The Future of Employment (Part IV, Final)

The Future of Employment (Part IV)

How Susceptible are Jobs to Computerisation

by Carl Benedikt Frey and Michael A Osborne (September 17 2013)

V. Employment in the Twenty-First Century

In this section, we examine the possible future extent of at-risk job computerisation, and related labour market outcomes. The task model predicts that recent developments in Machine Learning (“ML”) will reduce aggregate demand for labour input in tasks that can be routinised by means of pattern recognition, while increasing the demand for labour performing tasks that are not susceptible to computerisation.

However, we make no attempt to forecast future changes in the occupational composition of the labour market. While the 2010~2020 US Bureau of Labour Statistics (“BLS”) occupational employment projections predict US net employment growth across major occupations, based on historical staffing patterns, we speculate about technology that is in only the early stages of development. This means that historical data on the impact of the technological developments we observe is unavailable. {21} We therefore focus on the impact of computerisation on the mix of jobs that existed in 2010. Our analysis is thus limited to the substitution effect of future computerisation.

Turning first to the expected employment impact, reported in Figure III, we distinguish between high, medium and low risk occupations, depending on their probability of computerisation (thresholding at probabilities of 0.7 and 0.3). According to our estimate, 47 percent of total US employment is in the high risk category, meaning that associated occupations are potentially automatable over some unspecified number of years, perhaps a decade or two. It shall be noted that the probability axis can be seen as a rough timeline, where high probability occupations are likely to be substituted by computer capital relatively soon. Over the next decades, the extent of computerisation will be determined by the pace at which the above described engineering bottlenecks to automation can be overcome. Seen from this perspective, our findings could be interpreted as two waves of computerisation, separated by a “technological plateau”. In the first wave, we find that most workers in transpo rtation and logistics occupations, together with the bulk of office and administrative support workers, and labour in production occupations, are likely to be substituted by computer capital. As computerised cars are already being developed and the declining cost of sensors makes augmenting vehicles with advanced sensors increasingly cost-effective, the automation of transportation and logistics occupations is in line with the technological developments documented in the literature. Furthermore, algorithms for big data are already rapidly entering domains reliant upon storing or accessing information, making it equally intuitive that office and administrative support occupations will be subject to computerisation. The computerisation of production occupations simply suggests a continuation of a trend that has been observed over the past decades, with industrial robots taking on the routine tasks of most operatives in manufacturing. As industrial robots are becoming more advanced, wit h enhanced senses and dexterity, they will be able to perform a wider scope of non-routine manual tasks. From a technological capabilities point of view, the vast remainder of employment in production occupations is thus likely to diminish over the next decades.

More surprising, at first sight, is that a substantial share of employment in services, sales and construction occupations exhibit high probabilities of computerisation. Yet these findings are largely in line with recent documented technological developments. First, the market for personal and household service robots is already growing by about twenty percent annually (MGI, 2013). As the comparative advantage of human labour in tasks involving mobility and dexterity will diminish over time, the pace of labour substitution in service occupations is likely to increase even further. Second, while it seems counterintuitive that sales occupations, which are likely to require a high degree of social intelligence, will be subject to a wave of computerisation in the near future, high risk sales occupations include, for example, cashiers, counter and rental clerks, and telemarketers. Although these occupations involve interactive tasks, they do not necessarily require a high degree of social intelligence. Our model thus seems to do well in distinguishing between individual occupations within occupational categories. Third, prefabrication will allow a growing share of construction work to be performed under controlled conditions in factories, which partly eliminates task variability. This trend is likely to drive the computerisation of construction work.

In short, our findings suggest that recent developments in ML will put a substantial share of employment, across a wide range of occupations, at risk in the near future. According to our estimates, however, this wave of automation will be followed by a subsequent slowdown in computers for labour substitution, due to persisting inhibiting engineering bottlenecks to computerisation. The relatively slow pace of computerisation across the medium risk category of employment can thus partly be interpreted as a technological plateau, with incremental technological improvements successively enabling further labour substitution. More specifically, the computerisation of occupations in the medium risk category will mainly depend on perception and manipulation challenges. This is evident from Table III, showing that the “manual dexterity”, “finger dexterity” and “cramped work space” variables exhibit relatively high values in the medium risk category. Indeed, even with recent technological developments, allowing for more sophisticated pattern recognition, human labour will still have a comparative advantage in tasks requiring more complex perception and manipulation. Yet with incremental technological improvements, the comparative advantage of human labour in perception and manipulation tasks could eventually diminish. This will require innovative task restructuring, improvements in ML approaches to perception challenges, and progress in robotic dexterity to overcome manipulation problems related to variation between task iterations and the handling of irregular objects. The gradual computerisation of installation, maintenance, and repair occupations, which are largely confined to the medium risk category, and require a high degree of perception and manipulation capabilities, is a manifestation of this observation.

