Economic growth under

transformative AI
A guide to the vast range of possibilities for output growth,
wages, and the labor share

Philip Trammell and Anton Korinek

Global Priorities Institute | October 2020

GPI Working Paper No. 8-2020

 Economic Growth under Transformative AI:
A guide to the vast range of possibilities
for output growth, wages, and the labor share∗

Philip Trammell† and Anton Korinek‡

February 3, 2023

∗
Thanks to Tim Blomfield and Nicholas Emery for contributions to the literature sum-
maries and organization.
†
Global Priorities Institute and Department of Economics, University of Oxford.
Contact: [email protected].
‡
Department of Economics and Darden School of Business, University of Virginia,
NBER and CEPR.
 1

Contents
1 Introduction 2

2 Economics background 5
2.1 Production and factor shares . . . . . . . . . . . . . . . . . . . 5
2.2 Substitution . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.3 Exogenous growth . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.4 (Semi-)endogenous growth . . . . . . . . . . . . . . . . . . . . 11

3 AI in basic models of good production 14

3.1 Capital productivity in isolation . . . . . . . . . . . . . . . . . 14
3.2 Imperfect substitution . . . . . . . . . . . . . . . . . . . . . . 14
3.3 Perfect substitution . . . . . . . . . . . . . . . . . . . . . . . . 18
3.4 Substitutability in robotics production . . . . . . . . . . . . . 22
3.5 Growth impacts via impacts on saving . . . . . . . . . . . . . 25

4 AI in task-based models of good production 29

4.1 Introducing the task-based framework . . . . . . . . . . . . . . 29
4.2 Task automation . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.3 Task creation . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4.4 Task replacement . . . . . . . . . . . . . . . . . . . . . . . . . 35

5 AI in technology production 38
5.1 Learning by doing . . . . . . . . . . . . . . . . . . . . . . . . . 39
5.2 Automated research . . . . . . . . . . . . . . . . . . . . . . . . 40
5.3 AI assistance in research . . . . . . . . . . . . . . . . . . . . . 45
5.4 Growth impacts via impacts on technology investment . . . . 47

6 AI in both good and technology production 48

7 Overview of the possibilities 51

8 Conclusion 55

A Proofs 59
 2

1 Introduction
At least since Herbert Simon’s 1960 prediction that artificial intelligence
would soon replace all human labor, many economists have understood that
there is a possibility that sooner or later artificial intelligence (AI) will dra-
matically transform the global economy. AI could have a transformative
impact on a wide variety of domains; indeed, it could transform market struc-
ture, the value of education, the geopolitical balance of power, and practically
anything else.
We will focus on three of the clearest and best-studied classes of potential
transformations in economics: the potential impacts on output growth, on
wage growth, and on the labor share, i.e. the share of output paid as wages. On
all counts we will focus on long-run impacts rather than transition dynamics.
Instead of attempting to predict the future, our focus will be on surveying
the vast range of possibilities identified in the economics literature.
Standard growth models imply that the potential impact of AI on the
growth rate of output could take the form of
• a decrease to the growth rate, even perhaps rendering it negative;
• a permanent increase to the growth rate, as the Industrial Revolution
increased the global growth rate from near zero to something over two
percent per year;
• a continuous acceleration in growth, with the growth rate growing
unboundedly as time tends to infinity (what, following Aghion et al.
(2019), we will call a “Type I growth explosion”); or even
• an acceleration in the growth rate rapid enough to produce infinite
output in finite time (a “Type II growth explosion”).1
Basic physics suggests that the last of these scenarios is impossible, of
course—as is eternal exponential or super-exponential growth. The relevant
possibility is that AI induces a growth path that resembles these benchmarks
for some time, until production confronts limiting factors, such as land or
energy, that have never or have not recently been important constraints on
growth. We will thus also consider how increases in growth may eventually
be choked off by such limiting factors.
1
Type II growth explosions, as we will call them, are the literal mathematical singular-
ities after which views of rapid AI-induced growth are sometimes called “singularitarian”.
The term “singularity” is used in several different ways throughout the literature covered
here, however, so we will generally avoid it—except in the context of the “singularity
condition” of Mookherjee and Ray (2017) (§3.4).
 3

The potential impact of AI on the labor market includes predictions that

• wages fall;
• wages rise, but less quickly than output, producing a declining labor
share;
• wages rise in line with output, producing a constant labor share; or
even
• wages rise more quickly than output, producing a rising labor share.
In the respective limiting cases, AI could result in a future in which wages
are (literally or asymptotically) zero or near zero; very high in absolute
terms but approximately zero percent of total output; or high both in
absolute and in relative terms.

As this discussion illustrates, the space of possibilities is vast. At the same

time, as Simon’s failed prediction testifies, transformational impacts from AI
are by no means certain to transpire on any particular time horizon. Most
studies to date of the economics of AI, therefore, have focused on the most
immediate, moderate, and likely impacts of AI.2 These include marginal
shifts in output and factor shares; impacts on marketing and statistical dis-
crimination; impacts on regional inequality; and industry-specific forecasts
of AI-induced growth and labor displacement over the next few decades.
Empirical estimates of the future economic importance of AI have drawn
inferences from foreseeable industry-specific applications of AI, from totals
spent on AI R&D, and from comparisons between AI and past technologi-
cal developments in computing, internet connectivity, or related fields (see
e.g. Chen et al. (2016)). Industrial and R&D-based inferences may however
severely underestimate the field’s transformativeness if technological devel-
opment has substantial external effects, as is often assumed; the importance
of the atomic bomb is not well approximated by the cost of the Manhattan
project. Even in the absence of externalities, furthermore, inferences from
2
Summaries of the economic implications of AI have been put out by most major
consulting firms and by several governments and academic institutes. A proper review of
these reviews would require a document in its own right, but for comparison, the review
that appears to go furthest in discussing AI’s transformative possibilities is that from
Accenture (Purdy and Daugherty, 2016). The most radical scenario the authors consider
is one in which AI comes to serve as a “new factor of production” complementing both
labor and capital. They forecast that, in this scenario, the result will be what they call
“a transformative effect on growth”, by which they mean a doubling of growth rates in
developed countries up to 2035.
 4

R&D expenditures implicitly discount future output according to the time

preference of the research funders, who may assign negligible value to the
impact of their technologies on the distant future. And common reference-
class-based projections preclude the possibility that AI ultimately proves to
be truly transformational, less like the economic impact of broadband than
like that of the Industrial Revolution or the evolution of the human species
itself.
In recent years, economists have begun to engage earnestly in formal
theoretical explorations of a wide array of the transformative possibilities
of AI, including those outlined above. We aim to summarize the findings of
these explorations.3 In the process, we hope not only to state the conclusions
of various models, but also to give the reader some mathematical intuition
for the most important mechanisms at play. (Indeed we have altered some of
the models slightly, to clarify their implications regarding transformativeness
and their relationships to the other models discussed here.)
This document is intended for anyone comfortable with a moderate
amount of mathematical notation and interested in understanding the chan-
nels through which AI could have a transformative impact on wages and
growth. Readers with backgrounds in economics will hopefully come to bet-
ter understand the possibilities which concern singularitarians, and readers
with singularitarian backgrounds will hopefully come to better understand
the relevant tools and insights of economics.
The rest of this document proceeds as follows. §2 consists of an overview
of the economics that will be relevant for understanding the subsequent sec-
tions. §3 discusses models in which AI is added to a standard production
function. §4 discusses models in which AI is added to a “task-based” produc-
tion function. §3 and §4 both implicitly take place in a setting of exogenous
productivity growth; §5 discusses models in which productivity growth is at
least partially endogenous and AI can feature in its production. §7 compares
the results found in §3-5. Finally, §8 concludes.
3
Sandberg (2013) presented an “overview of models of technological singularity” before
the past decade of economist engagement with AI and its transformative potential. Most of
the models he summarizes therefore do not attempt to spell out how artificial intelligence,
or indeed any particular transformative technology, would interact with standard economic
models to produce the results in question. The models summarized here fill this gap.
 5

2 Economics background
Literature on the economics of AI extends prior literature on the economics
of production and growth. Those without backgrounds in economics may
therefore find it helpful to review the latter briefly. Those with backgrounds
in economics may want to skim the section to familiarize themselves with the
notation and terminology we introduce.

2.1 Production and factor shares

Throughout this review, we say that output at a time t is determined by a
production function F (·) and the list of input (or “factor”) quantities avail-
able at t. If for simplicity we categorize all production factors as either labor
L or capital K, and think of all output as a single good Y , we can write
Yt = F (Kt , Lt ). We will temporarily abandon the time subscripts and ex-
plore production in a static setting; they will return when we explore growth.
A production function F (·) will be assumed to be continuously differen-
tiable, increasing, and concave in each argument. It is also assumed to exhibit
constant returns to scale (CRS); doubling the earth, with all its labor and all
its capital, would presumably double output. Finally, its inputs are assumed
to be complements: the marginal productivity of each is increasing in the
supply of the other.4
There can be “capital-augmenting” technology, denoted A, and “labor-
augmenting” technology, denoted B. Increases in some factor-augmenting
technology make the use of the given factor more efficient, so that we can
proceed as if we had more of it. With technology, that is, our production
function takes the form Y = F (AK, BL). Note that, by CRS, a technological
advance that multiplies both A and B by a given factor will multiply output
by this factor as well.
The “marginal product” of a factor is the derivative of output with re-
spect to that factor. Both inputs are assumed to be paid their marginal
products. That is, the wage rate is FL (AK, BL), and the capital rental rate
is FK (AK, BL). This makes sense if we imagine that production is taking
place in a competitive market, with lots of identical firms facing this produc-
tion function. If a factor employed at one firm is not being paid its marginal
product, another firm will offer to pay more for it.
4
More formally, it is assumed that FLK > 0, with subscripts here denoting partial
derivatives (and that FKL = FLK > 0, by Young’s Theorem).
 6

These factor payments will equal total output—i.e.

KFK (AK, BL) + LFL (AL, BL) = F (AK, BL) (1)

—whenever F (·) is CRS, by Euler’s Homogeneous Function Theorem. If

output exceeded the sum of factor payments, we would have to explain what
was done with the output not paid as wages or rents; and if factor payments
exceeded output, we would have to explain how the deficit was filled. The
fraction of output paid out as wages—i.e.

LFL (AK, BL)/F (AK, BL) (2)

—is termed the “labor share”. Likewise, the “capital share” is the fraction
of output that accrues to the owners of capital.
We will assume through most this document that everyone is employed.
In reality, of course, many cannot work, or will not work if the wage rate is
sufficiently low. Model scenarios in which the wage rate falls dramatically
may therefore be more accurately interpreted as scenarios in which unem-
ployment is widespread. For our purposes this will not be an important
distinction.

2.2 Substitution
Given a production (or utility) function and a list of factor (or consumption
good) prices, suppose a purchaser spends a fixed budget so as to maximize
output (or utility). Suppose furthermore that the production plans (or pref-
erences) are homothetic: that is, that the production (utility) function is
CRS, or a monotonic transformation of one that is CRS. The elasticity of
substitution for some factor i is then, intuitively, the value ϵ such that, if the
relative price of i falls by a small proportion (say, 1%), the relative quantity
of i purchased—Xi /Xj , for j ̸= i—will rise by an ϵ-times larger proportion
(ϵ%).5
Conversely, then, given a list of factor quantities, the elasticity of substi-
tution is the value ϵ such that, in order for the relative quantity of i sold to
5
Note that, absent homotheticity, elasticity of substitution is not defined: a 1% rise in
the price of i and a 1% fall in the price of other items may have different impacts on the
relative quantities purchased, even though they would produce the same impact on the
relative price of i.
 7

increase by a small proportion (say 1%), the relative price of i must fall by
a (1/ϵ)-times larger proportion ((1/ϵ)%).
Suppose that goods 1 and 2 are divided among many owners, and suppose
there are many interested purchasers. (The sets of owners and potential
purchasers need not be disjoint, but it may be easier to imagine that they
are.) The price of 1 should then equal the market-clearing price: the price
such that, given the price of 2, precisely the entire supply of 1 is purchased.
If the price of 1 were higher than this, anyone who owns some of it would
do well to sell her unsold goods at just less than the market price. If it were
lower, i.e. if purchasers would have been willing to buy more than the entire
supply of 1 at the given price, then anyone who owns some 1 could charge
more than the given price and still sell all she has.
Now consider the consequences of exogenously increasing the relative
quantity of good 1 by 1%, say by increasing its absolute quantity by 1%
across each of its owners and leaving the quantity of good 2 unchanged. The
market-clearing relative price of good 1 will then fall by (1/ϵ)%. As we can
see, marginally increasing the relative abundance of a good results in smaller
relative expenditure on that good—i.e. its owners receive a smaller share of
total income—precisely when the elasticity of substitution between it and
other goods, on the current margin, is less than 1.
For illustration: food and other goods are not very substitutable. When
food was much scarcer, its owners were able to command such higher prices
for it that people spent larger shares of their incomes on it. On the other
hand, industrially produced goods and handmade goods are very substi-
tutable. As the former grew more plentiful, following the Industrial Revolu-
tion’s explosion in manufacturing, people spent larger shares of their incomes
on them.
Goods are perfect complements if ϵ = 0. In this case, output (or utility)
is some constant times the minimum of the goods’ quantities, in some ratio.
(Consider left shoes and right shoes, in the 1:1 ratio, or bicycle frames and
wheels, in the 1:2 ratio.) However the goods’ relative prices change, the
relative quantities purchased will stay fixed at the given ratio.
The case of perfect substitutability is approached in the limit as ϵ → ∞.
In this case a positive quantity of each good is only purchased if their prices
are equal; if their prices differ, only the cheaper is purchased.

For ease of notation, let us define the “substitution parameter” ρ ≜ (ϵ − 1)/ϵ.

Note that the cases ϵ < 1, ϵ = 1, and ϵ > 1 correspond to ρ < 0, ρ = 0, and
 8

ρ > 0 respectively.
A production function exhibits constant elasticity of substitution (CES)
if its elasticity of substitution does not depend on the factor prices and quan-
tities. A two-factor CES production function that is also CRS, with factor-
augmenting technology, must take the form

Y = [(AK)ρ + (BL)ρ ]1/ρ (3)

if ρ ̸= 0 and
Y = (AK)a (BL)1−a , (4)
for some a ∈ (0, 1), if ρ = 0.6 In the second case, the function is called
“Cobb-Douglas”. When ρ ≤ 0, output requires strictly positive quantities of
both factors.
If we have more than two factors i, natural extensions of the above func-
tions look similar, with the Cobb-Douglas exponents ai still summing to 1.
When ρ ̸= 0, the share of output paid to factor X, with factor-augmenting
technology C, equals (CX/Y )ρ . When ρ = 0, a factor’s share equals the
exponent on that factor. In general, the share of factor X is decreasing in
CX/Y when ρ < 0, independent of CX/Y when ρ = 0, and increasing in
CX/Y when ρ > 0.
Observe that there is a deep similarity between cases (a) in which there
is a single consumption good, but multiple factors to its production, and (b)
in which consumer utility is defined over multiple consumption goods, each
of which employs a single production factor. Consumers in the latter cases
function like factories in the former cases, as if consumer goods were inputs
to the production of utility.

