The Simple Macroeconomics of AI [pdf]

The Simple Macroeconomics of AI

∗

Daron Acemoglu

Massachusetts Institute of Technology

April 5, 2024

Abstract

This paper evaluates claims about the large macroeconomic implications of new advances in AI. It

starts from a task-based model of AI’s eﬀects, working through automation and task complementarities. It

establishes that, so long as AI’s microeconomic eﬀects are driven by cost savings/productivity improvements

at the task level, its macroeconomic consequences will be given by a version of Hulten’s theorem: GDP and

aggregate productivity gains can be estimated by what fraction of tasks are impacted and average task-level

cost savings. Using existing estimates on exposure to AI and productivity improvements at the task level,

these macroeconomic eﬀects appear nontrivial but modest—no more than a 0.71% increase in total factor

productivity over 10 years. The paper then argues that even these estimates could be exaggerated, because

early evidence is from easy-to-learn tasks, whereas some of the future eﬀects will come from hard-to-learn

tasks, where there are many context-dependent factors aﬀecting decision-making and no objective outcome

measures from which to learn successful performance. Consequently, predicted TFP gains over the next 10

years are even more modest and are predicted to be less than 0.55%. I also explore AI’s wage and inequality

eﬀects. I show theoretically that even when AI improves the productivity of low-skill workers in certain tasks

(without creating new tasks for them), this may increase rather than reduce inequality. Empirically, I ﬁnd

that AI advances are unlikely to increase inequality as much as previous automation technologies because

their impact is more equally distributed across demographic groups, but there is also no evidence that AI

will reduce labor income inequality. AI is also predicted to widen the gap between capital and labor income.

Finally, some of the new tasks created by AI may have negative social value (such as design of algorithms

for online manipulation), and I discuss how to incorporate the macroeconomic eﬀects of new tasks that may

have negative social value.

JEL Classiﬁcation: E24, J24, O30, O33.

Keywords: Artiﬁcial Intelligence, automation, ChatGPT, inequality, productivity, technology adop-

tion, wage.

∗

Paper prepared for Economic Policy. I am grateful to Can Ye¸sildere for phenomenal research assistance,

to Leonardo Bursztyn, Mert Demirer, Lauren Fahey, Shakked Noy, Sida Peng, Julia Regier, and Whitney

Zhang for useful comments, and to participants in the Economic Policy conference and my discussants there,

David H´emous and Benoˆıt Coeur´e, for comments and suggestions. I thank Pamela Mishkin and Daniel Rock

for generously sharing their data on AI exposure. I am also heavily indebted to my collaborators on several

projects related to these topics, David Autor, Simon Johnson and Pascual Restrepo, from whom I learned a

great deal and who have also given me very useful comments on the current draft. All remaining errors are

mine. The online Appendix is available upon request.

1 Introduction

Artiﬁcial intelligence (AI) has captured imaginations. Promises of rapid, even unparalleled,

productivity growth as well as new pathways for complementing humans have become com-

monplace. There is no doubt that recent developments in generative AI and large language

models that produce text, information and images—and Shakespearean sonnets—in response

to simple user prompts are impressive and even spellbinding. ChatGPT, originally released

on November 30, 2022, soon became the fastest spreading tech platform in history, with an

estimated 100 million monthly users only two months after launch.

AI will have implications for the macroeconomy, productivity, wages and inequality, but

all of them are very hard to predict. This has not stopped a series of forecasts over the last

year, often centering on the productivity gains that AI will trigger. Some experts believe

that truly transformative implications, including artiﬁcial general intelligence (AGI) enabling

AI to perform essentially all human tasks, could be around the corner.

Other forecasters

are more grounded, but still predict big eﬀects on output. Goldman Sachs (2023) predicts

a 7% increase in global GDP, equivalent to $7 trillion, and a 1.5% per annum increase in

US productivity growth over a 10-year period. Recent McKinsey Global Institute (2023)

forecasts suggest that generative AI could oﬀer a boost as large as $17.1 to $25.6 trillion to

the global economy, on top of the earlier estimates of economic growth from increased work

automation. They reckon that the total impact of AI and other automation technologies

could produce up to a 1.5 − 3.4 percentage point rise in average annual GDP growth in

advanced economies over the coming decade.

Are such large eﬀects plausible? And if there are going to be productivity gains, who will

be their beneﬁciary? With previous automation technologies, such as robotics, most gains

Korinek and Suh (2024) predict a “baseline” GDP growth of 100% over the next 10 years, and also

entertain the possibility of much higher “aggressive” AGI growth rates, such as a 300% increase in GDP.

Many others are seeing recent developments as a conﬁrmation of the forecasts in Kurzweil (2005) about the

impending arrival of “singularity” and “explosive” economic growth (Davidson, 2021).

Three caveats are in order. First, although most recent advances are in generative artiﬁcial intelligence,

the economic forces explored here apply to other types of AI, and estimates of exposed tasks I use come on

the basis of anticipated improvements in a range of AI-related technologies, including computer vision and

software building on large language models. Hence, I consider the numbers here to apply to all of artiﬁcial

intelligence and thus typically refer to “AI”, unless there is a reason to emphasize generative AI.