Our model predicts that the second wave of computerisation will mainly depend on overcoming the engineering bottlenecks related to creative and social intelligence. As reported in Table III, the “fine arts”, “originality”, “negotiation”, “persuasion”, “social perceptiveness”, and “assisting and caring for others”, variables, all exhibit relatively high values in the low risk category. By contrast, we note that the “manual dexterity”, “finger dexterity” and “cramped work space” variables take relatively low values. Hence, in short, generalist occupations requiring knowledge of human heuristics, and specialist occupations involving the development of novel ideas and artifacts, are the least susceptible to computerisation. As a prototypical example of generalist work requiring a high degree of social intelligence, consider the ONET tasks reported for chief executives, involving “conferring with board members, organization officials, or staff members to discuss issues, coordinate activities, or resolve problems”, and “negotiating or approving contracts or agreements”. Our predictions are thus intuitive in that most management, business, and finance occupations, which are intensive in generalist tasks requiring social intelligence, are largely confined to the low risk category. The same is true of most occupations in education, healthcare, as well as arts and media jobs. The ONET tasks of actors, for example, involve “performing humorous and serious interpretations of emotions, actions, and situations, using body movements, facial expressions, and gestures”, and “learning about characters in scripts and their relationships to each other in order to develop role interpretations”. While these tasks are very different from those of a chief executive, they equally require profound knowledge of human heuristics, implying that a wide range of tasks, involving social intelligence, are unlikely to become subject to computerisation in the near future.

The low susceptibility of engineering and science occupations to computerisation, on the other hand, is largely due to the high degree of creative intelligence they require. The ONET tasks of mathematicians, for example, involve “developing new principles and new relationships between existing mathematical principles to advance mathematical science” and “conducting research to extend mathematical knowledge in traditional areas, such as algebra, geometry, probability, and logic”. Hence, while it is evident that computers are entering the domains of science and engineering, our predictions implicitly suggest strong complementarities between computers and labour in creative science and engineering occupations; although it is possible that computers will fully substitute for workers in these occupations over the long-run. We note that the predictions of our model are strikingly in line with the technological trends we observe in the automation of knowledge work, even within occupational categories. For example, we find that paralegals and legal assistants – for which computers already substitute – in the high risk category. At the same time, lawyers, which rely on labour input from legal assistants, are in the low risk category. Thus, for the work of lawyers to be fully automated, engineering bottlenecks to creative and social intelligence will need to be overcome, implying that the computerisation of legal research will complement the work of lawyers in the medium term.

To complete the picture of what recent technological progress is likely to mean for the future of employment, we plot the average median wage of occupations by their probability of computerisation. We do the same for skill level, measured by the fraction of workers having obtained a bachelor’s degree, or higher educational attainment, within each occupation. Figure IV reveals that both wages and educational attainment exhibit a strong negative relationship with the probability of computerisation. We note that this prediction implies a truncation in the current trend towards labour market polarization, with growing employment in high and low-wage occupations, accompanied by a hollowing-out of middle-income jobs. Rather than reducing the demand for middle-income occupations, which has been the pattern over the past decades, our model predicts that computerisation will mainly substitute for low-skill and low-wage jobs in the near future. By contrast, high-skill and high-wage occupat ions are the least susceptible to computer capital.

Our findings were robust to the choice of the seventy occupations that formed our training data. This was confirmed by the experimental results tabulated in Table II: a GP classifier trained on half of the training data was demonstrably able to accurately predict the labels of the other half, over one hundred different partitions. That these predictions are accurate for many possible partitions of the training set suggests that slight modifications to this set are unlikely to lead to substantially different results on the entire dataset.

V.A. Limitations

It shall be noted that our predictions are based on expanding the premises about the tasks that computer-controlled equipment can be expected to perform. Hence, we focus on estimating the share of employment that can potentially be substituted by computer capital, from a technological capabilities point of view, over some unspecified number of years. We make no attempt to estimate how many jobs will actually be automated. The actual extent and pace of computerisation will depend on several additional factors which were left unaccounted for.

First, labour saving inventions may only be adopted if the access to cheap labour is scarce or prices of capital are relatively high (Habakkuk, 1962). {22} We do not account for future wage levels, capital prices or labour shortages. While these factors will impact on the timeline of our predictions, labour is the scarce factor, implying that in the long-run wage levels will increase relative to capital prices, making computerisation increasingly profitable (see, for example, Acemoglu, 2003).

Second, regulatory concerns and political activism may slow down the process of computerisation. The states of California and Nevada are, for example, currently in the process of making legislatory changes to allow for driverless cars. Similar steps will be needed in other states, and in relation to various technologies. The extent and pace of legislatory implementation can furthermore be related to the public acceptance of technological progress. {23} Although resistance to technological progress has become seemingly less common since the Industrial Revolution, there are recent examples of resistance to technological change. {24} We avoid making predictions about the legislatory process and the public acceptance of technological progress, and thus the pace of computerisation.