2.3 Exogenous growth

In practice, of course, the substitution parameter between labor and capital
may not be constant. It may be high when labor is more abundant than
capital and low otherwise, or vice-versa, for example. It may also change
over time, for reasons independent of factor quantities. That said, it has
been estimated in a variety of contexts and times to be substantially neg-
ative (see e.g. Oberfield and Raval (2014) or Chirinko and Mallick (2017)).
6
If it is not also CRS, it must be a monotonic transformation of the above. Note that,
either way, CES implies homotheticity.
 9

When exploring preliminary hypotheses about growth, factor shares, and

other macroeconomic variables, therefore, it can be helpful to start with a
model in which production is CES (and CRS) with ρ < 0.
Output per person in the developed world has grown substantially over
the past few centuries, following a roughly exponential trajectory. In a two-
factor production function without technology growth, the only possible ex-
planation for this would be the capital stock growing more quickly than the
population. Capital accumulation cannot be the primary force driving long-
run growth, however, for at least two reasons.
First: if ρ = 0, capital accumulation can produce unbounded output,
albeit at a growth rate that slows to zero; and if ρ > 0 (or if production
is not CES but the substitution parameter is permanently bounded above
zero), capital accumulation can in principle sustain a positive growth rate.7
But if ρ < 0 (constant or bounded below), as the stock of capital per unit
of labor grows, capital’s marginal productivity falls to the point that output
growth slows to a halt. That is, a lack of labor per unit of capital definitively
constrains the growth of output per person. As we can see from the first
equation above, when ρ < 0, the capital term tends to zero as the quantity
of capital increases. In the limit, then, if capital is far more plentiful than
labor, output tends to BL, and output per person tends to B.
Second, historically, the capital share throughout the developed world has
been roughly constant (at about 1/3). As noted in §2.2, however, if ρ < 0,
an unboundedly increasing stock of capital per unit of labor should decrease
the capital share to zero.
Capital-augmenting technology growth just increases the effective capital
stock, so if ρ < 0 it cannot produce long-run output growth either, for the
same reasons.8

As long as (effective) capital is accumulating, the way to get long-run per-

7
After accounting for capital depreciation, ρ may have to be strictly positive for capital
accumulation to allow long-run growth. We will ignore capital depreciation throughout
most of this document for simplicity.
8
More generally, as should now be intuitive, when highly complementary production
factors undergo different rates of accumulation or productivity growth, output’s growth
rate converges to that of the slowest-growing factor and its share goes converges to 1.
Likewise with respect to complementary consumption goods, each requiring a different
input. This is sometimes known as the “Baumol condition”, after Baumol’s (e.g. Baumol
(1967)) seminal analyses of the increasing share of output spent on low-productivity-
growth sectors, such as live entertainment.
 10

capita output growth in this framework, as shown by Uzawa (1961), is to

introduce labor-augmenting technology growth. To illustrate this, suppose
for simplicity that A is fixed, that B grows at some constant exponential
rate gB , and that a constant proportion s of output is saved as capital each
period.9 If s is high enough that capital accumulation can keep up with the
growing effective labor force, and if the labor supply is also constant at L,
the result is a growth path in which output Yt , capital Kt , and “effective
labor” Bt L all grow at rate gB . (Recall that, by CRS, equal proportional
increases to Kt and Bt L will produce an equal proportional increase to Y .)
If s is too small, output will be constrained by capital accumulation.
(This is clearest when s = 0.) In this case, Yt eventually approximately
equals AKt . Then
Kt+1 = Kt + sYt ≈ Kt + sAKt (5)
(using discrete-time notation for clarity), so capital and output both grow at
asymptotic rate sA.10 The requirement that capital accumulation keep up
with the growing effective labor force is thus the requirement that sA ≥ gB .
We will call this condition “sufficient saving”. Note that given any fixed
gA > 0 and gB , the sufficient saving condition will eventually be met.
Letting the labor force grow at some positive rate gL makes no interesting
difference. In this case, so long as sA ≥ gB + gL , Bt Lt grows at rate gB + gL ,
and Yt and Kt do likewise. Regardless of population growth, factor shares
are constant over time,11 so wages and capital rents per person grow at rate
gB .
The empirical causes of technology growth remain highly uncertain. An
exogenous growth model is one that does not attempt to model these causes,
but simply takes constant exponential growth in B for granted.
9
The saving rate is in fact historically (very roughly) constant, at least in developed
countries over the past century or so. There are relatively plausible ways to microfound
this phenomenon, if we wish to develop our model in more detail, and we will touch on
some of these in §3.5. For most of this document, however, we will simply take a constant
saving rate for granted.
10
Models without labor, in which Yt = AKt , are termed “AK models”. An AK economy
is of course only ever constrained by capital accumulation and exhibits growth at rate sA.
11
It follows from the identity sYt = Kt+1 − Kt that sYt /Kt = gK,t . Since in the long
run a constant saving rate maintains gY = gK , the long-run capital share (AK/Y )ρ equals
(sA/gY )ρ . If there is insufficient saving (so that gY = sA < gB ), or in the edge case of
sA = gB = gY , the capital share tends to 1.
 11

2.4 (Semi-)endogenous growth

On an endogenous growth account, on the other hand, growth in B is mod-
eled as the output of some deliberate effort, such as technological research.
That is, technology, like final output, is generated from inputs such as la-
bor, capital, and the stock of existing technology. In the most commonly
used research-based growth model, that of Jones (1995), the growth of B in
absolute terms is given by
Ḃt = θBtϕ (St Lt )λ (6)
for θ > 0, λ ∈ (0, 1], and (to generalize from Jones (1995) itself) unrestricted
values of ϕ. St denotes the fraction of the labor force working as researchers
(or “scientists”) at t. Intuitively, λ > 1 corresponds to cases in which re-
searchers complement each other, and λ < 1 corresponds to cases in which
some sort of “duplicated work” or “stepping on toes” effect predominates.
Output is given by Yt = F (Kt , Bt (1 − St )Lt ), as before.
In this setting, though output is CRS with respect to capital and effective
labor at any given time, it exhibits increasing returns to scale in population
across time. We therefore cannot continue to assume that all inputs to
production are paid their marginal products. In particular, we cannot
assume that technological innovators are compensated for all the additional
future output that their research produces on the margin; this sum of
marginal products would exceed total output! (Indeed, Nordhaus (2004)
estimates that innovative firms accrue on average only about 2% of the value
they produce.) It would be beyond the scope of this section to summarize
theories regarding the empirical or optimal number of researchers, or the
empirical or optimal level of worker pay. For now, to introduce endogenous
technological development without having to consider its interactions with
the rest of the framework, we might simply assume that a government
sets St and pays St Lt workers to do research. Their wages are equal to
non-research workers’ wages, in this stylization, and they are financed by
lump-sum taxes levied equally across the population.

To maintain a constant rate of output growth, we must maintain a constant

rate of labor productivity growth, by the reasoning laid out in §2.3. However,
the growth rate of B at t, denoted gB,t , is by definition equal to Ḃt /Bt .
Holding L and S fixed, therefore, we now have
gB,t = θBtϕ−1 (SL)λ . (7)
 12

If ϕ < 1, as we can see, gB,t falls to 0 as Bt grows. If ϕ > 1, gB,t rises to

infinity. Only in the knife-edge case of ϕ = 1 do we get exponential growth
with a constant number of researchers.12
As a matter of fact, the number of researchers has grown dramatically
over the past few centuries. Both the population and the fraction of the
population working in research have grown. For simplicity, and because the
number of researchers cannot grow indefinitely in a fixed population, let us
here ignore the second trend and suppose that S is fixed, with Lt growing
at a constant rate gL . In this case labor productivity growth gB is constant
over time iff

θBtϕ−1 (SLt )λ = θB0ϕ−1 exp(gB (ϕ − 1)t)S λ Lλ0 exp(gL λt) (8)

is constant over time; that is, iff the change in the number of researchers just
offsets the change in the difficulty of producing proportional productivity
increases. This in turn will obtain iff (ϕ − 1)gB + λgL = 0, or13

λgL
gB = . (9)
1−ϕ
Because we are holding 1 − St fixed, the number of non-research workers
will grow at rate gL . So the number of effective non-research workers will
grow at rate gB + gL . As we have seen, under a constant saving rate, capital
and output will grow at this rate too, and output per person will grow at
rate gB .

Observe that the steady-state rate of labor productivity growth is here un-
defined when ϕ = 1. The calculation also breaks down when ϕ > 1, absurdly
predicting a negative rate. This is because the value assumed to exist in
the derivation, namely a steady-state productivity growth rate gB under a
growing research workforce, does not exist when ϕ ≥ 1. When ϕ = 1, it
is straightforward to see that a positive growth rate in the number of re-
searchers produces an increasing rate of labor productivity growth. When
12
Indeed, some reserve the term “endogenous” for growth models in which ϕ = 1, since
an unexplained process of exponential population growth is needed when ϕ < 1. Models
of this form with ϕ < 1 are then termed “semi-endogenous”.
13
It follows from gB,t = θBtϕ−1 (SLt )λ that the corresponding growth path is stable. If
Bt is “too high”, growth subsequently slows, since ϕ − 1 < 0. Likewise, if Bt is “too low”,
growth accelerates.
 13

ϕ > 1, even a constant number of researchers produces ever-increasing labor

productivity growth as well.
The ever-increasing labor productivity growth rate gB,t that follows when
ϕ = 1 and gL > 0 translates into an increasing output growth rate up to the
point that gB,t +gL = sA. At that point capital accumulation cannot keep up
with the growth of the effective labor force, and production is constrained by
capital. If capital-augmenting technology can be developed in parallel with
labor-augmenting technology, however, the two factors can both grow at an
increasing rate, and output therefore can as well. That is, we have a Type I
growth explosion.
When ϕ > 1, moreover, even a constant number of researchers is enough
to produce “infinite output in finite time”, i.e. a Type II growth explosion.
The intuition for this is perhaps easiest to grasp when ϕ = 2, θ = 1, and
(SL)λ = 1, so that we have gB,t = Bt . Suppose gB,0 is such that B doubles
every time period. Thus B1 = 2B0 , so gB,1 = 2gB,0 . At this doubled growth
rate, B doubles every half-period; B1.5 = 2B1 . By repeated applications
of the same reasoning, the labor-augmenting technology level approaches
a vertical asymptote at t = 2. If capital-augmenting technology follows a
similar process, output approaches a vertical asymptote at t = 2 as well.
The potential for endogenous growth processes to produce explosive
growth is striking. However, since the researcher population growth rate has
long been positive and the productivity growth rate has long been roughly
constant, and in fact declining over recent decades (Gordon, 2016), we can
infer that at least historically ϕ < 1. Indeed, the most extensive study of the
topic to date—that done by Bloom et al. (2020)—estimates ϕ = −2.1. An
estimate of ϕ ∈ (0, 1) would indicate that, when we have access to a large
stock of existing technologies, these aid in the development of new technolo-
gies, but offer diminishing marginal aid. An estimate of ϕ < 0 implies that
when there is a large stock of existing technologies it is harder to develop
new technologies—perhaps because so much of the low-hanging technological
fruit has already been developed, with this “fishing out” effect outweighing
the effect of technological assistance in technological development.
 14

3 AI in basic models of good production

3.1 Capital productivity in isolation
At face value, AI promises to make capital more productive. This would
most naturally be modeled in the standard framework as an increase to A,
which would amount to effective capital accumulation. As Acemoglu and
Restrepo (2018a) point out, and as we have seen, this on its own would not
be predicted to have very transformative economic effects. It would increase
output and wages somewhat. But given ρ < 0 and a fixed or only slow-
growing labor supply, labor is the primary bottleneck to output, and any
increases to wages would come ever more from an increase in the labor share
rather than an increase in output. Indeed, the only “transformative” effect
of capital productivity is that, as A → ∞, all else equal, the labor share
should rise to 1. This is of course the opposite of the intuitive trend, which
is also the observed trend in the labor share in recent decades, especially in
the industries that have undergone most automation.
The models below, therefore, are all designed to shed light on the conse-
quences of increasing the productivity of capital in combination with various
structural changes to the production function that AI might also precipitate.

3.2 Imperfect substitution

Nordhaus (2021) explores the transformative possibility of AI in the stan-
dard model of good production without adding anything explicit about AI.
Instead, he posits that AI changes some of the model’s parameters “behind
the scenes”. This process has two steps.
First, suppose that AI raises the substitution parameter between labor
and capital (or certain kinds of capital, such as computers) so that it is per-
manently bounded above 0. In this case, capital accumulation is sufficient for
exponential output growth, even without population growth or technological
development of any kind.
For illustration, consider our CES production function with ρ > 0 and
technology represented but held fixed, and supposing the saving rate s is
constant. If the capital supply grows more quickly than than the labor supply,
Yt will come to approximately (in proportional terms) equal AKt , and capital
and output will accumulate exponentially at rate sA. More generally, if labor-
augmenting technology grows exogenously at some rate gB ≥ 0, the output
 15

growth rate following the substitutability change shifts from min(sA, gB ) to

max(sA, gB ). The substitutability change thus increases the growth rate as
long as sA > gB .
Second, suppose that At grows without bound. It does not matter
whether this technology growth is due to AI or to forces that predated (but
were less relevant before) the substitutability change. It also does not mat-
ter whether technology grows exogenously at some exponential rate gA > 0,
as Nordhaus posits, or is the output of human research effort as in (6)—in
which case, even under constant population, At rises without bound. In all
cases, the growth rate of output will tend to sAt , which, with At , will itself
be growing indefinitely. We will thus have a Type I growth explosion.
Under both transformative scenarios—the one-time growth rate increase
that can occur without capital-augmenting technological development and
the growth explosion that occurs with it—capital per worker will grow to
infinity. Since ρ > 0, the capital share will now tend to 1 rather than 0.
For any fixed value ρ < 1 however, capital and labor are still complements;
we still have FLK > 0. Absolute wages will therefore grow rapidly as the
effective capital stock grows, as long as ρ is bounded below 1. In fact, with
gA > 0, wages will grow superexponentially, though less quickly than output
or effective capital. Absolute wages will stagnate only if ρ rapidly rises to 1
(i.e. if ϵ rapidly grows to infinity)—or if capital and labor become perfect
substitutes, in which case ρ = 1 (i.e. ϵ is infinite). The latter case is explored
further in §3.3.

Nordhaus also discusses an analogous possibility: that AI will transform

consumption growth via the “demand side” of the economy, rather than the
“supply side”.
To explore this scenario, instead of dividing the space of goods into two
production inputs and one output, let us divide it into one input (“capital”
K) and two outputs (which might be called “standard consumption” Y and
“computer-produced consumption” Z). Capital grows exogenously at rate
gK . Given capital stock Kt , the production of the two consumption goods
must satisfy
Yt + Zt /Dt = Kt . (10)
That is, each unit of capital can produce either 1 unit of standard consump-
tion or Dt units of computer-produced consumption per unit time (without
being used up). 1/Dt is the relative price of Z at t: it is the number of units
 16

of Y that must be given up at t per unit of Z.

Consumers’ utility functions all equal U (·), defined over Y and Z. U (·)
has the same features a production function was assumed to have: it is
differentiable, increasing, and concave in each argument; the preferences it
represents are homothetic (recall §2.2); and its inputs are complements. In
response to consumer demand, production is allocated between Y and Z to
maximize utility.
Suppose Dt grows exponentially at rate gD . (This might be thought of as
Moore’s Law: famously, the number of computations that can be purchased
with a given amount of capital seems to double approximately every eigh-
teen months.) The relative price of Z then falls exponentially at rate gD .
With each proportional fall in this relative price, the relative quantity of Z
produced will rise by σt gD , where σt denotes the elasticity of substitution be-
tween the goods in the consumer utility function on the margin that obtains
at t.
Now let St denote the proportion of capital allocated to computing. The
relative quantity of Z produced equals Dt St /(1−St ). Considering the growth
rate of this term, by the reasoning above we have

gD + gS,t − g1−S,t = σt gD (11)

=⇒ gS,t − g1−S,t = (σt − 1)gD (12)

If σ is bounded above by σ̄ < 1, this is always negative. Over the long run,
gS must be negative and g1−S must be zero, since both terms are in the long
run non-positive. Thus we have, in the long run, gS ≤ (σ̄ − 1)gD . Finally,
since Z = DSK, we have

gZ ≤ σ̄gD + gK . (13)

On the other hand, if we maintain σt ≥ 1, then, as t → ∞, the fraction of

capital allocation to computing does not fall to zero. (Indeed, if σt is bounded
above 1, approximately all of capital is ultimately allocated to computing.)
So gZ = gD + gK .
Letting Ct ≜ Yt + Zt denote total consumption, the AI-relevant implica-
tions are straightforward. If computer-produced consumption is not currently
very substitutable for other consumption (σ bounded below 1), but develop-
ments in AI render it more substitutable (such that σ is then at least 1), then
the consumption growth rate could rise from something perhaps not much
higher than gK to fully gD + gK . This would not be a growth explosion, as
 17

we are using the term. But given the speed of Moore’s Law, it would be a
dramatic shift.
The definition of a “unit of consumption” is somewhat arbitrary, in the
presence of changing relative prices, and an exploration of the relevant work
on quantity indices would be outside the scope of this survey. In short,
though, the analysis of the previous paragraph effectively defines total con-
sumption in a given year in units determined by goods’ relative prices in the
starting year, whereas one might instead hold that units of Z contribute ever
less to “consumption” as the relative price of Z falls. Doing so only exacer-
bates the result above. Low substitution elasticity not only slows growth in
the production of Z but also generates a rapid fall in the relative price of Z,
which further slows the growth of measured consumption. On this account-
ing, therefore, the posited increase in the substitutability between “Y ” and
“Z” does even more to increase the measured consumption growth rate.
The wage and labor share are not defined in this model, since capital is
the only factor of production. As should be clear, however, an analogous
model with labor would behave similarly, as long as instead of simply
positing growth in capital gK , we also posit equal growth gB = gK in
labor-augmenting technology. Then the labor share will be constant (by
CRS), and consumption-denominated wages will grow at the consumption
growth rate.

