MIT INITIATIVE ON THE DIGITAL ECONOMY RESEARCH BRIEF 2018 Vol.

WHAT CAN MACHINEs LEARN AND WHAT dOES

IT MEAN FOR OCCUPATIONS AND THE ECONOMY?
By Erik Brynjolfsson, Tom Mitchell, and Daniel Rock

Concern about automation’s impact on employment is growing

as rapid advances in machine learning (ML), many based on deep
neural networks, are poised to generate significant economic IN THIS BRIEF
value and transform numerous occupations and industries. The
question of whether—or to what extent--machines will replace 1. Machine learning (ML) technologies will grow more
human labor looms large today. pervasive.

Our research suggests that ML technologies will indeed grow 2. J obs are bundles of work tasks. The suitability for
more pervasive, but within job categories, what we define as the machine learning (SML) for work tasks varies greatly.
“suitability for machine learning” (SML) of work tasks varies greatly.
We further propose that our SML rubric, illustrating the variability 3. ML will rarely automate entire jobs. More often, it
in task-level SML, can serve as an indicator for the potential will lead to the reengineering of processes and the
reorganization of a job or an occupation because the set of tasks reorganization of tasks.
that form a job can be separated and re-bundled to redefine the
job. Evaluating worker activities using our rubric, in fact, has the 4. Analysis suggests that ML will affect very different parts
benefit of focusing on what ML can do instead of grouping all of the workforce, including many professional jobs,
forms of automation together. compared with earlier waves of automation.

Debates about the effects of artificial intelligence (AI) on work 5. A shift is needed in the debate about the impact of
should shift away from the common focus on full automation of artificial intelligence on work: Away from the focus on
many jobs and pervasive occupational replacement, and toward full automation of many jobs, and toward the redesign
the redesign of jobs and reengineering of business processes. of jobs and processes.

MACHINE LEARNING AND THE WORKFORCE Adoption of robots, in particular, has been connected to reduced
employment and wages in local labor markets (Acemoglu and
Machine learning, a sub-field of AI, studies the question, “How Restrepo 2017). A recent study by the McKinsey Global Institute
can we build computer programs that automatically improve even suggested that about half of the work activities people
their performance at some task through experience?” Recent perform could be automated with current technology (Manyika et
rapid progress in ML has made it possible for machines to al. 2017). While automation is already having significant effects on
match or surpass humans in certain types of tasks, especially many parts of the workforce and advances in ML are impressive,
those involving image and speech recognition, natural language we remain far from a world of Artificial General Intelligence (AGI)
processing, and predictive analytics. So far, the realized economic that replaces human work across entire occupations.
effects of ML are small relative to its potential. As is common,
there is a time lag of years, or even decades, before technological To delve deeper, we focused on which work tasks within
advances generate substantial economic value: Entrepreneurs occupations will be most affected by ML, and which will be
and innovators take time to adopt new technologies, co-invent relatively unaffected. When considering this question, a key insight
complementary technologies, discover new business processes, must be maintained: An occupation can be viewed as a bundle of
and reconfigure existing work. This is especially true of General- tasks, some of which offer better applications for technology than
Purpose Technologies (GPTs) like AI. GPTs become pervasive, others (Autor, Levy, and Murnane 2003). As with other studies of
improve over time, and generate complementary innovation task automation, the impact of ML on employment is a function
(Bresnahan and Trajtenberg 1995). of SML for specific work activities. We find that ML’s potential will
affect a different set of tasks than earlier technologies for task
automation.
By contrast, most of the recent progress
in ML performance has been made Our research examines the channels by which ML can affect the
by a specific class of algorithms workforce. We apply Brynjolfsson and Mitchell’s (2017) rubric for
evaluating the potential for applying ML to 2,069 work activities,
called deep neural networks, or more 18,156 tasks, and 964 occupations in the O*NET database. From
generally, deep learning systems. this, we build measures of SML for labor inputs in the U.S. economy.
We then discuss measures of the potential for reorganization.

IDE.MIT.EDU 1
MIT INITIATIVE ON THE DIGITAL ECONOMY RESEARCH BRIEF 2018 Vol. 4

WHAT CAN MACHINEs LEARN AND WHAT dOES

IT MEAN FOR OCCUPATIONS AND THE ECONOMY?
By Erik Brynjolfsson, Tom Mitchell, and Daniel Rock

