3) Humans are often good at fulfilling the complementary, non-substituted roles because those roles rely on hard-to-automate skills.

  • (4) Thus, even in cases of widespread technological substitution, the demand for human labour is not always reduced.


How does this chain of reasoning stack up?



3. Threats to the Complementarity Effect
There is certainly something to it: work processes clearly do rely upon complementary inputs to generate economic value. There is plenty of room for positive sum interactions between humans and robots. But it is not all a bed of roses. Autor himself acknowledges that there are three factors which modulate the scale and beneficial impact of the complementarity effect. They are:

Capacity for complementarity: In order to benefit from the complementarity effect, workers must be able to perform the complementary roles. If workers are only capable of performing the substitutable role, they will not benefit. For instance, it is possible (maybe even likely) that many bank tellers were not good at relationship management. They undoubtedly lost their jobs to ATMs (or so their roles diminished and pay packets cut).

Elasticity of labour supply: Elasticity is an economic concept used to describe how responsive demand or supply is to changes in other phenomena (usually price). Elasticity of labour supply refers to how much the supply of labour increases (or decreases) in response to changes in the price demanded for labour. This modulates complementarity in the following way: Workers capable of fulfilling the complementary roles may not benefit from the increased demand for their labour if it is possible for other workers to flood the market and fulfil complementary tasks. This may have happened with the rise in lower paid personal service workers in the wake of computerisation in the late 20th century. I’ll talk about this more in the next entry.

Output elasticity of demand and income elasticity of demand: This refers to how much demand for a particular product or service increases or decreases in response to increases in productivity and income. In essence, if there is more of a product or service being supplied, and people have more money that they can spend on that product or service, will demand actually go up? The answer varies and this affects the impact of technology on employment. In the case of agricultural produce, demand probably won’t go up. There is only so much food and drink people require each day. This likely explains why the percentage of household income spent on food has steadily declined over the past century despite huge technologically-assisted gains in agricultural productivity. Contrariwise, demand for healthcare has dramatically increased in the same period, despite the fact that this is in an area that has also witnessed huge technologically-assisted gains in productivity. Why? Because people want to be healthier (or avoid disease) and this is a sufficiently fuzzy concept to facilitate increased demand.


This last factor is crucial and provides another part of the response to Autor’s challenge. Part of the reason why there are still so many jobs is that people’s demands don’t remain static over time. On the contrary, their consumption demands usually increase along with increases in income and productivity. Autor provides an arresting illustration of this. He argues that an average US worker living in 2015 could match the standard of living of the average worker in 1915 by simply working for 17 weeks a year. So why do they work for so much longer? Because they’re not satisfied with that standard of living: they’ve tasted the possibility of more and they want it.

Something strikes me about this analysis of technology and employment. The complementarity effect is, no doubt, real. But its ability to sustain demand for human labour in the medium-to-long term seems to depend on one crucial assumption: that technology will remain a narrow, domain-specific phenomenon. That there will always be this complementary space for human workers. But what if we can create general artificial intelligence? What if robot workers are not limited to routine, narrowly-defined tasks? In that case, they could fill the complementary roles too, thereby negating the increased demand for human workers. Indeed, this was one of the central theses of Brynjolfsson and McAfee’s book. They were concerned about the impact of exponential and synergistic technological advances on human employment. They would argue that Autor’s lack of pessimism is driven by a misplaced fealty to historical patterns.

Think about it this way. Suppose there are ten complementary inputs required for a particular work process. A hundred years ago all ten inputs were provided by human workers. Ninety years ago machines were invented that could provide two of these inputs. That was fine: humans could switch to one or more of the remaining eight inputs. Then, fifty years ago, more machines were invented. They could provide two more of the inputs. Humans were limited to the remaining six, but they were happy with this because there was increased demand for those inputs and they paid better. All was good. But then, a few years ago, somebody invented new machines that not only replaced four more of the inputs, but also did a better job than the older machines on the four previously-replaced inputs. Suddenly there were only two places left for human labour to go. But still people were happy because these roles were the most highly skilled and commanded the highest incomes. The complementarity effect continued to hold. Now, fast forward into the future. Suppose somebody invents a general machine learning algorithm that fulfills the final two roles and can be integrated with all the pre-existing machines. A technological apotheosis of sorts has arrived: the technological advances of the past hundred years have all come together and can now completely replace the ten human inputs. People didn’t realise this would happen: they were tricked by the historical pattern. They assumed technology would only replace one or two inputs and that they could fill the complementary space. They neglected both the combined impact of technology, and the possibility of exponential growth.

That was the type of scenario Brynjolfsson and McAfee were warning us about and it seems unaffected by Autor’s claims for the complementarity effect. To link it back to the argument presented in the previous section, it seems like the possibility of general machine intelligence (and/or the synergistic effects of many technological advances) could cast premise (2) into doubt.

To be fair to him, Autor has a response (of sorts) to this. He is sceptical about the prospects for general machine intelligence and the likelihood of machine learning having a significant displacement effect. This features heavily in his defence of the comparative advantage argument. I’ll be looking at that in a future entry.

No comments:

Post a Comment