Finally, Nordhaus constructs various tests of the hypothesis that we are

headed for a growth increase via the channels discussed above. If we are
in fact headed for a supply-driven growth increase, for example, we should
expect to find a rising growth rate and a rising capital share. If we are
headed for a demand-driven growth increase, we should expect to find
a rising share of global income spent on computer-produced goods. A
thorough discussion of his empirical conclusions would be beyond the scope
of this survey, but he concludes that on balance the evidence disconfirms
these hypotheses.

Models in which labor and capital must be combined in more complex ways
tend to produce the same broad conclusion. If labor and capital are suffi-
ciently substitutable, then increasing capital productivity can increase the
capital share, but it will still increase the absolute wage rate. Berg et al.
(2018) detail a variety of such models. We will not work through them here.
 18

3.3 Perfect substitution

We have seen that, if labor and capital are the only two factors of production,
then whenever the elasticity of substitution between them is finite, increases
to the quantity of effective capital cause absolute wages to grow. Thus, if
the elasticity shifts from less than 1 to greater than 1—a shift which can
allow for faster capital accumulation—wage growth can accelerate, even as
the labor share falls.
As we will see, however, prospects for wages look worse in cases of
perfect substitutability. In this case, if there are only two production
factors, the returns to each must be linear. Increases to the quantity of
effective capital thus have no impact on the wage rate. If there are multiple
production factors, and if some grow scarce relative to effective “capital plus
labor”, increases to the quantity of effective “capital plus labor”—driven by
increases in effective capital—drive wages down.

A model of the beginning of perfect substitution between labor and capital

can be presented most simply as one in which human-substitute robots are
simply at first expensive, and then cheap, in units of human labor hours.
This is because, as noted in §2.2, when goods are perfect substitutes toward
some end, they are only ever both purchased in positive quantities when their
prices are the same. Even if it were already feasible to produce robots fully
substitutable for human labor, therefore, we would only see any produced,
and observe their effects, once their rental rate had fallen below what would
otherwise have been the wage rate. In other words, perhaps the substitutabil-
ity does not need to rise; perhaps it is perfect, and all that needs to change
is a relative price.
To illustrate this dynamic, consider the following simple model, inspired
by Hanson (2001). Equipment Q, labor L, and land W are employed in a
Cobb-Douglas production function,

Y = F (Q, L, W ) = Qa Lb W 1−a−b . (14)

The output good can be consumed or invested as capital K. Capital can

serve either as equipment or as robotics, which functions as labor, whereas
the human workforce H is fixed and can only serve as labor. The productivity
of capital—the number of units of effective capital generated by one unit of
converted output—is denoted A. That is, if p denotes the fraction of capital
 19

employed as equipment, output is

Y = (pAK)a (H + (1 − p)AK)b W 1−a−b . (15)
At rises exogenously without bound. For simplicity we will assume that a
constant and sufficient fraction s of output is saved as capital. Because the
substitution parameter between equipment and labor is not less than (in fact
is equal to) 0, the accumulation of effective equipment is enough to sustain
output growth.
Early in time, when effective capital is scarce, all capital is used as equip-
ment; p = 1. Indeed, at the rate at which capital can be converted from
equipment to robotics, it would be valuable instead to use some human labor
as equipment, if that were possible. Capital then grows (using discrete-time
notation for clarity) such that
Kt+1 = Kt + s(At Kt )a H b W 1−a−b (16)
⇒ gK,t = (Kt+1 − Kt )/Kt = sAat Kta−1 H b W 1−a−b .
As we can see from the right hand side, capital growth will approach a steady
state such that
a
agA + (a − 1)gK = 0 ⇒ gK = gA . (17)
1−a
We will thus have output growth of gY = a(gA + gK ) = gA a/(1 − a).
As the equipment stock grows, wages rise. As the productivity of capital
rises and effective equipment grows more abundant, however, there comes
a time past which it is optimal to split further capital between equipment
and robotics. The labor growth rate then jumps to the rate that keeps its
marginal productivity equal to that of equipment, and the output growth
rate jumps accordingly.14 In particular, with capital now filling the roles of
both equipment and labor, we now have gY = gK = gA (a + b)/(1 − a − b), by
the same calculation as above.15
14
As Yudkowsky (2013) points out, we might interpret this as a model in which AI
comes in the form of “emulations”—a theoretical technology on which Hanson has written
extensively—which are always technically feasible but which are, at first, prohibitively
expensive, because effective capital is sufficiently scarce.
15
Note that, were it not for the inclusion of the non-accumulable factor land, there
would be no steady-state growth rate; in solving for it, we would have to divide by 0.
Instead, the economy would be, asymptotically, an AK economy with exogenous capital
productivity growth. As we saw in the previous section, we would have a Type I growth
explosion.
 20

Hanson estimates the growth implications of crossing the robotics

cost threshold using a slightly more realistic model with roughly realistic
estimates of the parameters involved. The productivity of capital is assumed
to double (i.e. the cost of effective capital is assumed to halve) every two
years, in a conservative approximation to Moore’s Law. Before capital
begins to be used as robotics, output in the model grows at a relatively
familiar rate of 4.3% per year. After, the growth rate is 45%.

In the model above, because the production function is Cobb-Douglas, the

labor share—the share of output paid in compensation for human and/or
robotic labor—is independent of the factor quantities. As human labor con-
stitutes an ever smaller share of total labor, however, the human labor share
falls to zero.
Furthermore, even the absolute wage FL falls to zero. To see this, note
that in a CRS production function, the marginal productivities of equipment
and labor are kept equal (FQ,t = FL,t ) when the quantities of the two fac-
tors grow at the same rate. We can thus rearrange our formula regarding
competitive CRS factor payments:
Yt − FW,t W
FL,t Qt + FL,t Lt + FW,t W = Yt ⇒ FL,t = . (18)
Qt + Lt
With a constant share of output accruing to land as well, but the quantity
of land fixed, land rent per unit of land—i.e. the land rental rate FW —must
grow at the same rate as output. Y and FW will thus both grow at gY , and
Q and L will both grow at rate gA + gK = gA + gY > gY . The right-hand
ratio will then fall to zero.
In any CRS production function without labor-augmenting technology,
what happens to the marginal productivity of labor, and thus wages, depends
on the quantity of effective labor relative to that of the other effective factors
of production. This relative quantity need not rise; it could fall, if labor’s
complements grow productive and plentiful more quickly than its substitutes,
or stay fixed if they grow at the same rate.
Consider the following model, very similar to the above, but in which
technology augments only equipment, not capital used as robotics:
Y = F (Q, L, W ) = Qa Lb W 1−a−b = (pAK)a (H + (1 − p)K)b W 1−a−b . (19)
The growing stock of equipment implies that, as above, wages rise before
the substitutability cost threshold is crossed. Furthermore, we will still have
 21

gY = gK , and thus
a
gY = a(gA + gK ) ⇒ gY = gA (20)
1−a
before the threshold is crossed. Finally, the threshold will still eventually
be crossed: if all capital were used as equipment indefinitely, the marginal
productivity of capital used as equipment AFQ = aY /K would fall below
that of labor FL = bY /H.
After the threshold is crossed, we will have
a
gY = a(gA + gK ) + bgK ⇒ gY = gA . (21)
1−a−b
Note that this is still a growth rate increase, though not as large as that in
the first model above.
To see what happens to wages, however, observe that when p is chosen
so that invested output is split optimally between equipment and robotics,
it will satisfy AFQ = FL , or
aY bY H +K a
= ⇒ p= . (22)
pK H + (1 − p)K K a+b
As K → ∞, we have p → a/(a + b) < 1, and therefore gL = gK = gY . As
above, because the production function is Cobb-Douglas, the labor share
is constant. Now, however, the quantity of effective labor grows no more
quickly than output, so labor payments per labor quantity—i.e. wages per
human worker—merely stagnate.

Korinek and Stiglitz (2019) offer another illustration of this phenomenon, in

the context of a somewhat similar model. As usual we will simplify here to
highlight the intuition.
Suppose that Y is produced as in the second model of this section (i.e.
the model just above), except that the substitution parameter between land
and the other two factors is bounded below 0. Though land is in fixed supply,
it is at first plentiful enough that its factor share is low. The saving rate is
fixed.
At first, as capital accumulates, it is split between use as robotic labor
and use as equipment, so that the relative quantities of labor and equipment
are unchanged. The capital and labor shares are roughly constant, but the
absolute wage stagnates, as we have seen. In time, however, land becomes a
 22

binding constraint. The share of output received as land rents approaches 1,

and the absolute wage falls to 0.
As should be clear, the same logic could apply to many more complex
models. Embed any production function in a “surrounding” production func-
tion with a fixed-supply and low-substitutability resource such as land, and
in the long run, even if all of the original production function’s resources
grow abundant, the resource in fixed supply constrains growth and its own-
ers receive approximately all output.

3.4 Substitutability in robotics production

Like Korinek and Stiglitz, Mookherjee and Ray (2017) develop a model in
which capital can replace human labor without technological progress. Unlike
Korinek and Stiglitz, they do not simply assume that capital can be used as
robotics, but make the robot production function explicit and identify a
condition under which human labor replacement can occur. A simplification
of their model is as follows.
The final good Y is produced using capital K and labor L in a typical
two-factor production function F (·), with a substitution parameter bounded
below 0. Labor is supplied by human work H and robotics R, which are
perfect substitutes. Robotics is better thought of as the provision of robot
services than as robots, because it must be used as it is produced; it cannot
be accumulated. If we would like to think of it as a kind of physical capital,
we would say that it exhibits full depreciation.
Robotics is also produced using capital and labor, using a standard but
perhaps different production function f (·), also with a substitution parameter
bounded below 0. Whereas one unit of robotics is defined as that which
replaces 1 human worker in the output production function, however, one
unit of robotics replaces some D ∈ (0, 1) human workers in the robotics
production function. For each input X ∈ {K, H, R}, SX is defined (assuming
X > 0) to be the fraction of X that is allocated to the production of robotics
rather than the final good. For simplicity, the population of human workers
is fixed and there is no technological progress. More formally, output and
robotics at t are

Yt = F ((1 − SK,t )Kt , (1 − SH,t )H + (1 − SR,t )Rt ); (23)

Rt = f (SK,t Kt , SH,t H + DSR,t Rt ).
 23

As usual, a constant fraction of output is saved as capital.

Early on, when capital is scarce and human labor relatively plentiful,
there may be no reason to produce robotics at all. As capital accumulates
and output begins to be constrained by human labor, however, the marginal
output productivity of capital falls to zero. It may therefore at some point be
worthwhile to allocate some positive fractions SK and SH of available capital
and human labor to robotics production. To be precise, it will necessarily
start being worthwhile iff fL (k, 0) > 1 given k > 0, i.e. if, given some cap-
ital (which is eventually near-worthless in final good production), marginal
contributions of labor can create robotics at a ratio of more than 1:1. Let
us call this the “robotization condition”. Note that it is a relatively weak
condition; fL (k, 0) = ∞ given k > 0 if f (·) is CES, for example.
If robotics production relies on capital and human labor—i.e. if we set
SR = 0—it too will ultimately be constrained by lack of labor:

as SK,t Kt → ∞, Rt → SH,t R̄, where R̄ ≜ lim f (x, H). (24)

x→∞

Output in turn is constrained by total labor, despite the possibility of

robotics:
(1 − SH,t )H + Rt
as (1 − SK,t )Kt → ∞, Yt → Ȳ , (25)
H + R̄
where Ȳ ≜ lim F (x, H + R̄).
x→∞

∗
Long-run output will be maximized by setting SH = SH , the value that
∗ ∗ ∗
maximizes (1 − SH )H + SH R̄—i.e. SH = 1 if the robotization condition is
∗
met, because in this case R̄ > H, and SH = 0 if not. Long-run output will
∗ ∗
then approach an upper bound of Ȳ ((1−SH )H +SH R̄)/(H + R̄). The human
labor share will approach 1, either because it is the scarce input to output
directly or because robotics is the scarce input to output and human labor
is the scarce input to robotics. The absolute wage will of course stagnate.
In short, robotization can raise the output ceiling, but it cannot on its own
produce a sustainably positive growth rate.
One might expect that, if we do not fix SR = 0, it will eventually be
optimal to use robotics in the production of robotics. In fact, this will only
happen if D is large enough that, as the quantity of capital allocated to
robotics production grows large, one unit of robotics can produce more than
one unit of robotics: that is, if limk→∞ fL (k, H) > 1/D; or equivalently,
 24

because f (·) is CRS, if fL (k, 0) > 1/D for k > 0. The identification of such a
condition in their more general setting is Mookherjee and Ray’s key insight,
and they call the condition the “von Neumann singularity condition”, after
the work by Burks and Von Neumann (1966) on self-replicating automata.
It is of course very closely analogous to the robotization condition above, but
stronger since we are imposing D < 1. We might take this to be the natural
case; robotics production is presumably harder to automate than most other
tasks are.16
Suppose that this condition is met, and that SK K is large enough that
fL (SK K, H) > 1/D. Then there is an optimal quantity of robotics to allocate
to robotics production, so as to maximize net robotics production. This is the
quantity such that use of a marginal unit of robotics on robotics production
increases robotics output by exactly one unit. That is, it is optimal to set
SR > 0 such that SR satisfies fL (SK K, H + DSR∗ R) = 1/D.
The value of R depends on SR , since more inputs to robotics production
will correspond to higher robotic output. Nevertheless we know that a unique
SR ∈ (0, 1) satisfying the above equality exists, for a given SK K. To see this,
recall that fL (SK K, H + DSR R) > 1/D at SR = 0, by supposition. And
we must have limSR →1 fL (SK K, H + DSR R) < 1/D, or else the quantity of
robotics output R and thus also SR R would grow without bound, fixing SK K,
as SR → 1; but in this case fL → 0, by the assumption that the substitution
parameter in the robotics production function is bounded below 0. By the
concavity and continuous differentiability of f (·), therefore, there is a unique
SR∗ : fL (SK K, H + DSR∗ R) = 1/D.
∗
Under the singularity condition, SK and SR approach constants SK and
∗ 17
SR , strictly between 0 and 1, as the capital stock grows. Growth proceeds
as in an AK model, with the rate of capital accumulation, final good output
growth, and robotics output growth all asymptotically constant and propor-
tional to the saving rate.18 The wage level is constant and lower than it is in
16
Presumably at least some other tasks are more difficult to automate, however. As
Mookherjee and Ray present in the original paper, their central result does not depend on
robotics production being more difficult to automate than all other tasks, just on robotics
production being sufficiently difficult to automate.
17
See Appendix A.1 for a proof.
18 ∗ ∗
The model will approximate an AK model with A = Y /K = F ((1–SK ), (1–SR )R/K),
∗ ∗
where R/K likewise satisfies R/K = f (SK , SR R/K). Technological progress that allows
capital to produce more robotics increases long-run R/K and functionally “increases A”,
though it will never exceed its upper bound of F (1, ∞). This will be finite, by the as-
sumption that F (·)’s substitution parameter is bounded below 0.
 25

the absence of the singularity condition, since the ratio of capital to labor in
robotics production is still asymptotically constant but now positive rather
than zero. The share of income accruing to human labor falls to zero.
As with Hanson (2001), moderate tweaks to this model could result
in absolute human wages rising or falling, rather than merely stagnating.
Also, in the presence of population growth or labor-augmenting technology
growth, introducing automation can increase the growth rate of final good
output (or final good output per capita) from the rate of effective labor
(or labor-augmenting technology) growth to something much higher, given
sufficient saving.

As we saw in §3.3, when human work must compete with robotics for which it
is perfectly substitutable, the standard result is that the human labor share
falls to 0 and the wage stagnates or changes (likely falls) as well. Above,
however, we saw that modeling robotics production explicitly, rather than
stipulating a frictionless conversion of the final output good into robotics,
allows for a channel through which human work can remain necessary. A
positive human labor share can be maintained, even when robotics can fully
substitute for human work in the final good production function, when human
work cannot be fully substituted for in robotics production.
Korinek (2018) presents another model in which humans and robots must
in some sense compete and in which robots are not simply “capital that can
function as labor” but items that must be produced and sustained. He too
finds that the human labor share, and in his case also the wage rate, can fall
to 0 unless human labor remains necessary for the maintenance of the robot
population. But we will not explore his model further here.

3.5 Growth impacts via impacts on saving

In some of the models we have considered, saving has been an important
determinant of growth. The saving rate, however, has been assumed to be
exogenous. This leaves open another channel through which developments
in AI could impact growth: by changing the rate of return to saving, more
advanced AI could change the rate of saving and thus the growth rate.
This scenario can be illustrated most simply using a model from Korinek
and Stiglitz (2019). Suppose labor and capital are perfectly substitutable.19
19
Equivalently, we could say that output is produced by a single factor, labor, which
 26

Labor can only be supplied by humans. Activity unfolds in discrete time,

and capital depreciates fully every period; it cannot accumulate. (Capital
depreciation simplifies the exposition but is not necessary for the central
result.) Given saving rate st , that is, output and capital growth are then

Yt = AKt + BLt , Kt+1 = st Yt . (26)

We begin with K = 0. Output per capita is thus B and, without saving,

does not grow. If A < 1, there is no incentive to save, and doing so cannot
generate growth; foregoing each unit of consumption at t would offer someone
only A < 1 additional units of consumption at t + 1, starting from the same
baseline of B. For any A > 1, on the other hand, individuals with no or
sufficiently low pure time preference will want to save some fraction st > 0
of their incomes; not to do so would be to miss the opportunity to give up
marginal consumption at baseline B in exchange for a larger quantity of
marginal consumption also at baseline B. More precisely, it should be clear
that positive saving will be optimal, given any pure time discount factor
β < 1, as long as A > 1/β. Furthermore, under certain assumptions about
the shape of individuals’ utility functions, the induced saving rate will be
some constant s > 1/A, independent (at least roughly) of the absolute output
level. In the long run, as the relative contribution of effective human labor
BLt grows negligible, we will have Yt ≈ AKt . And since

Kt+1 = sYt ≈ sAKt (27)

⇒ (Kt+1 − Kt )/Kt ≈ sA − 1,

capital and therefore output will grow at asymptotic rate sA − 1 > 0.

In short, an increase in A—induced, perhaps, by AI developments which
render robots cost-effective replacements for human labor—can trigger
saving and can thus increase the growth rate of output and output per
capita. Here, an A-increase raises per capita output growth from zero to
a positive number, leaves the wage rate constant at B, and pushes the
human labor share to zero; but other impacts on wages, the labor share,
and growth are possible. The point is just that, in addition to the ways in
which increases in A can sometimes directly impact the growth rate, they
can sometimes do so indirectly by impacting s.

can be supplied both by humans and by robots.

 27

There is another mechanism through which developments in AI could impact

the saving rate. If the saving rate is heterogeneous across the population,
then growth will depend on how income is distributed between high- and low-
savers. Developments in AI could thus affect the growth rate by affecting
the income distribution. In principle, this effect could have implications for
growth in either direction. Here, we will focus on the especially interesting
and counterintuitive possibility that AI slows and even reverses growth by
transferring wealth from those with low to those with high propensity to
consume.
This scenario is illustrated most simply by Sachs and Kotlikoff (2012),
though the same mechanism is explored in more detail by Sachs et al. (2015).
If some investment goods are sufficiently substitutable for labor, automation
raises capital rents but lowers wages. If saving for the future comes dispropor-
tionately out of wage income, for whatever reason, then this wage-lowering
can cause future output to fall.
Consider an overlapping generations (OLG) economy with constant pop-
ulation size. Each person lives for two periods. The young work, investing
some of their income; the old live off their investments. More precisely, out-
put is a symmetric Cobb-Douglas function of labor and capital. The output
good can be consumed or invested as capital K. Capital can be used either
as equipment Q or as robotics, which serves as labor, and it is split between
these uses until their marginal products are equal. The human workforce H
is fixed and can only serve as labor. Unlike in the Hanson model, however,
the productivity of equipment is fixed. A denotes the productivity of robotics
only.
Formally, if p is the share of capital used as equipment,
Yt = F (Qt , Lt ) = (pt Kt )1/2 (H + (1 − pt )At Kt )1/2 . (28)
Capital at t is financed by those who were young at t − 1, who put aside half
their wage incomes as investment.20 The old at t consume all their wealth:
not only their investment income, FQ,t pt Kt + FL,t (1 − pt )At Kt , but even the
capital stock Kt , which is liquidated after use in production. The economy
is in a zero-growth steady state when investment is just replenished each
period: that is, when FL,t /2 = Kt .
20
Let rt denote the interest rate at t: that is, here, FQ,t+1 , or equally At+1 FL,t+1 .
Suppose that period utility is logarithmic in consumption and that the young choose the
saving rate st to maximize lifetime utility ln((1 − st )FL,t ) + ln(st FL,t (1 + rt )). Then the
chosen saving rate will always equal 1/2.
 28

Now suppose robotics grows slightly more productive, so that At+1 > At .
Let GA ≜ 1 + gA denote At+1 /At . For a single period, total output and
the incomes of the old grow. The young see a fall in wages, however, and
investment therefore falls as well. This fall in investment outweighs the fact
that some of the investment, namely that in robotics, is now more productive.
Output therefore falls. The wage rate falls too, due both to the abundance
of robotics and to the lack of investment in equipment.
More formally: in the new equilibrium, the marginal product of robotics is
again equal to that of equipment. Because the “relative cost” of robotics has
now been multiplied by 1/GA and because here ρ = 0, the relative quantity of
labor must be GA times higher. Letting asterisks denote the new equilibrium
outcomes, L∗ /Q∗ = GA Lt /Qt . So

1  L∗ −1/2 1  Lt −1/2
FL∗ = , FL,t = (29)
2 Q∗ 2 Qt
−1/2
⇒ FL∗ = FL,t GA < FL,t .

The new rate of return on investment is r∗ = At+1 FL∗ = GA At FL∗ , the new
wage is FL∗ , and the saving rate remains 1/2. The consumption of the old
thus equals half their income while young, times 1 + r∗ :
FL∗ 1/2 FL,t
(1 + GA At FL∗ ) = (1 + At GA FL,t ) 1/2 (30)
2 2GA
−1/2 FL,t FL,t
= (GA + At FL,t ) < (1 + At FL,t ) .
2 2
The new equilibrium therefore features lower output in all subsequent peri-
ods, and lower consumption for both young and old.
If robotics productivity continues to grow at rate gA , the wage rate, and
1/2
thus the consumption of the young, will continue to fall to 0 at rate GA −
1. The consumption of the old (and thus also output) will fall at a falling
rate, and will eventually stabilize above 0, as the increasing productivity of
invested equipment ever more closely compensates for the falling absolute
amount invested.21 The human labor share thus falls to 0.
Again, the direction of these impacts is sensitive to whether the “winners”
from advances in AI save more or less than the “losers”. As Berg et al. (2018)
21
Technically, if flow utility is logarithmic in consumption (see footnote 20), lifetime
utility falls to negative infinity as the consumption of the young falls to zero.
 29

point out, for instance, those who make most of their incomes from wages
currently empirically exhibit lower saving rates than those who make most
of their incomes from capital rents, so the mechanism identified by Sachs
and Kotlikoff should if anything increase output growth. In any event, the
key point is just a reiteration of the well-known fact that, in a neoclassical
growth model with finitely lived agents and no (or imperfect) intergenera-
tional altruism, the rate of saving is not necessarily optimal. Accordingly,
policymakers must always consider not only the impact of a policy or techno-
logical development on short-term output, but also its impact on the saving
rate. When a given development produces a suboptimal saving rate, it should
be counterbalanced by investment subsidies or by transfers from those with
high to those with low propensity to consume: in this model, from the old
to the young.

4 AI in task-based models of good produc-

tion
4.1 Introducing the task-based framework
In §3, we imagined that capital and labor were each employed in a single
sector. In the Cobb-Douglas case, we held the exponent a on capital fixed.
We then explored the implications of changing ρ, the substitutability of cap-
ital and other durable investments for human labor, and of independently
changing the growth rates of factor-augmenting technology.
In reality, however, capital and labor are of course employed heteroge-
neously, and this heterogeneity seems likely to shape the economic impacts
of developments in AI. Indeed, sectors with high rates of automation have
historically experienced stagnating or declining wages (Acemoglu and Au-
tor, 2012; Acemoglu and Restrepo, 2020), even as wages on average have
increased.
Here, therefore, we will explore a model of CES automation from Zeira
(1998), which makes room for this sort of heterogeneity. (We will follow the
exposition and extension of Zeira’s model given by Aghion et al. (2019).)
As we will see, this model amounts roughly to assuming a fixed substitution
parameter ρ and either a changing capital exponent a, in the Cobb-Douglas
case, or impacts on factor-augmenting technology which are sensitive to ρ
when ρ ̸= 0.
 30

Let us begin with the ρ = 0 case. Suppose output is given by a Cobb-Douglas

combination of a large number n of factors Xi , for i = 1, . . . , n:

Y = X1a1 · X2a2 · · · · · Xn1−a1 −···−an−1 . (31)

At such a fine-grained level, these “factors” might better be thought of as in-

termediate production goods (Zeira, 1998), or even as individual tasks (Ace-
moglu and Autor, 2011). We will refer to them as tasks.
Fraction a of the tasks are automatable, in that they can be performed by
capital or labor, and fraction 1−a are not, in that they can only be performed
by labor. Given capital and labor stocks K and L, if all automatable tasks
are indeed automated (performed exclusively by capital), K/(na) units of
capital will be spent on each automatable task and L/(n(1 − a)) units of
labor on each non-automated task. With just a little algebra, we have

Y = AK a L1−a , (32)

a two-factor Cobb-Douglas production function with an unimportant

coefficient A.22

Now consider a general CES production function with a continuum of pro-

duction factors Yi from i = 0 to 1, instead of just two:
Z 1 1/ρ
Y = Yiρ di (33)
0

Tasks i ≤ β ∈ (0, 1) are automatable.

Let K and L denote the total supplies of capital and labor, and Ki and
Li the densities of capital and labor allocated to performing some task i (so
Yi = Ki + Li ). Suppose again that all automated tasks are indeed performed
exclusively by capital. Since the tasks are symmetric and the marginal prod-
uct of each task is diminishing (∂ 2 Y /∂Xi2 < 0), the density of capital applied
to each task i ≤ β will be equal (assuming that production proceeds effi-
ciently), as will the density of labor applied to each i > β. And since we
must have Z β Z 1
Ki di = K and Li di = L, (34)
0 β

22
A = a–a (1–a)a–1 , which ranges from 1 (at a = 0 or 1) to 2 (at a = 1/2).
 31

we know that Ki = K/β ∀i ≤ β and Li = L/(1 − β) ∀i > β. We can thus

write our production function as

Y = [β(K/β)ρ + (1 − β)(L/(1 − β))ρ ]1/ρ (35)

= [β 1−ρ K ρ + (1 − β)1−ρ Lρ ]1/ρ
= F (AK, BL) = [(AK)ρ + (BL)ρ ]1/ρ

where A = β (1−ρ)/ρ and B = (1 − β)(1−ρ)/ρ . This is simply a two-factor CES

production function.
We have assumed that all automatable tasks are indeed performed ex-
clusively by capital. This will obtain so long as there is more capital per
automatable task than labor per non-automatable task, i.e. as long as
K L
> . (36)
β 1−β
In this case, even when capital is spread across all automatable tasks, we
have FK < FL , so there is no incentive to use labor on a task that capital
can perform. Let us call this condition the “automation condition”. For any
fixed β, if capital accumulates indefinitely and the labor supply stays fixed
or grows more slowly, the condition will eventually hold.

4.2 Task automation

Let us now explore the implications of task automation in more detail,
across the regimes of CES production with ρ = 0 and < 0. The case of ρ > 0
will be covered in §4.3.

As we have seen, under Cobb-Douglas production, task automation raises

a along the range from 0 to 1. Recall the Cobb-Douglas production func-
tion given factor-augmenting technology, (4). Since a constant saving rate
imposes gY = gK , the growth rate in this case satisfies

gY = a(gA + gK ) + (1 − a)(gB + gL )
a
⇒ gY = gA + gB + gL . (37)
1−a
The impact of a one-time increase to a, or of increases only up to some
bound strictly below 1, is therefore straightforward. The capital share rises
 32

with a. If gA > 0, the growth rate increases, ultimately raising the wage rate;
otherwise the growth rate is unchanged, and the impact on the wage rate is
ambiguous.
Given asymptotic complete automation, with 1 − a falling to 0 expo-
nentially or faster, the model approximates an AK model. If gA = 0, the
growth rate rises to sA. If gA > 0, the growth rate rises without bound, and
wages too rise superexponentially.

Automation, as we have defined it, allows capital to perform more tasks. One
might therefore imagine that it is equivalent to the development of some sort
of capital-augmenting technology. Aghion et al. (2019) observe, however, that
automation in the above model is actually equivalent to the development of
labor -augmenting—and capital-depleting!—technology, as long as ρ < 0 and
the automation condition holds. To see this, recall our production function:

Y = [(AK)ρ + (BL)ρ ]1/ρ , (38)

where A = β (1−ρ)/ρ and B = (1 − β)(1−ρ)/ρ . As β rises from 0 to 1, therefore,

A falls from unboundedly large values to 1, and B in turn rises from 1 without
bound.
The reason for this result is that, as β rises, capital is spread more thinly
across the widened range of automatable tasks, and labor is concentrated
more heavily in the narrowed range of non-automatable tasks.23 Automation
therefore allows capital to serve as a better complement to labor. A marginal
unit of labor is spread across fewer non-automatable tasks, producing a larger
increase to the supply of each; given the abundance of capital, this then
produces a larger increase to output. Conversely, under this allocation, labor
serves as a worse complement to capital, requiring capital to spread itself
over more tasks (and only partially compensating for this effect by supplying
the remaining tasks more extensively).
As explained in §2.3, when ρ < 0, labor-augmenting technology is the key
to sustained output growth. Let us spell that out in this context. Suppose
that, by some exogenous process, a constant fraction of the remaining non-
automatable tasks are made automatable each period, so that (1−βt ) → 0 at
a constant rate g1−β < 0. Then B will grow at rate gB = g1−β (1 − ρ)/ρ > 0.
23
Here we are only considering increases in β up to the point that capital per au-
tomatable task no longer exceeds labor per non-automatable task, so that the automation
condition is satisfied.
 33

A is asymptotically constant at 1, so gA ≈ 0. If the saving rate s is constant

and high enough to maintain the automation condition, we get our familiar
“balanced growth path”. The capital stock, effective labor supply, and output
all grow at asymptotic rate gY = gB + gL , output per capita grows at rate
gB , and the labor share is asymptotically constant and positive.24
Automation can thus increase the growth rate of output per capita, and
have other transformative consequences.25 In the model above, because the
automation rate −g1−β is the only driver of growth, introducing it increases
the growth rate from 0 to g1−β (1 − ρ)/ρ. In the presence of growth from
other sources, automation can increase the growth rate further. Consider for
instance what follows if we have Bt = Dt (1 − β)(1−ρ)/ρ , with β constant but
Dt growing exogenously at rate gD . Given saving sufficient to maintain the
automation condition, output per capita grows at rate gD . The implications
of introducing automation at rate −g1−β then depend on whether saving is
still sufficient to maintain the automation condition. If it is, the per-capita
growth rate increases to gD + g1−β (1 − ρ)/ρ, and the labor share falls to an
asymptotic positive value, as observed above. For more on a model of task
automation with (something close to) direct labor-augmenting technology
growth gD > 0, see §4.4.
Now suppose again that gD = 0, but now suppose that saving is not
sufficient to maintain the automation condition—as it cannot be if, for
instance, all tasks become automatable. In this case some automatable
tasks will not be automated. Here, things proceed roughly as in a model
of full substitutability. The growth rate is capped at sA, the wage rate
equals the capital rental rate and stagnates, and the labor share equals
L/(L+K). Assuming sA > gL , the labor share falls at rate gL −gK = gL −sA.

24
The capital share here equals βt1−ρ (Kt /Yt )ρ . As β → 1, the capital share rises to an
upper bound of (K/Y )ρ , where K/Y is the long-run capital-to-output ratio, as long as
this exists and is finite. It follows from sYt = Kt+1 − Kt that sYt /Kt = gK,t ; and since
gK = gY = gB + gL , we have K/Y = s/(gB + gL ). The labor share will thus fall to
1 − (s/(gB + gL ))ρ . This is nonnegative because sA ≥ gB + gL , by sufficient saving, and A
is asymptotically 1; and it is strictly positive as long as we are not in the knife-edge case
of sA = gB + gL .
25
One might however take the position that what we have here been calling automation
is not a new force on the horizon, promising to augment pre-existing drivers of growth,
but a microfoundation for the process of labor-augmenting technological development we
have observed for centuries. On this view, advances in AI will continue to push β ever
closer to 1, but this process will simply continue the existing trend.
 34

Finally, consider the implications of exogenous growth in capital-augmenting

technology A. When not all tasks are automated, increases to A only increase
the effective capital stock. If gA > 0 and s > 0, even if the automation
condition is not yet met, the growth rate sA will grow until it is met. The
capital stock will then grow at the output growth rate, the effective capital
stock will grow faster by gA , and the capital share will fall to 0, roughly as
explained in §3.1.
When all tasks are automated, on the other hand, output Yt asymptoti-
cally equals At Kt , and sustained exponential growth in A produces a Type I
growth explosion.

4.3 Task creation

Let us begin with the Aghion et al. (2019) model of automation and introduce
a process of task creation. The resulting model will be somewhat akin to that
developed by Hémous and Olsen (2014).
As before, output is a CES production of a range of tasks. Each task is
performed by labor and/or capital, with tasks above an automation threshold
β requiring labor. Now, however, new and initially non-automated tasks can
be created. The range of tasks thus runs from i = 0 to N , with tasks i ≤ β
automatable, and not only β but also N can be increased. By the same
reasoning as in §4.1, if there is enough saving that the automation condition
is met, output is
Y = [(AK)ρ + (BL)ρ ]1/ρ (39)
where A = β (1−ρ)/ρ and B = (N − β)(1−ρ)/ρ .
If ρ < 0, then increases to β holding N fixed act like labor-augmenting
technology, as in Aghion et al. (2019). By the same token, however, increases
to N holding β fixed act like labor-depleting technology; they require labor
to “spread itself too thinly”. It will never be productive to create new tasks,
and automation and growth will simply proceed as in §4.2.

If ρ > 0, on the other hand, it is increases to N , holding β fixed, that function

as labor-augmenting technology. In particular, they asymptotically produce
gB = gN (1 − ρ)/ρ. As explained in §3.2, growth then proceeds at a rate of
max(sA, gB ), where s denotes the saving rate. Increasing the rate of task
creation can thus increase the growth rate.
More importantly, increases to β, regardless of N , function as advances
 35

in capital-augmenting technology. Recall that effective capital accumulation

is enough for growth when ρ > 0. By raising the “ceiling” N and allowing
for future automation to raise β, task creation can thus have radical effects.
To see this, first suppose that N increases exogenously at a constant
proportional rate gN , and that N − β is constant. This is essentially the case
explored by Nordhaus (2021) and summarized in §3.2: given ρ > 0, capital
accumulation and capital-augmenting technology growth combine to produce
a Type I growth explosion. The labor share will fall to 0, even while wages,
like output (though more slowly than output), grow superexponentially.
If gA = gB , i.e. if gβ = gN so that a constant fraction of tasks is
always automated, the outcome is similar; indeed, conditions are even
more favorable to labor. Upon endogenizing the task automation and
creation processes, the gβ = gN condition turns out to hold under relatively
natural-seeming circumstances. For a few more words on this, see the end
of the following section.

Finally, as discussed at the end of §3.3, suppose we embed production func-

tion (39) in a “surrounding” production function. That is, suppose that the
good denoted Y , which we had been referring to as the final output good,
must instead be combined with a fixed-supply resource (such as land W ) in
order to produce the final output Z.
What follows depends centrally, as we have seen, on the substitution
parameter ρ′ between Y and W . If ρ′ > 0, essentially nothing changes; the
land share is asymptotically 0, growth in Z approximately equals growth in
Y , and so forth. If ρ′ < 0, on the other hand, output approaches an upper
bound. Then the relative quantity of Y rises without bound, by capital
accumulation or asymptotic task automation; the land share rises to 1; and
the wage rate falls to 0.
This is more similar to the case explored by Hémous and Olsen, though
they take the fixed-supply resource to be skilled labor, whereas we are ignor-
ing skill differences throughout this review. In any event, we will not explore
this further here.

4.4 Task replacement

Acemoglu and Restrepo (2018b, 2019a) develop a similar model of task cre-
ation, but combine it with a process of task replacement. Here is a simplifi-
cation.
 36

Instead of ranging from 0 to N , task indices i now range from N − 1

to N . Capital is equally productive at all tasks it can perform, but labor
productivity at task i is Bi = Di (1 − β)(1−ρ)/ρ , where Di = exp(gD i), for
some gD > 0. In an exogenous growth setting, both β ∈ (N − 1, N ) and N
grow over time at a constant exogenous absolute rate—let us say, without
loss of generality, at one unit per unit time. The fraction of tasks not au-
tomatable is thus constant at N − β, but the productivity of human labor at
the non-automatable tasks grows at exponential rate gB = gD . With enough
saving, all automatable tasks are automated. Output and capital grow at
rate gD + gL , in line with the effective labor supply, and wages grow at rate
gD . The labor share is constant, as in any CES model with labor-augmenting
technology growth and sufficient saving.
Moving from asymptotic automation in the original setting of §4.2 to this
model of task automation, creation, and replacement thus increases the out-
put growth rate iff gD > g1−β (1 − ρ)/ρ. As usual, if we imagine starting
from a world without task creation and replacement, introducing this pro-
cess raises the growth rate from 0 to a positive number; and if we imagine
starting from a world with some other source of exogenous growth in labor
productivity, introducing this process raises the growth rate, given enough
saving to maintain the automation condition.
This model is nearly equivalent to a task-based model in which the
task-range is fixed at the unit interval, β is fixed, and B grows exogenously
at rate gD . Its framing is motivated by the empirical observation that we
have long seen the automation of existing tasks go hand-in-hand with the
creation of new, high-productivity tasks that, at least temporarily, only
humans can perform (Goldin and Katz, 2009; Acemoglu and Autor, 2012),
and continue to see this pattern in the present (Autor, 2015; Acemoglu and
Restrepo, 2019b). The result is a near-complete turnover of job types over
time, rather than a mere encroachment of automation onto human territory.
In this sense, this promises to be a more realistic model of automation than
one without task replacement. As we have just seen, more realistic models
of this type are also compatible with balanced growth.

This balanced growth result is, however, sensitive to the assumption that ad-
vances in automation technology (increases in β) and task creation (increases
in N ) proceed at the same rate.
If task creation oustrips automatability, the labor share rises, and as
β − N + 1 → 0 asymptotically, the labor share rises to 1. In this case, we
 37

approach a state in which labor performs all tasks. Capital is relegated to

an ever-shrinking band of the lowest-labor-productivity tasks. Since output
and, given a constant saving rate, capital grow at the same rate as effective
labor, while capital is used ever less efficiently, capital rents fall and the
capital share falls to 0. In equilibrium, output grows at rate gD + gL , and
wages grow at the labor productivity growth rate gD , as before.
Now suppose that automatability outstrips task creation, and in partic-
ular that it does so at a constant rate gN −β < 0. What follows depends
on the extent to which capital accumulation keeps up with this process. If
s ≤ gD + gL , the automation condition will not be met in the long run,
and capital and effective labor will be perfect substitutes on the margin.
Output will thus equal Kt + Dt Lt , and the stock of capital grows at the
same rate as that of effective labor when s(Kt + Dt Lt )/Kt = gD + gL ,
i.e. when Kt /(Kt + Dt Lt ) = s/(gD + gL ). In the long run we thus have
gK = gY = gD + gL , the labor share and the fraction of tasks not automated
approach 1 − s/(gD + gL ) (ranging from 1 at s = 0 to 0 at s = gD + gL ), and
wages grow at rate gD .
Now suppose that s ∈ (gD + gL , gD + gN −β (1 − ρ)/ρ + gL ], so that capital
accumulation outpaces labor and labor productivity growth in isolation but
not in combination with automation. Now the fraction of tasks automated
grows over time: if it stayed constant, capital per automated task would
ultimately exceed effective labor per non-automated task, since s > gD + gL ,
and it would be profitable to reallocate some capital to automatable but
non-automated tasks. But the fraction of tasks automated does not catch up
with the automatability frontier: if all automatable tasks were automated,
effective labor per non-automated task would ultimately exceed capital per
automated task, since s < gD +gN −β (1−ρ)/ρ+gL , and it would be profitable
to reallocate some labor to currently automated tasks. Thus the automation
condition is not met in this scenario either, and capital and effective labor
are still perfect substitutes on the margin. Since the stock of capital grows
more quickly than that of effective labor, in the long run output grows at
rate s, wages grow at rate gD , and the labor share again falls to 0.
Finally, if s > gD + gN −β (1 − ρ)/ρ + gL , the automation condition is met.
Growth proceeds at gD + gN −β (1 − ρ)/ρ + gL and the labor share approaches
a positive constant, as in the model of §4.2 with direct labor-augmenting
technology growth gD > 0. Empirically the saving rate is currently far higher
than the growth rate of effective labor, so unless automation accelerates
dramatically, this is the most relevant case for consideration.
 38

Now let us briefly and informally consider the implications of endogenizing

the automation technology and task creation processes. Suppose that, in
addition to the labor force, there is a pool of researchers who allocate their
efforts between increasing β and increasing N . Upon doing either, they earn
a patent right to some of the gains that result.
This scenario, as detailed by Acemoglu and Restrepo (2018b), produces
intuitive equilibrating pressures, suggesting that we might expect to observe
automation technology and task creation proceeding at the same rate, with-
out having to assume this ad hoc. Excessive development of automation
technology results in tasks that are automatable but not automated, be-
cause of an insufficient ratio of capital to effective labor. This eliminates
the immediate value of further automation technology. Excessive task cre-
ation, on the other hand, increases the value of automation technology, by
inefficiently relegating capital to a narrower range of tasks.
The full range of possibilities here, however, is essentially the same as in
the exogenous growth case. The proportion of tasks automated can fall to 0,
if the saving rate is sufficiently low; there can be asymptotically complete au-
tomation if the saving rate is sufficiently high; and there is partial automation
in intermediate cases. The primary novelty of the endogenous research case
is that here, which case obtains can depend on the researchers’ productivity
at developing automation technology relative to their productivity at task
creation. In particular, increases in productivity at developing automation
technology, relative to productivity at task creation, can increase the equi-
librium automation rate and decrease the equilibrium labor share. Also, the
growth rate here is not exogenous but depends on the level of productivity
at both researcher tasks, as well as on the size of the researcher population.

5 AI in technology production
Throughout the discussion so far (except for the brief note at the end of the
previous section), technological development has been exogenous, when it has
appeared at all. Even in this circumstance, developments in AI have proved
capable of delivering transformative economic consequences. When techno-
logical development is endogenous, and in particular when more advanced
AI can allow it to proceed more quickly, the resulting process of “recursive
self-improvement” can generate even more transformative consequences.
 39

5.1 Learning by doing

This recursive effect can be seen most simply in a model with Cobb-Douglas
production. Let us interpret the production as task-based, with fraction a of
tasks automated, as in §4.1. The model below is inspired by the exploration
of learning by doing in Hanson (2001).
The labor supply grows at exogenous rate gL > 0, but capital productivity
growth proceeds endogenously. We will use a modification of the endogenous
growth model presented in §2.4. In that model, technology growth is a func-
tion of existing technology and “researcher effort”. Here, instead, we will not
introduce research and will say simply that capital productivity grows as a
function of the existing technology and output. That is,

Yt = (At Kt )a L1−a
t , (40)

where
Ȧt = θAϕt Ytλ ⇒ gA,t = θAϕ−1
t Ytλ (41)
for some ϕ < 1 and λ > 0. One might interpret this as a model in which
the production process itself contributes to the generation of productivity-
increasing ideas.
Given a constant saving rate, gK = gY . So, from our production function,
a
gY = a(gA + gY ) + (1 − a)gL ⇒ gY = gA + gL . (42)
1−a
From our formula for gA,t , the steady state of gA (if it exists) will be that
which satisfies (ϕ − 1)gA + λgY = 0. Substituting for gY in this expression
and solving for gA , we have

λ(1 − a)
gA = gL . (43)
(1 − a)(1 − ϕ) − λa

This exponential growth path will exist as long as the denominator is positive:
that is, as long as
1−ϕ
a< . (44)
1−ϕ+λ
 40

In this case, output growth will be given by26

(1 − a)(1 − ϕ)
gY = gL . (45)
(1 − a)(1 − ϕ) − λa

Otherwise, the recursive process by which proportional increases to At gener-

ate proportional increases to Yt , which in turn generate proportional increases
to At+1 (using discrete-time notation for clarity), results in the proportional
increases at t + 1 being larger than those at t. The growth rates of A and Y
thus increase without bound.
The transformative potential of automation is now straightforward. In-
creases in a increase the long-run growth rate without bound, as a approaches
(1 − ϕ)/(1 − ϕ + λ). Past this threshold, increases in a trigger a growth ex-
plosion.
The growth explosion type can be determined by substituting (40) into
expression (41) for gA,t :
(1−a)λ
gA,t = θAtϕ−1 Ytλ = θAϕ−1
t Aaλ aλ
t K t Lt . (46)
a/(1−a)
Since27 Kt ∝ Yt = (At Kt )a Lt1−a , rearranging gives us Kt ∝ At Lt . Sub-
stituting for Kt into (46), we have
ϕ−1+λa/(1−a)
gA,t ∝ θAt Lλt . (47)

When a > (1 − ϕ)/(1 − ϕ + λ), the exponent on At is positive, producing a

Type II growth explosion. When a = (1 − ϕ)/(1 − ϕ + λ), we have gA,t ∝ Lλt ;
the technology growth rate itself grows asymptotically at rate λgL , producing
a Type I growth explosion.

5.2 Automated research

Cockburn et al. (2019) taxonomize AI systems as belonging to three broad
categories: symbolic reasoning, robotics, and deep learning. Symbolic
reasoning systems, they argue, have proven to have few applications.
26
Note that, in this case, effective capital growth gA + gY will always equal gL ((1 −
a)(1 − ϕ) + λ(1 − a))/((1 − a)(1 − ϕ) − λa) > gL . With effective capital growing more
quickly than labor, the automation condition will always eventually be met for any fixed
a.
27
The “∝” symbol means “is asymptotically proportional to”.
 41

Robotics—by which they broadly mean capital that can substitute for
human labor in various ways, instead of complementing it—has of course
had many applications. It has also been the subject of a substantial
majority of the theoretical literature on the economics of AI, including all
that discussed in this survey so far. They propose however that the most
transformative possibilities come from deep learning systems, by which they
mean systems that can learn as human researchers can: artificial systems
that can participate directly in the process of technological development.
Citing Griliches’s (1957) discussion of the implications of “inventing a
method of invention”, they argue that deep learning systems (in their use
of the term) will have qualitatively more radical consequences than mere
robotics. As we will see, this appears to be correct.

Following Aghion et al. (2019), let us focus directly on the implications of

technology production by using an even simpler production function than
usual:
Yt = At (1 − S)Lt . (48)
Labor L is the only factor of production. S is the constant proportion of
people who work in research as opposed to final good production. Output
technology A, however, is developed using a CES function of both labor and
capital K. Building on the standard technology production function from
§2.4 (where Ȧt = θAϕt (St Lt )λ ), we have
Ȧt = Aϕt [(Ct Kt )ρ + (Dt SLt )ρ ]λ/ρ , ρ ̸= 0; (49)
Ȧt = Aϕt (Ct Kt )λa (Dt SLt )λ(1−a) , ρ = 0,
for some permitted values of ρ, λ, and, in the Cobb-Douglas case, a.
Ct and Dt denote capital- and labor-augmenting technology levels re-
spectively, in the research context. Including them removes the need for a θ
coefficient. The inclusion of factor-augmenting technology terms is unusual
in a technology production function, and perhaps somewhat unsatisfying, as
it amounts to introducing an explicit technology production function in the
final good sector only to leave technology growth in the new “research” sector
potentially unexplained. That said, one must introduce a wrinkle along these
lines to study the implications of the (exogenous) asymptotic automation of
research tasks—keeping in mind the Aghion et al. (2019) result, summarized
in §4.2, that if ρ < 0, asymptotic automation can amount to growth in what
is here denoted D. The implications of growth in both C and D, and across
 42

all values of ρ, have then been included for generality. To explore the case
in which no technology growth is exogenous, simply posit that C and D are
constant throughout the discussion below.
As usual, we will assume a constant saving rate, so that capital accumu-
lation in the long run tracks output.

Suppose ρ < 0. Recall that in this case sustained growth in effective capital
and in effective research labor are both necessary to sustain growth in output
technology, and growth will be driven by whichever factor grows more slowly.
Observe that when growth is constrained by effective capital accumulation,
we have gA ∝ Atϕ−1 (Ct Kt )λ , and that when it is constrained by effective
research labor growth, we have gA ∝ Aϕ−1 t (Dt Lt )λ .
Regarding ϕ:
• Recall from the reasoning of §2.4 that, if ϕ < 1, we have gA = λ(gC +
gK )/(1−ϕ) on the capital-constrained path and gA = λ(gD +gL )/(1−ϕ)
on the labor-constrained path. In the former case, since gK = gY = gA +
λ
gL , we can substitute for gK and rearrange to get gA = 1−ϕ−λ (gC + gL ).
A capital-constrained path with this growth rate exists when ϕ < 1−λ.
λ λ
Thus, if ϕ < 1 − λ, gA = min( 1−ϕ−λ (gC + gL ), 1−ϕ (gD + gL )). A
one-time increase to Ct , Dt , Lt , or S, as long as S remains below 1, does
not affect the output technology growth rate. A permanent increase
to gC , gD , or gL , on the other hand, does increase the growth rate of
output technology and thereby output per capita.
• If ϕ ∈ (1 − λ, 1),28 output technology growth cannot be constrained by
capital accumulation; such a scenario would imply gA ∝ Aϕ−1t (Ct Kt )λ ∝
Aϕ−1+λ
t (Ct Lt )λ , contradictorily producing superexponential growth in
λ
output technology, output, and capital. We have gA = 1−ϕ (gD + gL ).
• If ϕ = 1, suppose labor remains fixed at L and labor-augmenting tech-
nology remains fixed at D, while effective capital accumulates. In the
long run we then have gA = DSL. A one-time increase to D, S, or
L increases the growth rate of output technology and thereby output
(again, as long as S remains below 1). If we begin from a state in
which gD = gL = 0 and introduce positive labor or labor-augmenting
technology growth, the result is a Type I growth explosion.
28
The dynamics of the knife-edge ϕ = 1 − λ case are somewhat complex and will be
omitted for clarity.
 43

• If ϕ > 1, we have a Type II growth explosion regardless of the other

parameters, as explained in §2.4.
As noted above: recall from §4.2 that, given ρ < 0, increases to D can be
interpreted as increases to the fraction of research tasks that have been
automated.

Suppose ρ = 0. Technology growth is then

gA,t = Aϕ−1
t (Ct Kt )λa (Dt SLt )λ(1−a) (50)
∝ emt ,
where
m ≜ gA (ϕ − 1) + λa(gC + gK ) + λ(1 − a)(gD + gL ). (51)
(Though there is no conceptual distinction between capital- and labor-
augmenting technology in the Cobb-Douglas case, both variables have been
retained, for easier comparison with the other cases.)
From our assumption of a constant saving rate, gK = gY = gA + gL . So
m = gA (ϕ − 1 + λa) + λagC + λ(1 − a)gD + λgL . (52)
Regarding ϕ:
• If ϕ < 1 − λa, we will have, in equilibrium, the constant output tech-
λ
nology growth rate that sets m = 0. This is gA = 1−λa−ϕ (agC + (1 −
a)gD + gL ). One-time increases to C, D, S, or L do not change the
growth rate, but increases to gC , gD , or gL do.
• If ϕ = 1 − λa, we have steady growth only if C, D, and L are in the
long run constant, since we are assuming that these growth terms are
all nonnegative. Fixing A0 = K0 = 1, the output technology growth
rate is C λa (DSL)λ(1−a) . A one-time increase to C, D, S, or L increases
the growth rate. If C, D, or L grow unboundedly, we have a Type I
growth explosion.
• If ϕ > 1 − λa, we have a Type II growth explosion regardless of the
other parameters.
Recall from §4.1 that increases to a can be interpreted as increases to the
fraction of research tasks that have been automated. They can thus induce
Type I and Type II growth explosions, if ϕ ∈ (0, 1).
 44

Suppose ρ > 0. Recall that in this case sustained growth in effective capital
or in effective research labor suffice to sustain growth in output technol-
ogy, and growth will be driven by whichever factor grows more quickly.
Now, when growth is driven by effective capital accumulation, we have
gA ∝ Atϕ−1 (Ct Kt )λ , and when it is driven by effective research labor growth,
we have gA ∝ Aϕ−1t (Dt Lt )λ .
Regarding ϕ:
• If ϕ < 1 − λ, the capital- and labor-driven technology growth rates
λ λ
equal 1−ϕ (gC + gK ) and 1−ϕ (gD + gL ), respectively. In the former case,
since gK = gY = gA + gL , we can substitute for gK and rearrange to get
λ λ λ
gA = 1−ϕ−λ (gC + gL ). Thus gA = max( 1−ϕ−λ (gC + gL ), 1−ϕ (gD + gL )).
Growth rate increases require increases to gC , gD , or gL .
• If ϕ > 1 − λ,29 we have a Type II growth explosion regardless of the
other parameters.

An intuition for these results is as follows. As explained briefly in §2.4, the

growth of some variable X exhibits a Type II growth explosion if its growth
rate takes the form gX ∝ X ψ for some ψ > 0. When ρ < 0, capital accumu-
lation cannot accelerate technological development, which is bottlenecked by
its slower-growing factor, namely effective labor. Output technology growth
is then gA ∝ Aϕ−1 , so the Type II growth explosion requires ϕ > 1. When
ρ > 0, on the other hand, capital accumulation at rate gY = gA effectively
multiplies gA by a factor of Aλ , so gA ∝ A1−ϕ+λ . The Type II growth explo-
sion therefore requires only ϕ > 1 − λ.
Note that our analysis of the ρ > 0 case also covers the case in which the
development of output technology is fully automated. Simply use ρ = 1.
It also covers a common interpretation of the possibility of “recursive self-
improvement”. If A represents cognitive ability, and enhanced intelligence
(human or artificial) speeds the rate at which intelligence can be improved
such that gA,t ∝ Atϕ−1 Ktλ , then explosive growth obtains iff ϕ > 1 − λ (since,
again, K grows in line with A). If we remove capital accumulation entirely
and say that the development of intelligence depends only on the intelligence
level, such that gA,t ∝ Atϕ−1 , then explosive growth obtains iff ϕ > 1.
29
As under ρ < 0, the ϕ = 1 − λ case has been omitted for simplicity.
 45

5.3 AI assistance in research

In §4, we discussed several papers which use a microfoundation of the
output production function as a basis for exploring the implications of a
certain kind of automation. Somewhat analogously, Agrawal et al. (2019)
use Weitzman’s (1998) microfoundation of the process of technological
development as a basis for exploring the implications of a certain way in
which advances in AI might assist in technological development.

Let Y = A(1−S)L, as before, and hold S fixed but posit labor growth gL > 0.
Given A existing “technological ideas”, a researcher has access to only Aϕ ,
for some ϕ ∈ (0, 1), perhaps due to some sort of cognitive limitation.30 New
ideas are made from combinations of existing ideas. Given access to Aϕ
ϕ
ideas, a researcher therefore faces 2A idea-combinations. Of these, not all
can generate new technological ideas, perhaps due to some other sort of
cognitive limitation. Instead, each researcher’s idea-generation function is
“isoelastic” in ideas available:
ϕ
(2A )α − 1
Ȧ = θ , α > 0;
α (53)
ϕ
Ȧ = θ ln(2A ) = θ ln(2)Aϕ , α = 0,

for some θ > 0 and some α ∈ [0, 1].31

Suppose that θ = 0 (or that θ → 0 as A → ∞), and that total research
output is linear in the number of researchers raised to some positive power
λ, as in the growth model of §2.4. Then collective technological development
is given (at least asymptotically) by

λgL
Ȧ = θ ln(2)Aϕ (SL)λ ⇒ gA = . (54)
1−ϕ

This is just the standard Jones model, with a coefficient of ln(2) rescaling θ.
Now suppose that α > 0 (or that α is bounded below by α > 0 as
A → ∞). Then collective technological development is bounded below (at
30
The model requires ϕ > 0 such that the fishing-out effect does not predominate. As
discussed in §2.4, Bloom et al. (2020) estimate ϕ = −2.1.
31
The formula for α = 0 is the limiting case of the formula for α > 0, as α → 0.
 46

least asymptotically) by
ϕ
(2A )α − 1
Ȧ = θ (SL)λ (55)
α
ϕ
(2A )α − 1
⇒ gA = θ (SL)λ .
Aα
It follows from the second term that, for large A, gA increases more than
polynomially in A. That is, gA increases quickly enough in A to produce
a Type II growth explosion. If the above model approximates reality,
therefore, we presumably currently have α = 0.

Given α = 0, let us now consider the potential impacts of artificial research

assistance.
If it allows for a one-time increase to θ, this amounts to a one-time increase
to the supply of effective researchers. This puts us on a higher growth path,
but it does not increase the growth rate or have any other transformative
effects. But if AI assistance improves with time in this way, allowing for θt
to grow at some positive exponential rate, this amounts to an increase in the
growth rate of effective researchers. It can thus increase the growth rate of
technology and thereby output.
If AI assistance allows researchers to access more of the stock of exist-
ing ideas, it amounts to a one-time increase in ϕ. (One might argue that
this is what internet library access and accurate search engines have already
enabled.) As we can see, this increases the growth rate as well.
Most transformatively, if AI tools help researchers search through
the ever-growing “haystacks” of possible idea-combinations for valuable
“needles”, they could permanently increase α. Agrawal et al. (2019) argue
that this is precisely the sort of activity to which AI systems are best suited:
they are already being profitably used to identify promising combinations
of chemicals in pharmaceutical development, for example. (See Agrawal et
al. (2018) for a more thorough defense of this argument.) If this turns out
to hold across the board, the result is stark: as shown above, a permanent
increase to α produces a Type II growth explosion.

Agrawal et al. (2019) also explore the potential impacts of AI assistance

in research teams, rather than in assisting individual researchers. Seeber
et al. (2020) do the same, in the context of a much more applied and less
 47

formal inquiry. Neither analysis appears to reveal channels for transformative

growth effects substantively different from those presented above.

5.4 Growth impacts via impacts on technology invest-

ment
Throughout §5, we have taken technology production to be endogenous, in
the sense that it has required explicit inputs of capital and labor. Never-
theless we have taken the level of investment in technological development—
the fraction S of labor, and (in §5.2) the amount of capital, allocated to
research—to be exogenous. A final way in which AI could have a transfor-
mative impact, therefore, is by changing the levels of investment in, and effort
allocated to, technological development. As we have seen, at least in some
circumstances, these changes can change the growth rate, or can determine
the existence or type of a growth explosion.
This pathway to transformative impact is somewhat analogous to the
possibility, explored in §3.4, that developments in AI could affect the growth
rate by affecting the saving rate, even in an economy without endogenous
technological development. As in that case, this change could in principle be
positive or negative. (Indeed, one way AI could impact the extent to which
resources are devoted to technological development is by affecting the saving
rate, as long as capital is modeled as an input to technology production.) Also
as in that case, to the extent that the literature has explored this pathway
to transformative impact, it has focused on the counterintuitive possibility
that AI slows (though not, here, reverses) growth.
This could take place by accelerating the “Schumpeterian” process of
“creative destruction”. On this analysis, the incentive to innovate comes
from a temporary monopoly that the innovators enjoy, either by patents
or by trade secrets, during which they can extract rents from those who
would benefit by using the new technology in production. AI, however, could
make it easier for competitors to copy innovations. Relatedly, AI could also
ease the rapid development of technologies only negligibly more productivity-
enhancing than those they replace. Because these technologies would entirely
eliminate the markets for those they replace, their rapid development would
curtail the incentive for innovation. In the absence of this incentive, technol-
ogy growth can slow to a halt. This cannot cause output per capita to fall,
at least in most models, but it can cause output per capita to stagnate.
 48

This dynamic is explored more formally by Aghion et al. (2019) in the

context of the model of automated research laid out in §5.2, and by Acemoglu
and Restrepo (2018b) in the context of the model of automation and task
replacement laid out in §4.4. We will not work through it here.
As with the Sachs and Kotlikoff (2012) observation that AI can do dam-
age by lowering the saving rate, the insight here is not primarily an insight
about artificial intelligence. It is primarily a special case of the well-known
fact, mentioned briefly in §2.4, that though free and competitive markets
can generally be expected to appropriately compensate production factors
for a final good in a static setting, the same cannot be said about the in-
puts to technological development. Policymakers interested in growth must
always consider the impact of structural economic changes on the incentives
for technological innovation, therefore, and must adjust their funding or sub-
sidization of basic research in light of such changes as they unfold.

6 AI in both good and technology production

Naturally, the effects of AI are most transformative of all when it allows
capital to better substitute for labor in both good production and technol-
ogy production. Unfortunately, this pair of circumstances has been studied
even less extensively than the effects of AI in each sector separately. Nev-
ertheless, an analysis that begins with the research automation of §5.2, but
replaces the labor-only production function with a CES one, proves relatively
straightforward.
Suppose we replace the labor-only final good production function, (48)
from §5.2, with a CES production function in capital and labor. Let the
substitution parameter in the final good sector be denoted ρY , and that
in the research sector be denoted ρA . We will ignore factor-augmenting
technology in the good production function; output technology A will be
thought of as augmenting both. We will assume a constant and sufficient
saving rate s, and fractions of capital and labor used in the technology
sector—SK and SL respectively—strictly between zero and one.

We can now consider the growth regimes that obtain under different values
of ρY and ρA . For simplicity, we will not allow for labor growth or exogenous
sources of technology growth. To begin, let us list the cases we have already
implicitly covered.
 49

If ρY < 0, little changes from the case of §5.2. Output is still bottle-
necked by the scarce factor, namely labor. Output therefore asymptotically
resembles A(1 − SL )L, as before. How technology evolves depends on ρA , as
covered in §5.2.
If ρY = 0 but ρA < 0, we are in the well-worn territory of Cobb-Douglas
production—so, given capital depreciation, production per capita that grows
with technology—and technology that grows sub-exponentially (unless re-
search labor inputs grow exponentially).
If ρY > 0 but ρA < 0, we reach the Type I growth explosion discussed in
§3.2. Output in the absence of growth in A grows at rate sA, but A grows
without bound, even without growth in research labor inputs.

If ρY = ρA = 0, however, the ability of capital to contribute to both good

production and technology production generates possibilities we have not yet
considered. As we will see, the resulting growth path is highly sensitive to
the other parameters.32
In particular, let

Yt = At ((1 − SK )Kt )a , (56)

K̇t = sYt , and (57)
Ȧt = θAϕt (SK Kt )λ , (58)

where SK ∈ (0, 1), a ∈ (0, 1), s > 0, θ > 0, ϕ < 1, and λ > 0. Also, define
λ
γ≜ . (59)
(1 − a)(1 − ϕ)
Then
• If γ > 1, Y exhibits a Type II growth explosion.
• If γ = 1, Y grows exponentially, with
 θ λ 1−a
λ SK  1+λ−a
lim gY,t = s 1+λ−a . (60)
t→∞ 1 − a (1 − SK )a
• If γ < 1, Y grows power-functionally.
Note that the production function of (56) is Cobb-Douglas with an implicit
constant labor stock normalized to 1, and/or a constant land stock also
32
What follows is in essence an elaboration on Aghion et al. (2019), §4.1, Example 3.
 50

normalized to 1, and (given CRS) the exponents on labor and land summing
to 1 − a. Note likewise that technology production, as described by (58),
may be interpreted as Cobb-Douglas with inputs other than capital fixed.
A proof of the above can be found in Appendix A.2, but an intuition for
the exponential growth threshold provided by γ = 1 is as follows. If growth
in A and Y were driven by exogenous exponential growth in K, we would
have, in steady state,
λ
gA = gK (61)
1−ϕ
and thus
 λ 
gY = + a gK . (62)
1−ϕ
But gK is not exogenous: future growth in K roughly equals past growth in
Y , since capital accumulation is driven by saving a proportion of output. If
λ
1−ϕ
+ a > 1, therefore, a given growth rate in K generates a higher growth
rate of Y , and this higher growth rate is subsequently exhibited by K. The
λ
growth rate of K therefore grows over time. Likewise, if 1−ϕ + a < 1, a
given growth rate in K generates a lower growth rate of Y , and this lower
growth rate is subsequently exhibited by K. The growth rate of K there-
λ
fore falls. Finally, observe that 1−ϕ +a > 1 iff γ > 1, and likewise for < and =.

If we replace the ρY = ρA = 0 model with one in which (a) ρY > 0 and/or

(b) ρA > 0, nothing changes except that, respectively,
(a) the exponent on capital in good production effectively rises from a to
what, in a fully specified Cobb-Douglas model, would have been the sum
of the exponents on capital and labor in good production; and/or

(b) the exponent on capital in idea production effectively rises from λ to

what, in a fully specified Cobb-Douglas model, would have been the sum
of the exponents on capital and labor in idea production.
Let us denote these new exponents ã and λ̃.
In the absence of a natural resource constraint, ã = 1, by the assumption
of CRS good production. Since, from (59), we have lima↑1 γ = ∞, it follows
that, absent significant natural resource constraints, ρY > 0 and ρA ≥ 0
always produce a Type II growth explosion.
 51

By contrast, we do not in general assume that λ̃ = 1, i.e. that research

outputs exhibit constant returns to scale in research inputs. Furthermore,
even if we did, a value of λ = 1 is not sufficient (or, for that matter, necessary)
for γ > 1. An assumption of ρA > 0 therefore has no qualitative implications
beyond those of the ρA = 0 case.

7 Overview of the possibilities

The table below summarizes the transformative scenarios we have considered.
They have been rearranged slightly for clarity, and some near-redundant pos-
sibilities have been removed, but they primarily follow the order in which they
are presented in §3–6. Relevant literature is cited below each scenario. Note
that, in keeping with the presentation so far, the cited literature introduces
the models that allow for the scenarios in question, but does not always
discuss the transformative scenarios on which we have focused.
We have not considered all possible AI scenarios, as this table makes
clear. Nevertheless we have hopefully sampled the possibilities thoroughly
enough that the reader is now comfortable filling some of the gaps.
 52

Human
labor Human
Scenario33 Growth34 share35 wages36
LS in production & = →1 +
capital-augmenting tech growth
§3.1 Acemoglu and Restrepo (2018a)

HS in consumption goods ++ C ++
§3.2 Nordhaus (2021)

HS in production ++ →0 ++
§3.2 Nordhaus (2021)

HS (not PS) in production & I →0 I

capital-augmenting tech growth
§3.2 Nordhaus (2021)

PS in production & I →0 L
capital-augmenting tech growth
§3.2

PS in production, capital-augmenting ++ →0 →0
tech growth, & MS land constraint

33
“PS”, “HS”, “MS”, and “LS” stand for perfect, high, moderate, and low substitutabil-
ity, and refer to substitution parameters ρ = 1, > 0, = 0, and < 0 respectively. Unless
otherwise noted, the “HS” case allows for perfect substitutability. In the scenarios with
endogenous research, “negative”, “positive”, “low”, “intermediate”, and “high [research]
feedback” refer to research feedback exponents ϕ < 1 − λ, > 1 − λ, < 1, = 1, and > 1
respectively.
34
+ and – refer to cases in which AI shifts the output path up or down without changing
the growth rate, e.g. by increasing or decreasing the plateau level in a circumstance where
output plateaus regardless of AI. --, ++, I, and II refer to cases in which AI allows for
decreases to the long-run growth rate, increases to the long-run growth rate, Type I growth
explosions, and Type II growth explosions. = refers to cases in which AI does not change
the long-run output level or growth rate.
35
C means that AI pushes the human labor share to some positive constant, not neces-
sarily lower or higher than the value it would take in the absence of AI.
36
L means that human wages are driven to some low but constant rate (typically the
rental rate of effective capital). C means that they are pushed to some positive constant,
not necessarily lower or higher than they would be in the absence of AI. All other symbols
are defined as in the Growth column.
 53

§3.3 Hanson (2001)

PS in production, equipment-augmen- ++ →0 L
ting tech gr., & MS land constraint
§3.3

PS in production & LS land constraint = →0 →0

(regardless of tech)
§3.3 Korinek and Stiglitz (2019)

HS in final good production, HS in ++ →0 L

robotics production
§3.4 Mookherjee and Ray (2017), Korinek and Stiglitz (2019)

HS in final good production, LS in + C C

robotics production
§3.4 Mookherjee and Ray (2017), Korinek (2018)

PS in production & one-off ++ →0 =

capital-augmenting tech increase →
saving increase
§3.5 Korinek and Stiglitz (2019)

PS in production & capital-aug. tech − →0 →0

growth → saving decrease
§3.5 Sachs and Kotlikoff (2012), Sachs et al. (2015)

MS in production & asymptotic or full I →0 I

task automation
§4.2 Aghion et al. (2019)

LS in production & asymptotic task ++ C ++

automation
§4.2 Aghion et al. (2019)

LS in production & task automation ++ C ++

and replacement
§4.4 Acemoglu and Restrepo (2018b)

HS in production & task automation I →0 I

and creation
 54

§4.3 Hémous and Olsen (2014)

Learning by doing, w/intermed. ++

feedback and/or automation
§5.1 Hanson (2001)

Learning by doing, with suffic. II

feedback and/or automation
§5.1 Hanson (2001)

LS in tech production, low research ++

feedback, & asymptotic research task
automation; or HS in tech production,
negative research feedback, & research
capital productivity growth
§5.2 Aghion et al. (2019)

LS in tech production, intermed. I

research feedback & asymp. research
task automation; or HS in tech prod.,
zero research feedback, & research
capital productivity growth
§5.2 Aghion et al. (2019)

LS in tech production & high research II

feedback or HS in tech production &
positive research feedback
§5.2 Aghion et al. (2019)

AI-assisted multiplication of ++
combinatorial idea discovery
§5.3 Agrawal et al. (2019)

AI-assisted elasticity-change in idea II

discovery
§5.3 Agrawal et al. (2019)

AI-diminished innovation incentives −−

§5.4 Aghion et al. (2019), Acemoglu and Restrepo (2018b)
 55

HS in production & MS or HS in idea II

production (for any value of research
feedback)
§6 Aghion et al. (2019)

The human labor share and wage are technically undefined in the models
of endogenous technology production, since, as noted in §2.4, we cannot
straightforwardly assume that the factors of technology production will tend
to be paid their marginal products (or anything else in particular). As often
presented, however, human labor is the lone factor of final good production
in these models, and the technology being produced is labor-augmenting.
Taken literally, therefore, the wage rate in these models should grow in line
with technology and so with output. That is, it should exhibit growth rate
decreases, increases, Type I growth explosions or Type II growth explosions
as listed above.

8 Conclusion
The set of models discussed here cover a wide range of AI’s possible long-
run macroeconomic impacts. It can hopefully serve as a bridge between
the tools of economics—whose use is typically restricted to shorter-term and
smaller-scale possibilities (but need not be)—and the longer-term and larger-
scale questions posed by futurists, who typically do not draw on the tools of
economics (but could, we believe, sometimes learn from doing so).
Nevertheless, of course, many topics relevant to the economics of AI, and
even of transformative AI, could not be covered here.

Wage distribution is a—perhaps even the—central concern of the literature

on the economics of AI, including much of the literature cited here. It is like-
wise a central concern of the less long-term-focused reviews of the economics
of AI cited in §1. Indeed, wages and skill levels are of course empirically
highly unequal. And this inequality has indeed increased in the recent past,
a development many attribute to the rise of automation. Nevertheless, we
have consistently referred to all wages and human abilities as homogeneous.
The choice to focus on average wages and on the overall labor share is
motivated in part by the supposition that, if we are truly considering the
 56

long run, the likeliest transformative possibility is that AI will outsmart us

all. In this event, human talents will not save us; if we retain positive wages
or a positive labor share, we will do so only because AI is put to use making
us more productive, or because some tasks, like those performed by clergy or
by hospice nurses, remain resistant to automation. Otherwise, as Freeman
(2015) colorfully puts it, “[w]ithout ownership stakes, workers will become
serfs working on behalf of robots’ overlords”.
To be clear, however, this view may by all means be incorrect. We
cannot rule out AI-induced scenarios in which, even in the long run, income
is concentrated not entirely in the hands of the robot owners but also at
least to some extent in the hands of the most skilled human individuals.

The subject most conspicuously present, despite receiving relatively little

direct attention in economics literature, is the possibility of lasting changes
to the growth regime: growth rate shifts and Type I and II growth explosions.
As this review perhaps illustrates, the lack of attention currently given to
these possibilities is not a necessary consequence of all plausible economic
modeling. Rather, it is the result of a widespread norm of focusing only on
model scenarios in which long-run growth is constant. Even Aghion et al.
(2019), who take the singularitarian growth potential of AI most seriously,
focus less on scenarios in which labor and capital are highly substitutable in
technology production on the grounds that, as long as ϕ > 0, “in this case
researchers are not a necessary input and so standard capital accumulation
is enough to generate explosive growth. This is one reason why the case of
ρ < 0 . . . is the natural case to consider.” Expressed motivations along these
lines appear throughout the literature.
Even outside discussions of AI, it is rare for economic growth litera-
ture to consider substantial and permanent growth rate increases, let alone
growth explosions. This is presumably because such models would violate
the “Kaldor fact” (1957) of constant exponential per-capita growth at 2–3%
per year, which has roughly held in the industrialized world since roughly
the Industrial Revolution. But on a longer timeframe, models of increasing
growth would not be ahistorical; the growth rate was far lower before the
Industrial Revolution, and before the Agricultural Revolution it was lower
still. Empirical forecasts on the basis of these longer-run facts commonly
predict radical future increases to growth, including substantial one-time
rate increases and Type I and II growth explosions (see e.g. Hanson (2000)
and Roodman (2020)). Some of the models underlying such forecasts lack
 57

economic foundations (e.g. that in Hanson (2000)), making it difficult to as-

sess whether the forces driving historical superexponential growth are still at
work today. Notably, however, some have economic foundations (e.g. that in
Kremer (1993)) that imply that continued increases to the pool of effective
workers or researchers, as advances in robotics and AI would permit, would
continue the superexponential process. For a more thorough survey of the
singularitarian literature, at least as of 2013, see Sandberg (2013).
In short, if there were a sufficiently compelling theoretical reason
to believe that the observed long-run trend of increasing growth will
soon permanently halt, then we should of course dismiss transformative
growth scenarios. But as many models of the economics of AI confirm,
there is no shortage of mechanisms, once we allow ourselves to look for
them, through which advances in automation could have transformative
growth consequences. Futurists and economists interested in the long term
might therefore do well to collaborate more on this point of common interest.

Finally, the subject most conspicuously absent here that features most heav-
ily in futurist discussion about AI is the most transformative macroeconomic
possibility of all: the risk of an AI-induced existential catastrophe (see e.g.
Bostrom (2017)). Unlike the possibility of transformative growth effects,
AI risk appears to be absent from the economics literature not primarily
“by choice” but because there is no particularly obvious mechanism through
which accelerating automation or capital productivity, within existing models
of production or growth, can pose a danger.
It is possible, of course, to write down economic models in which produc-
tion and/or technological development pose catastrophic risks in the abstract,
as e.g. Jones (2016) and Aschenbrenner (2020) have done. As outlined above,
however, the only growth-slowing AI possibilities economists have considered
to date are those mediated by impacts on saving (Sachs and Kotlikoff, 2012)
and on innovation incentives (Aghion et al., 2019; Acemoglu and Restrepo,
2019). These scenarios are very far from those that motivate most concern
about AI risk. The latter typically feature superintelligent agents, with goals
not fully aligned with ours, who take control of the world.
The tools of economics can shed at least some light on these concerns as
well. Most simply, to the extent that AI development poses such a risk, AI
safety is a global and intergenerational public good. Through that lens, much
of the analysis of public goods, and in particular many of the tools developed
by environmental economists for the pricing and provision of climate risk
 58

mitigation, could apply to AI safety.

More subtly, to the extent that AI risk arises from AIs’ ability to control
resources independently of human input, models in which the human labor
share remains positive and significant should give us comfort. If human
work remains a bottleneck to growth—say, if AI accelerates growth only by
giving human workers instructions which they must physically perform—then
humanity can in principle impoverish any robot overlords by going on strike.
More worrying are models in which a unit of capital can grow, do research
into capital-augmenting technology, and recursively self-improve all without
human input.
A thorough analysis of the links between the economics of AI and the
issue of AI safety remains an important topic for further exploration.
 59

A Proofs
A.1 Asymptotically positive fractions of capital and
robotics used in robotics production
We will work within the framework of §3.4.
Consider a time t at which SR,t > 0, and let m > 1 denote Kt′ /Kt for some
t > t. From t to t′ , the capital input to robotics production is multiplied
′

by mSK,t′ /SK,t . Because f (·) is CRS, to maintain the condition that fL,t′ =
fL,t = 1/D the labor input to robotics production must also be multiplied
by mSK,t′ /SK,t , and robotics production will then also be multiplied by this
factor. We thus have

H + DSR,t′ Rt′ = (H + DSR,t Rt )mSK,t′ /SK,t and (63)

Rt′ = Rt mSK,t′ /SK,t . (64)

Because both inputs to robotics production are multiplied by a common

quantity, fK is constant across periods. It follows that if the capital input
to final good production grows proportionally more (less) than the labor
input, the marginal productivity of capital in final good production falls
(rises), and the marginal contribution of capital to final good production via
robot production rises (falls). Thus, to maintain the condition that capital
is allocated efficiently, the capital and labor inputs to final good production
must be multiplied by a common quantity across periods:
1 − SK,t′ Rt′ 1 − SR,t′
m = . (65)
1 − SK,t Rt 1 − SR,t

Substituting (64) into (63) and (65) and solving for SK,t′ and SR,t′ , we find
that, as m → ∞,

∗ DRt (1 − SR,t )
SK,t′ → SK ≜ SK,t and (66)
DRt (1 − SR,t ) − (1 − SK,t )H
SR,t′ → SR∗ ≜ SR,t + H/(DRt ). (67)

SK and SR are thus asymptotically constant and nonzero. Furthermore,

since Rt = fK,t SK,t Kt +(H +SR,t DRt )/D, we must have H +SR,t DRt < DRt .
∗
It follows that SK and SR∗ are strictly less than 1.
 60

A.2 Growth paths given Cobb-Douglas production

and research
As in §6, suppose

Yt = At ((1 − SK )Kt )a , (68)

K̇t = sYt , and (69)
Ȧt = θAϕt (SK Kt )λ , (70)

where A0 > 0, K0 > 0, SK ∈ (0, 1), a ∈ (0, 1), s > 0, θ > 0, ϕ < 1, and
λ > 0, and where (68)–(70) are defined for t ∈ [0, ∞)—or, if the system
exhibits a Type II growth explosion at some time t∗ , for t ∈ [0, t∗ ).
Observe first that, for all t,

gKt = s(1 − SK )a At Kta−1 and (71)

λ ϕ−1 λ
gAt = θSK At Kt . (72)

Let ĝK (≜“ggK ”) denote the proportional growth rate of gK itself, and let ĝA
be defined likewise. It then follows from (71) and (72) that, for all t,

ĝKt = gAt + (a − 1)gKt and (73)

ĝAt = (ϕ − 1)gAt + λgKt . (74)

If, for any time τ , ĝKτ > 0 and ĝAτ > 0, then

gAτ + (a − 1)gKτ > 0

=⇒ gAτ > (1 − a)gKτ ; (75)
(ϕ − 1)gAτ + λgKτ > 0
1−ϕ
=⇒ gKτ > gAτ ; (76)
λ
and thus
(1 − a)(1 − ϕ)
gAτ > gAτ (77)
λ
=⇒ γ > 1 (78)

since gAτ > 0 ∀τ by construction.

 61

Likewise, if for any τ we have ĝKτ < 0 (= 0) and ĝAτ < 0 (= 0), then
γ < 1 (= 1, respectively).

For any τ ,

ĝKτ = 0 ⇐⇒ ĝAτ = 0. (79)

The “⇒” direction follows from (73). If ĝKτ = 0, then σgAτ = (1 − a)gKτ ; so
if the right-hand side is constant around τ , so is the left. The “⇐” direction
follows likewise from (74).
Also, ĝK and ĝA are continuous in t wherever they are defined. So by
the intermediate value theorem, if either term is negative at some time and
positive at another time, it must equal zero at an intermediate time. By
(79), we must then have γ = 1.
It follows that, if γ ̸= 1, either

1. ĝKt > 0 and ĝAt > 0 ∀t,

2. ĝKt > 0 and ĝAt < 0 ∀t,

3. ĝKt < 0 and ĝAt > 0 ∀t, or

4. ĝKt < 0 and ĝAt < 0 ∀t,

with case 4 incompatible with γ > 1 and case 1 incompatible with γ < 1.
We will now show that cases 2 and 3 are also incompatible with γ ̸= 1.
Consider case 2. From ĝKt > 0 ∀t, and (73), it follows that

gAt > (1 − a)gKt ∀t. (80)

Recall that, by stipulation, gK always rising and gA is always falling. Thus

{gKt } is bounded above, for instance by gA0 /(1 − a), and {gAt } is bounded
below, for instance by (1 − a)gK0 . By the monotone convergence theorem
for functions, limt→∞ gKt and limt→∞ gAt are defined (and finite, and—since
gK0 = s(1 − SK )a A0 K0a−1 > 0—positive). Let us denote these limits gK ∗
and
∗
gA respectively.
By (73) and (74), it then follows that limt→∞ ĝKt and limt→∞ ĝAt are also
∗
defined (and finite). Since gK and gA∗ are finite and nonzero, as we have just
 62

shown, it must be that limt→∞ ĝKt = limt→∞ ĝAt = 0. Taking the limits of
terms (73) and (74), we then have

gA∗ = (1 − a)gK
∗
and (81)
∗ 1−ϕ ∗
gK = g , (82)
λ A
which jointly imply gA∗ = γgA∗ and thus γ = 1.
Case 3 can be shown to imply γ = 1 by a precisely analogous proof.
Thus γ > 1 implies case 1 and γ < 1 implies case 4.

Suppose γ > 1. By the statements of case 1 and expressions (73)–(74), we

have

gAt > (1 − a)gKt ∀t and (83)

1−ϕ
gKt > gAt ∀t. (84)
λ
By (83), and substituting by expressions (71) and (72),
2
gAt > (1 − a)gAt gKt
= θ̃Aϕt Ktλ+a−1 ∀t, (85)

where

θ̃ ≜ (1 − a)sθ(1 − SK )a SK
λ
. (86)

If the relationship of (83) were an equality at all t, then A would always grow
at precisely the same proportional rate as K 1−a . Noting that

A0 = A0 K0a−1 · K01−a , (87)

we would maintain this ratio between A and K 1−a , with

At = A0 K0a−1 Kt1−a ∀t. (88)

It thus follows from (83) that

At ≥ A0 K0a−1 Kt1−a ∀t (89)

1
=⇒ Kt ≤ K0 Aa−1
0 At
1−a
∀t (90)
 63

(with equality at t = 0 and strict inequality at t > 0). It likewise follows

from (84) that
ϕ−1 1−ϕ
Kt ≥ K0 A0 λ At λ ∀t. (91)

So, if λ + a − 1 ≤ 0, it follows from (85) and (90) that

2 − λ+a−1 ϕ+ λ+a−1
gAt > θ̃A0 1−a
K0λ+a−1 At 1−a
∀t. (92)

Given γ > 1, the exponent on At in (92) is positive. Likewise, if λ+a−1 > 0,

it follows from (85) and (91) that

2 − 1−ϕ (λ+a−1) ϕ+ 1−ϕ (λ+a−1)

gAt > θ̃A0 λ
K0λ+α−1 At λ
∀t. (93)

Again, given γ > 1, the exponent on At in (93) is positive. Either way,

therefore, A grows at worst hyperbolically, and so exhibits a Type II growth
explosion. It follows immediately that Y does as well.

If γ < 1, a proof that A grows at best power-functionally is precisely anal-

ogous, except that it uses inequality (91) in the λ + a − 1 ≤ 0 case and
inequality (90) in the λ + a − 1 > 0 case. By (74) and the case 4 stipulation
that ĝAt < 0 ∀t, we then have
1−ϕ
gKt < gAt ∀t, (94)
λ
implying that K also grows at best power-functionally. Thus Y grows at
best power-functionally as well.
Furthermore, it follows from (70) that if K were constant, A (and thus
Y ) would grow power-functionally. Since the possibility of capital accumula-
tion cannot decelerate output growth, Y does in fact grow power-functionally.

Let us last consider the case of γ = 1.

1
From (73), we know that if gKt > (<) 1−a gAt then ĝKt < (>)0. Likewise,
β
from (74), we know that if gAt > (<) 1−ϕ gAt then ĝAt < (>)0. When γ = 1,
however,
β
= 1 − a, (95)
1−ϕ
 64

so

gK0 ≥ gA0 /(1 − a) (96)

⇐⇒ gKt ≥ gAt /(1 − a) ∀t. (97)
∗
By the reasoning following (80), the limits gK ≜ limt→∞ gKt and gA∗ ≜
limt→∞ gAt are defined, finite, and positive. Furthermore, by the continu-
∗
ity of ĝK and ĝA in gK and gA , we must have gK = gA∗ /(1 − a). Thus
∗
gK 1
∗
= (98)
gA 1−a
a
s(1 − SK ) 2−ϕ a−1−λ 1
=⇒ lim λ
At Kt = (99)
t→∞ θSK 1−a

λ 1−a
 θSK  1+λ−a
=⇒ lim At Kta−1
= by γ = 1 (100)
t→∞ s(1 − a)(1 − SK )a
 θ λ 1−a
∗ λ SK  1+λ−a
=⇒ gK = s 1+λ−a by (71). (101)
1 − a (1 − SK )a

Finally,

lim gY t = gA∗ + agK

∗
(102)
t→∞
∗
= gK by (98). (103)
 65

References
Acemoglu, Daron and David Autor, “Skills, tasks and technologies: Implica-
tions for employment and earnings,” in “Handbook of Labor Economics,” Vol. 4,
Elsevier, 2011, pp. 1043–1171.

and , “What Does Human Capital Do? A Review of Goldin and Katz’s ‘The
Race Between Education and Technology’,” Journal of Economic Literature,
2012, 50 (2), 426–463.

and Pascual Restrepo, “Modeling Automation,” AEA Papers and Proceed-

ings, 2018, 108, 48–53.

and , “The Race Between Man and Machine: Implications of Technology for
Growth, Factor Shares, and Employment,” American Economic Review, 2018,
108 (6), 1488–1542.

and , “Artificial Intelligence, Automation and Work,” in Ajay Agrawal,

Joshua Gans, and Avi Goldfarb, eds., The Economics of Artificial Intelligence:
An Agenda, National Bureau of Economic Research, 2019, pp. 197–236.

and , “Automation and New Tasks: How Technology Displaces and Rein-
states Labor,” Journal of Economic Perspectives, 2019, 33 (2), 3–30.

and , “Robots and Jobs: Evidence from US Labor Markets,” Journal of

Political Economy, 2020, 128 (6), 2188–2244.

Aghion, Philippe, Benjamin F Jones, and Charles I Jones, “Artificial

Intelligence and Economic Growth,” in Ajay Agrawal, Joshua Gans, and Avi
Goldfarb, eds., The Economics of Artificial Intelligence: An Agenda, National
Bureau of Economic Research, 2019, pp. 237–282.

Agrawal, Ajay, John McHale, and Alex Oettl, “Finding Needles in

Haystacks: Artificial Intelligence and Recombinant Growth,” in Ajay Agrawal,
Joshua Gans, and Avi Goldfarb, eds., The Economics of Artificial Intelligence:
An Agenda, National Bureau of Economic Research, 2019, pp. 149–174.

, Joshua Gans, and Avi Goldfarb, Prediction Machines: The Simple Eco-
nomics of Artificial Intelligence, Harvard Business Press, 2018.

Aschenbrenner, Leopold, “Existential Risk and Growth,” 2020. Global Priori-

ties Institute, Working Paper No. 6-2020.
 66

Autor, David, “Why Are There Still So Many Jobs? The History and Future of
Workplace Automation,” Journal of Economic Perspectives, 2015, 29 (3), 3–30.

Baumol, William J, “Macroeconomics of Unbalanced Growth: The Anatomy of

Urban Crisis,” The American Economic Review, 1967, 57 (3), 415–426.

Berg, Andrew, Edward F Buffie, and Luis-Felipe Zanna, “Should We Fear

the Robot Revolution? (The Correct Answer Is Yes),” Journal of Monetary
Economics, 2018, 97, 117–148.

Bloom, Nicholas, Charles I Jones, John Van Reenen, and Michael Webb,
“Are Ideas Getting Harder to Find?,” American Economic Review, 2020, 110
(4), 1104–1144.

Bostrom, Nick, Superintelligence, Oxford University Press, 2017.

Brynjolfsson, Erik and Andrew McAfee, The Second Machine Age: Work,
Progress, and Prosperity in a Time of Brilliant Technologies, WW Norton &
Company, 2014.

Burks, Arthur W and John Von Neumann, Theory of Self-reproducing Au-

tomata, University of Illinois Press, 1966.

Chen, Nicholas, Lau Christensen, Kevin Gallagher, Rosamond Mate,

and Greg Rafert, “Global Economic Impacts Associated with Artificial Intel-
ligence,” 2016. Analysis Group.

Chirinko, Robert S and Debdulal Mallick, “The Substitution Elasticity, Fac-

tor Shares, and the Low-frequency Panel Model,” American Economic Journal:
Macroeconomics, 2017, 9 (4), 225–253.

Cockburn, Iain M, Rebecca Henderson, and Scott Stern, “The Impact

of Artificial Intelligence on Innovation: An Exploratory Analysis,” in Ajay
Agrawal, Joshua Gans, and Avi Goldfarb, eds., The Economics of Artificial
Intelligence: An Agenda, National Bureau of Economic Research, 2019, pp. 115–
146.

Freeman, Richard B, “Who Owns the Robots Rules the World,” 2015. IZA
World of Labor.

Goldin, Claudia Dale and Lawrence F Katz, The Race Between Education
and Technology, Harvard University Press, 2009.
 67

Gordon, Robert J, The Rise and Fall of American Growth: The U.S. Standard
of Living since the Civil War, Princeton University Press, 2016.

Griliches, Zvi, “Hybrid Corn: An Exploration in the Economics of Technological

Change,” Econometrica, 1957, 25 (4), 501–522.

Hanson, Robin, “Long-term Growth as a Sequence of Exponential Modes,” 2000.

Working paper.

, “Economic Growth Given Machine Intelligence,” 2001. Working paper.

Hémous, David and Morten Olsen, “The Rise of the Machines: Automation,
Horizontal Innovation and Income Inequality,” 2014. Centre for Economic Policy
Research, Discussion Paper No. DP10244.

Jones, Charles I, “R&D-based Models of Economic Growth,” Journal of Political

Economy, 1995, 103 (4), 759–784.

, “Life and Growth,” Journal of Political Economy, 2016, 124 (2), 539–578.

Kaldor, Nicholas, “A Model of Economic Growth,” The Economic Journal,

1957, 67 (268), 591–624.

Korinek, Anton, “Humans, Artificially Intelligent Agents, and the Return of

Malthus,” 2018.

and Joseph E Stiglitz, “Artificial Intelligence and Its Implications for Income
Distribution and Unemployment,” 2019.

Kremer, Michael, “Population Growth and Technological Change: One Million

BC to 1990,” The Quarterly Journal of Economics, 1993, 108 (3), 681–716.

Mookherjee, Dilip and Debraj Ray, “Capital and Inequality in the Long Run:
Automation Without Technical Progress,” 2017. Working paper.

Nordhaus, William D, “Schumpeterian Profits in the American Economy: The-

ory and Measurement,” Technical Report, National Bureau of Economic Re-
search 2004.

, “Are We Approaching an Economic Singularity? Information Technology

and the Future of Economic Growth,” American Economic Journal: Macroeco-
nomics, 2021, 13 (1), 299–332.
 68

Oberfield, Ezra and Devesh Raval, “Micro Data and Macro Technology,”
2014. National Bureau of Economic Research, Working Paper No. 20452.

Purdy, Mike and Paul Daugherty, “Why Artificial Intelligence Is the Future
of Growth,” 2016. Accenture.

Roodman, David, “On the Probability Distribution of Long-term Changes in

the Growth Rate of the Global Economy: An Outside View,” 2020. Open
Philanthropy.

Sachs, Jeffrey D, R&D, Structural Transformation, and the Distribution of In-

come, University of Chicago Press, 2019.

and Laurence J Kotlikoff, “Smart Machines and Long-term Misery,” Tech-

nical Report 2012. National Bureau of Economic Research, Working Paper No.
18629.

, Seth G Benzell, and Guillermo LaGarda, “Robots: Curse or Blessing?

A Basic Framework,” 2015. National Bureau of Economic Research, Working
Paper No. 21091.

Sandberg, Anders, “An Overview of Models of Technological Singularity,” in

Max More and Natasha Vita-More, eds., The Transhumanist Reader: Classical
and Contemporary Essays on the Science, Technology, and Philosophy of the
Human Future, Wiley Online Library, 2013, pp. 376–394.

Seeber, Isabella, Eva Bittner, Robert O Briggs, Triparna de Vreede,

Gert-Jan De Vreede, Aaron Elkins, Ronald Maier, Alexander B Merz,
Sarah Oeste-Reiß, Nils Randrup, Gerhard Schwabe, and Matthias
Söllner, “Machines as Teammates: A Research Agenda on AI in Team Collab-
oration,” Information & Management, 2020, 57 (2), 103174.

Simon, Herbert A, The New Science of Management Decision, Harper & Broth-
ers, 1960.

Uzawa, Hirofumi, “Neutral Inventions and the Stability of Growth Equilibrium,”

The Review of Economic Studies, 1961, 28 (2), 117–124.

Weitzman, Martin L, “Recombinant Growth,” The Quarterly Journal of Eco-

nomics, 1998, 113 (2), 331–360.

Yudkowsky, Eliezer, “Intelligence Explosion Microeconomics,” 2013. Machine

Intelligence Research Institute, Technical Report.
 69

Zeira, Joseph, “Workers, Machines, and Economic Growth,” The Quarterly

Journal of Economics, 1998, 113 (4), 1091–1117.