One criterion for task SML is that the

ASSESSING SML
Successful application of ML is contingent on a variety of task
set of inputs and the corresponding set characteristics and contextual factors of work activities. While
of outputs for the task can be measured we find it daunting to imagine all the ways a task could be
sufficiently well that a machine can automated—matching wits with the collective ingenuity of all the
world’s entrepreneurs—the scope of tasks that are SML is much
learn the mapping between the two sets. more constrained and definable. Evaluating worker activities with
our rubric has the benefit of focusing on what ML can do and
In the case of ML, we find that:
avoiding grouping all forms of automation together.
1. Most occupations in most industries have at least some tasks
that are SML.
Suboptimal bundling of SML and non-SML tasks in jobs couldalso
2. Few occupations have all tasks that are highly SML.
prevent specialization and block potential productivity gains from
3. Unleashing ML potential will require significant redesign
ML. For instance, if the cost of ML capital (and SML task wage)
of the task content of jobs, as SML and non-SML tasks within
were zero, workers would prefer to switch to tasks that ML cannot
occupations become unbundled and re-bundled.
do. If firms only offer labor contracts that have a preset mixture
of SML and non-SML tasks, all of the labor effort put toward SML
tasks has an output opportunity cost. In other words, ML could be
MACHINE LEARNING AND TASK AUTOMATION doing those tasks, and the firm could increase profit if it were to
Previous-generation automation has had a significant impact reorganize jobs into new bundles of tasks.
on productivity and the workforce based on explicit rules or
manually written computer algorithms. However, applications One criterion for task SML is that the set of inputs and the
of automation were limited to areas where knowledge could corresponding set of outputs for the task can be measured
be codified as a computer program. Because of Polanyi’s sufficiently well that a machine can learn the mapping between
Paradox—the fact that we humans have “tacit knowledge”; we the two sets.
“know more than we can tell” (Polanyi 1966)— many tasks that
humans know how to do, such as visually inspecting parts, had As noted, we used the O*NET content model for 964
resisted automation because of our inability to codify this skill occupations in the U.S. economy joined to 18,156 specific tasks
in a computer program. at the occupation level, which are further mapped to 2,069 direct
work activities (DWAs) shared across occupations. Applying a
By contrast, most of the recent progress in ML performance rubric of 23 questions, each DWA is scored for its SML by seven
has been made by a specific class of algorithms called deep different people via the crowd-sourcing platform CrowdFlower.
neural networks, or more generally, deep learning systems.1 The rubric is applied to each DWA to generate initial SML scores.
With deep learning systems, ML models circumvent Polanyi’s High values of SML indicate where ML might have the greatest
Paradox by inferring the mapping function between inputs and potential to transform a job.
outputs automatically and analyzing large amounts of sample
data instead of being explicitly programmed. While not always
interpretable or explainable, these ML models open a new
set of possibilities for automation and complementarities to
labor. Software using deep neural nets can be extended to
new domains formerly closed to digitization by the high cost or
impossibility of writing explicit maps of inputs to outputs and
policies.

As a result, the types of tasks affected by ML tomorrow will be

quite different from those affected in past waves of automation.

1The AI Index Report at http://cdn.aiindex.org/2017-report.pdf contains Table 1-Suitability for Machine Learning: Summary Statistics
a series of benchmarks.

IDE.MIT.EDU 2
MIT INITIATIVE ON THE DIGITAL ECONOMY RESEARCH BRIEF 2018 Vol. 4

WHAT CAN MACHINEs LEARN AND WHAT dOES

IT MEAN FOR OCCUPATIONS AND THE ECONOMY?
By Erik Brynjolfsson, Tom Mitchell, and Daniel Rock

Table1 summarizes the SML measures for occupations, tasks, from earlier types of automation and it affects a very different set
and activities from our analysis. Table 2 presents the occupations of tasks.
with the five highest and five lowest values for SML on a scale
Furthermore, we find indicators that the next wave of automation
ranging from 1 to 5. Massage therapists seem fairly immune to
and reengineering may affect a different part of the labor force
machine learning technology, for instance, while concierges may
than the last one: The correlation coefficients with wage and total
be concerned. (Interestingly, the occupation “economist” scores
wage bill percentiles and within-occupation standard deviation of
close to average, with SML of 3.46) The variance of occupation-
SML are 0.17 and 0.002. However, it’s important to note that the ex
level SML is considerably lower than task-level SML.
ante potential of ML may differ from its ultimate implementation,
as other factors come to bear. We might see, for example, large-
scale ML platform companies contracted to automate aspects of
various jobs. The wage and employment effects of these contracts
are ambiguous given possible channels of demand elasticity,
complementary task efforts, and substitutes.

Additionally, although SML correlation with wage and total wage

expenditure percentiles is low, the actual implementation of ML
technologies by managers and integrators may not follow the SML
rankings. If technological change is directed, the implementation
of ML by managers and entrepreneurs will be focused on the high
wage bill tasks with higher SML.

If ML does substitute outright for some occupational tasks, re-

bundling residual tasks in new jobs may transfer risk from the
firm to its workers. This will affect job design, compensation,
and the organization of work. For instance, workers may need
Scale: 5 = maximum SML, 1 = minimum SML to be compensated for taking on bundles of tasks with harder-
to-measure average performance when machines handle
measurable tasks. Thus, over time, worker performance would
Our research reveals a high level of variability for the potential of
become harder to evaluate since the most measurable tasks tend
ML within jobs (within-occupation standard deviation of task SML
to be the most SML.
scores is 0.596). Jobs with higher within-occupation standard
deviation of SML have higher potential for reorganization.
Improvements in technologies have historically been the key driver
of increased industrial productivity. At the same time, they have
There are a number of important conceptual caveats to this
also disrupted employment and the wage structure systematically.
application of the SML rubric. First, the rubric focuses on technical
Our analysis suggests that in this era of technological progress,
feasibility. It is silent on the economic, organizational, legal,
ML will affect very different parts of the workforce than earlier
cultural, and societal factors that can have an important influence
technology waves. Furthermore, tasks within jobs typically show
on ML adoption. Matching the evolving state-of-the-art in ML in
considerable variability in SML, while few (if any) jobs can be fully
the future will require updating the rubric accordingly.2
automated using ML.

Machine learning technology can and will transform many jobs in

IMPLLICATIONS OF SML FOR THE WORKPLACE the economy, but automation of entire jobs will be less significant
It’s likely that ML will not have the same effects as the last waves than the reengineering of processes and the reorganization
of automation, which led to increased inequality and wage of tasks. The focus of researchers, as well as managers and
polarization as routine cognitive tasks were automated (Autor entrepreneurs, should, therefore, be not just on automation, but
and Dorn 2013). The correlation coefficients of SML with wage on job redesign.
percentile and wage bill percentiles, for instance, are very low:
2 Rubric details are available in the Supplementary Materials to Brynjolfsson
−0.14 and 0.10, respectively. ML is a very different technology
and Mitchell (2017).

IDE.MIT.EDU 3
MIT INITIATIVE ON THE DIGITAL ECONOMY RESEARCH BRIEF 2018 Vol. 4

WHAT CAN MACHINEs LEARN AND WHAT dOES

IT MEAN FOR OCCUPATIONS AND THE ECONOMY?
By Erik Brynjolfsson, Tom Mitchell, and Daniel Rock

Additional References ABOUT THE AUTHORS

Acemoglu, Daron, and Pascual Restrepo. 2017. “Robots and Jobs: Erik Brynjolfsson is the Director of the Initiative on the Digital Economy. He
Evidence from US Labor Markets.” Unpublished. also serves as the Schussel Family Professor of Management at MIT Sloan,
and Research Associate at the National Bureau of Economic Research,
Autor, David H., and David Dorn. 2013. “The Growth of Low-Skill Service and is currently a Visiting Scholar at Stanford University. At MIT, he teaches
Jobs and the Polarization of the US Labor Market.” American Economic courses on the Economics of Information and the Analytics Lab. He is the
Review 103(5): 1553–97. author of several books including, with co-author Andrew McAfee, most
recently, Machine, Platform, Crowd (2017), and The Second Machine Age
Autor, David H., Frank Levy, and Richard J. Murnane. 2003. “The Skill (2014)
Content of Recent Technological Change: An Empirical Exploration.”
Quarterly Journal of Economics 118(4): 1279–333. Tom Mitchell is the E. Fredkin University Professor at the Machine Learning
Department, School of Computer Science, at Carnegie Mellon University.
Brynjolfsson, Erik, and Tom Mitchell. 2017. “What Can Machine His research involves the development of new machine learning
Learning Do? Workforce Implications.” Science 358(6370): 1530–34. algorithms, theories and applications, and studies of their societal impacts.
Supplementary materials: http://science.sciencemag.org/content/sci/ Mitchell is a member of the U.S. National Academy of Engineering, and
suppl/2017/12/27/358.6370.1530.DC1/aap8062-Brynjolfsson-SM.pdf the American Academy of Arts and Sciences.

Manyika, James, Michael Chui, Mehdi Miremadi, Jacques Bughin, Katy Daniel Rock is a PhD Candidate at MIT Sloan and a Researcher at the MIT
George, Paul Willmott, and Martin Dewhurst. 2017. A Future That Works: Initiative on the Digital Economy. He studies how firms make and earn
Automation, Employment and Productivity. Chicago: McKinsey Global
returns to investments in technology, and is particularly interested in the
Institute.
economics of Artificial Intelligence.
Polanyi, Michael. 1966. “The Logic of Tacit Inference.” Philosophy 41(155):
1–18

REPORT
The full paper can be found here: http://ide.mit.edu/sites/default/
files/publications/pandp.20181019.pdf

MIT INITIATIVE ON THE DIGITAL ECONOMY SUPPORT THE MIT IDE

The MIT IDE is solely focused on the digital The generous support of individuals, foundations, and
economy. We conduct groundbreaking research, corporations are critical to the success of the IDE. Their
convene the brightest minds, promote dialogue, contributions fuel cutting-edge research by MIT faculty
expand knowledge and awareness, and implement and graduate students, and enables new faculty hiring,
solutions that provide critical, actionable insight for curriculum development, events, and fellowships.
people, businesses, and government. We are solving Contact Christie Ko (cko@mit.edu) to learn how you or
the most pressing issues of the second machine age, your organization can support the IDE.
such as defining the future of work in this time of
unprecendented disruptive digital transformation.

TO LEARN MORE ABOUT THE IDE, INCLUDING

UPCOMING EVENTS, VISIT IDE.MIT.EDU

IDE.MIT.EDU 4