Third, making predictions about technological progress is notoriously difficult (Armstrong and Sotala, 2012). {25} For this reason, we focus on near-term technological breakthroughs in ML and MR , and avoid making any predictions about the number of years it may take to overcome various engineering bottlenecks to computerisation. Finally, we emphasise that since our probability estimates describe the likelihood of an occupation being fully automated, we do not capture any within-occupation variation resulting from the computerisation of tasks that simply free-up time for human labour to perform other tasks.

Although it is clear that the impact of productivity gains on employment will vary across occupations and industries, we make no attempt to examine such effects.

VI. Conclusions

While computerisation has been historically confined to routine tasks involving explicit rule-based activities (Autor, et al, 2003; Goos, et al, 2009; Autor and Dorn, 2013), algorithms for big data are now rapidly entering domains reliant upon pattern recognition and can readily substitute for labour in a wide range of non-routine cognitive tasks (Brynjolfsson and McAfee, 2011; MGI, 2013). In addition, advanced robots are gaining enhanced senses and dexterity, allowing them to perform a broader scope of manual tasks (IFR, 2012b; Robotics-VO, 2013; MGI, 2013). This is likely to change the nature of work across industries and occupations.

In this paper, we ask the question: how susceptible are current jobs to these technological developments? To assess this, we implement a novel methodology to estimate the probability of computerisation for 702 detailed occupations. Based on these estimates, we examine expected impacts of future computerisation on labour market outcomes, with the primary objective of analysing the number of jobs at risk and the relationship between an occupation’s probability of computerisation, wages and educational attainment.

We distinguish between high, medium and low risk occupations, depending on their probability of computerisation. We make no attempt to estimate the number of jobs that will actually be automated, and focus on potential job automatability over some unspecified number of years. According to our estimates around 47 percent of total US employment is in the high risk category. We refer to these as jobs at risk – that is, jobs we expect could be automated relatively soon, perhaps over the next decade or two.

Our model predicts that most workers in transportation and logistics occupations, together with the bulk of office and administrative support workers, and labour in production occupations, are at risk. These findings are consistent with recent technological developments documented in the literature. More surprisingly, we find that a substantial share of employment in service occupations, where most US job growth has occurred over the past decades (Autor and Dorn, 2013), are highly susceptible to computerisation. Additional support for this finding is provided by the recent growth in the market for service robots (MGI, 2013) and the gradually diminishment of the comparative advantage of human labour in tasks involving mobility and dexterity (Robotics-VO, 2013).

Finally, we provide evidence that wages and educational attainment exhibit a strong negative relationship with the probability of computerisation. We note that this finding implies a discontinuity between the nineteenth, twentieth and the twenty-first century, in the impact of capital deepening on the relative demand for skilled labour. While nineteenth century manufacturing technologies largely substituted for skilled labour through the simplification of tasks (Braverman, 1974; Hounshell, 1985; James and Skinner, 1985; Goldin and Katz, 1998), the Computer Revolution of the twentieth century caused a hollowing-out of middle-income jobs (Goos, et al, 2009; Autor and Dorn, 2013). Our model predicts a truncation in the current trend towards labour market polarisation, with computerisation being principally confined to low-skill and low-wage occupations. Our findings thus imply that as technology races ahead, low-skill workers will reallocate to tasks that are non-susceptible to computer isation – that is, tasks requiring creative and social intelligence. For workers to win the race, however, they will have to acquire creative and social skills.


{21} It shall be noted that the BLS projections are based on what can be referred to as changes in normal technological progress, and not on any breakthrough technologies that may be seen as conjectural.

{22} For example, case study evidence suggests that mechanisation in eighteenth century cotton production initially only occurred in Britain because wage levels were much higher relative to prices of capital than in other countries (Allen, 2009b). In addition, recent empirical research reveals a causal relationship between the access to cheap labour and mechanisation in agricultural production, in terms of sustained economic transition towards increased mechanisation in areas characterised by low-wage worker out-migration (Hornbeck and Naidu, 2013).

{23} For instance, William Huskisson, former cabinet minister and Member of Parliament for Liverpool, was killed by a steam locomotive during the opening of the Liverpool and Manchester Railway. Nonetheless, this well-publicised incident did anything but dissuade the public from railway transportation technology. By contrast, airship technology is widely recognised as having been popularly abandoned as a consequence of the reporting of the Hindenburg disaster.

{24} Uber, a start-up company connecting passengers with drivers of luxury vehicles, has recently faced pressure from local regulators, arising from tensions with taxicab services. Furthermore, in 2011 the UK Government scrapped a 12.7 billion GBP project to introduce electronic patient records after resistance from doctors.

{25} Marvin Minsky famously claimed in 1970 that “in from three to eight years we will have a machine with the general intelligence of an average human being”. This prediction is yet to materialise.

References: See URL below.

Posted in Uncategorized

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s

%d bloggers like this: