by Ajay Agrawal
Proposition 1. Single- peaked density functions. If the distribution density
function, f, has a single peak at p = p, then (
/ p)
0 p < p.
Proposition 2 . Common distributions. If the distribution is normal, log-
normal, exponential or uniform, there exists a p*such that for 0 < p < p*, < 1 , and for p* < p, ε > 1 .
These propositions suggest that the model of demand derived from distri-
butions of preferences might be broadly applicable. The second proposition
is suffi
cient to create the inverted- U curve in employment as long as the price
starts above p* and declines below it.
10.2.3 Empirical
Estimates
This very simple model does not consider numerous factors that might
infl uence demand. It does not consider the role of close substitutes or the
eff ect of the business cycle on demand. New technology might create new
products that generate new demand, altering the distribution, or new sub-
stitutes that decrease demand. Global trade might alter downstream indus-
tries, aff ecting the demand for intermediate goods such as cloth or steel.
Nevertheless, the model appears to predict actual demand over a historical
timeframe reasonably well.
Assuming that the preference distribution is lognormal, I estimate the per
capita demand functions for these three commodities (see Bessen 2017 for
details). The model fi ts the data quite closely, realizing R- squareds of .982
or higher. Using these predictions, I obtain very rough estimates of the price
elasticity of demand at each end of the estimation sample (see table 10.1).
The demand was initially highly elastic but became highly inelastic.
Using estimated per capita demand, labor demand can be calculated
incorporating population size, import penetration, labor productivity, and
hours worked. These estimates are shown as the solid lines in fi gure 1. The
estimates appear to be accurate over long periods of time. There are notable
drops in employment during the Great Depression and excess employment
in motor vehicles during World War II. Finally, employment falls below the
Artifi cial Intelligence and Jobs: The Role of Demand 301
Table 10.1
Rough estimates of elasticity of demand
Cotton
Steel
Automotive
Year
Elasticity
Year
Elasticity
Year
Elasticity
1810
2.13
1860
3.49
1910
6.77
1995
0.02
1982
0.16
2007
0.15
estimates when globalization takes a bite out of employment in textiles after
1995, and steel after 1982.
Thus, even though this overly simple model does not account for all of the
factors that aff ect demand, it nevertheless provides a succinct explanation of
the inverted- U in employment in these manufacturing industries.
10.3 Implication
for
AI
10.3.1 The Importance of Demand
Although the model presented here appears to provide a good explana-
tion for how demand mediated the impact of technology in the past, what
is the relevance of this analysis for new technologies? There is, of course, no
guarantee that AI or other new technologies will be applied in markets with
preference distributions similar to those of the textile, steel, and automotive
industries.
The relevance of this history is more general. Specifi cally, the responsive-
ness of demand is key to understanding whether major new technologies
will decrease or increase employment in aff ected industries. Productivity-
enhancing technology will increase industry employment if product demand
is suffi
ciently elastic. If the price elasticity of demand is greater than one,
the increase in demand will more than off set the labor- saving eff ect of the
technology. And demand will likely be suffi
ciently elastic if the technology is
addressing large unmet needs aff ecting people with diverse preferences and
uses for the technology. This situation corresponds to the upper tail of the
distribution function. If, on the other hand, AI is targeted at more satiated
markets, then jobs will be lost in the aff ected industries, although not neces-
sarily in the economy as a whole.
The pace of change of a new technology is not suffi
cient by itself to deter-
mine the impact of that technology on employment. For example, a com-
mon view holds that faster technical change is more likely to eliminate jobs.
Some people argue that because of Moore’s Law, the rate of change will
be fast for AI and this will cause unemployment (Ford 2015). However, my
analysis highlights the importance of demand in mediating the impact of
automation. If demand is suffi
ciently elastic and AI does not completely
302 James Bessen
replace humans, then technical change will create jobs rather than destroy
them. In this case, a faster rate of technical change will actually create faster
employment growth rather than job losses.
The demand response to AI is, of course, an empirical question and,
therefore, an important part of the AI research agenda.
10.3.2 Research
Agenda
To understand the interaction between AI and demand over the next ten
or twenty years, empirical researchers will need answers to several specifi c
questions.
First, to what extent will AI replace humans and to what extent will it,
instead, merely augment human capabilities? That is, to what extent will
AI completely automate occupations and to what extent will it, instead,
merely automate some, but not all, tasks performed by an occupation. If
humans are completely replaced, demand no longer aff ects employment
because there isn’t any demand for humans. In the past, despite extensive
productivity growth, technology has almost always only partially automated
work. Consider what happened to the 271 detailed occupations used in the
1950 census by 2010. Most occupations listed then still exist in some form
(sometimes grouped diff erently) today. Some occupations were eliminated
for a variety of reasons. In many cases, demand for the occupational ser-
vices declined (e.g., boardinghouse keepers); in some cases, demand declined
because of technological obsolescence (e.g., telegraph operators). This, how-
ever, is not the same as automation. In only one case—elevator operators—
can the decline and disappearance of an occupation be largely attributed to
automation. Nevertheless, this sixty- year period witnessed extensive auto-
mation; it was just mostly partial automation.
This same pattern is likely to be true for AI over the next ten or twenty
years for the simple reason that although AI can outperform humans on
some tasks, today’s AI fails miserably at other tasks that humans perform.
A casual review of current developments suggests that over the near term
AI may be able to completely automate some jobs
of drivers and warehouse
workers, but most AI applications are targeted toward automating just some
subset of tasks performed by specifi c occupations. Nevertheless, a more rig-
orous empirical investigation is needed to measure the extent to which AI is
bringing or will bring complete versus partial automation.
To the extent that automation continues to be partial rather than complete
in the near term, demand will be key. This raises a second question: To what
extent will the eff ect of AI on demand and employment during the next ten
or twenty years be similar to the eff ect that AI and computer automation
generally had over the last several decades? Computers have been used to
automate work in activities such as accounting and loan making since the
1950s. The fi rst fully automatic loan application system was installed in
1972. In 1987, an artifi cial intelligence system was fi rst put into commercial
Artifi cial Intelligence and Jobs: The Role of Demand 303
operation in a system used to detect credit fraud. Since then, AI applications
have been used to automate a variety of tasks in other industries and occu-
pations, such as the electronic discovery of legal documents for litigation.
This means that we already have some evidence of the eff ects of AI and
computer automation generally. It does not seem that computer automa-
tion or AI has so far led to signifi cant job losses; the booming market for
electronic discovery applications, for instance, has been associated with an
increase in the employment of paralegals. A few studies have made esti-
mates of the employment impact of computer technology (Gaggl and
Wright 2017; Akerman, Gaarder, and Mogstad 2015), fi nding, if anything,
a modest increase in employment following technology adoption.6 Further
studies could deepen our understanding of the impact of computer automa-
tion on employment, and how this impact diff ers across occupations and
industries.
Also, we need to understand how AI applications in the near future will
diff er from those of the recent past. The model above provides a framework
to analyze this question. In particular, to the extent that the new applications
target the same services and industries as did the computer automation of
the recent past, then we should expect the elasticity of demand to remain
similar over the next ten or twenty years, perhaps with a modest decline.
That is, the elasticity of demand is not likely to change very quickly. On the
other hand, AI might introduce entirely new products and services that tap
into otherwise unmet needs and wants. In this case, there may be new and
unanticipated sources of employment growth. Research can help determine
the extent of change in the sorts of applications, occupations, and industries
aff ected by new AI applications that are also addressed by existing tech-
nologies. To the extent that AI creates wholly new applications, prediction
will be more diffi
cult. Indeed, in the past, predictions about technological
unemployment have reliably failed to anticipate major new applications of
technology and major new sources of demand.
A critical aspect of this research concerns the unevenness of the potential
impact of AI. While AI might not create overall unemployment in the near
future, it will likely eliminate jobs in some occupations while creating new
jobs in others. The need to retrain and transition workers to new occupa-
tions, sometimes in new locations, might be highly disruptive even though
the total employment rate remains high.
Finally, it is important to note that this analytical framework and research
agenda are very much limited to the next ten or twenty years for two rea-
sons. First, beyond a couple of decades, markets might well become satu-
rated. Suppose, for example, that demand is highly elastic for many fi nancial,
health, and other services today so that information technology increases
employment in these markets. If AI rapidly reduces costs or improves the
6. And, importantly, impacts that diff ered across skill groups.
304 James Bessen
quality of these services, the elasticity of demand will decline. That is, these
markets might see the kind of reversals in employment growth seen in fi g-
ure 10.1.
Second, in the future AI might very well be able to completely replace
many more occupations. Then the eff ect of AI on demand will no longer
matter for these occupations. For now, however, understanding how and
where AI aff ects demand is critical to understanding employment eff ects.
Appendix
Propositions
To simplify notation, let the wage remain constant at 1. Then
( p) = p f ( p) ,
1
F ( p)
so that
( p)
= f p + f 2 p + f =
f + f + 1 .
p
1
F
(1
F )2
1
F
f
1
F
p
Note that the second and third terms in parentheses are positive for p >
0; the fi rst term could be positive or negative. A suffi
cient condition for
(∂ε / ∂ p) ≥ 0 is
f
(10A.1)
+ f
0.
f
1
F
Proposition 1 . For a single peaked distribution with mode p, for p < p, f
0 so that (∂ε / ∂ p) ≥ 0.
Proposition 2. For each distribution, I will show that
0, lim = 0, lim = .
p
p
0
p
Taken together, these conditions imply that for suffi
ciently high price, ε > 1,
and for a suffi
ciently low price, ε < 1.
Normal Distribution
p
μ
f ( p) = 1 ( x), F( p) =
( x),
( p) = p
( x)
,
1
(
( x)) , x
where and are the standard normal density and cumulative distribution
functions respectively. Taking the derivative of the density function,
Artifi cial Intelligence and Jobs: The Role of Demand 305
f
+ f = x +
( x)
f
1
F
1
(
( x)).
A well- known inequality for the normal Mills’ ratio (Gordon 1941) holds
that for x > 0,7
( x)
(10A.2)
x
.
1
( x)
Applying this inequality, it is straightforward to show that (10A.1) holds
for the normal distribution. This also implies that lim = . By inspection,
p
ε(0) = 0.
Exponential Distribution
f ( p)
e p , F ( p)
1 e p ,
( p) = p, , p > 0.
Then,
f
+ f =
+ = 0,
f
1
F
so (10A.1) holds. By inspection, ε(0) = 0 and lim = .
p
Uniform Di
stribution
1
p
f ( p)
, F ( p)
,
( p) =
p , 0 < p < b,
b
b
b
p
so that
f
+ f = 1 > 0.
f
1
F
b
p
By inspection, ε(0) = 0 and lim = .
p
b
Lognormal Distribution
1
ln p
μ
f ( p)
( x), F( p)
( x),
( p) = 1
( x)
,
p
1
(
( x)) , x
so that
( p)
x
=
f + f + 1 =
1
+
+ 1 .
p
f
1
F
p
p
p
p (1
)
p
Canceling terms and using Gordon’s inequality, this is positive. And taking
the limit of Gordon’s inequality, lim = . By inspection, lim = 0.
p
p
0
7. I present the inverse of Gordon’s inequality.
306 James Bessen
References
Acemoglu, Daron, and Veronica Guerrieri. 2008. “Capital Deepening and Nonbal-
anced Economic Growth.” Journal of Political Economy 116 (3): 467– 98.
Akerman, Anders, Ingvil Gaarder, and Magne Mogstad. 2015. “The Skill Com-
plementarity of Broadband Internet.” Quarterly Journal of Economics 130 (4):
1781– 824.
Baumol, William J. 1967. “Macroeconomics of Unbalanced Growth: The Anatomy
of Urban Crisis.” American Economic Review 57 (3): 415– 26.
Bessen, James E. 2016. “How Computer Automation Aff ects Occupations: Tech-
nology, Jobs, and Skills.” Law and Economics Research Paper no. 15-49, Boston
University School of Law.
———. 2017. “Automation and Jobs: When Technology Boosts Employment.” Law
and Economics Research Paper no. 17-09, Boston University School of Law.
Boppart, Timo. 2014. “Structural Change and the Kaldor Facts in a Growth Model
with Relative Price Eff ects and Non- Gorman Preferences.” Econometrica 82 (6): 2167– 96.
Buera, Francisco J., and Joseph P. Kaboski. 2009. “Can Traditional Theories of
Structural Change Fit the Data?” Journal of the European Economic Association
7 (2– 3): 469– 77.