Second, I focus on the US economy because much of the existing evidence on microeconomic eﬀects of

AI and prevalence of exposed tasks is from the United States. The impact on other industrialized nations

should be similar, whereas the consequences for the developing world are harder to ascertain and require

much more in-depth research.

Third, some commentators use “productivity” to refer to output per worker (or average labor productivity),

while others mean total factor productivity (TFP). Throughout, I distinguish between aggregate TFP and

GDP eﬀects, and I use productivity improvement at the micro/task level as synonymous to cost savings.

accrued to ﬁrm owners and managers, while workers in impacted occupations experienced

negative outcomes (e.g., Acemoglu and Restrepo, 2020a). Could it be diﬀerent this time?

Some experts and commentators are more optimistic. A few “proof-of-concept” experimen-

tal studies document nontrivial productivity gains from generative AI, largely driven by

improvements for less productive or lower-performing workers (e.g., Peng et al., 2023; Noy

and Zhang, 2023; Brynjolfsson et al., 2023), and this has prompted some experts to be cau-

tiously optimistic (Autor, 2024), while others are forecasting a “blue-collar bonanza” (The

Economist, 2023).

This paper uses the framework from Acemoglu and Restrepo (2018, 2019b, 2022) to pro-

vide some insights for these debates, especially relevant for the medium-term (about 10-year)

macroeconomic eﬀects of AI. I build a task-based model, where the production of a unique

ﬁnal good requires a series of tasks to be performed, and these tasks can be allocated to ei-

ther capital or labor, which have diﬀerent comparative advantages. Automation corresponds

to the expansion of the set of tasks that are produced by capital (including digital tools and

algorithms). In this framework, AI-based productivity gains—measured either as growth of

average output per worker or as total factor productivity growth—can come from a number

of distinct channels (see Acemoglu and Restrepo, 2019a):

• Automation (or more precisely extensive-margin automation) involves AI models tak-

ing over and reducing costs in certain tasks. In the case of generative AI, various

mid-level clerical functions, text summary, data classiﬁcation, advanced pattern recog-

nition, and computer vision tasks are among those that can be proﬁtably automated.

• Task complementarity can increase the productivity in tasks that are not fully au-

tomated and may even raise the marginal product of labor. For example, workers

performing certain tasks may have better information or access to other complemen-

tary inputs. Alternately, AI may automate some subtasks, while at the same time

enabling workers to specialize and raise their productivity in other aspects of their job.

• Deepening of automation can take place, increasing the productivity of capital in tasks

that have already been automated. For example, an already-automated IT security

task may be performed more successfully by generative AI.

• New tasks may be created thanks to AI and these tasks may impact the productivity

of the whole production process.

New tasks in this framework also capture the possibility of productivity-enhancing reorganizing produc-

In this paper, I focus on the ﬁrst two channels, though I also discuss how new tasks en-

abled by AI can have positive or negative eﬀects. I do not dwell on deepening of automation,

because the tasks impacted by (generative) AI are quite diﬀerent than those automated by

the previous wave of digital technologies, such as robotics, advanced manufacturing equip-

ment and software systems.

I also do not discuss how AI can have revolutionary eﬀects by

changing the process of science (a possibility illustrated by new crystal structures discovered

by the Google subsidiary DeepMind and recent neural network-enabled advances in protein

folding), because large-scale advances of this sort do not seem likely within the 10-year time

frame and many current discussions focus on automation and task complementarities.

I show that when AI’s microeconomic eﬀects are driven by cost savings (equiva-

lently, productivity improvements) at the task level—due to either automation or task

complementarities—its macroeconomic consequences will be given by a version of Hulten’s

theorem: GDP and aggregate productivity gains can be estimated by what fraction of tasks

are impacted and average task-level cost savings. This equation disciplines any GDP and

productivity eﬀects from AI. Despite its simplicity, applying this equation is far from trivial,

because there is huge uncertainty about which tasks will be automated or complemented,

and what the cost savings will be.

Nevertheless, as an illustrative exercise, I use data from a number of recent studies, in

particular, Eloundou et al. (2023) and Svanberg et al. (2024), as well as the experimental

studies mentioned above, to obtain some back-of-the-envelope numbers. Eloundou et al.

(2023) provide the ﬁrst systematic estimates of what tasks will be impacted by generative

AI and computer vision technologies. Their methodology does not fully distinguish whether

the impact will take the form of automation or task complementarities, and does not provide

information on when we expect these impacts to be realized and how large their cost savings

will be.

For computer vision technologies, Svanberg et al. (2024) provide estimates of what

fraction of tasks that are potentially exposed to AI can be feasibly automated in diﬀerent

tion. The role of AI in enabling such reorganization is emphasized by, among others, Bresnahan (2019) and

Agrawal et al. (2023).

Eloundou et al. (2023) report negative statistical associations between their measure of exposure to AI,

which I use below, and measures of exposure to robots and manual routine tasks.

More speciﬁcally, I use the most granular information that Eloundou et al. (2023) present, which is their

“automation index”, coded with help from GPT-4. This index provides information on how much of the

activities involved in a task/occupation can be performed by AI. Although this index has somewhat greater

emphasis on automation, it does not systematically distinguish between automation and task complemen-

tarities. As I discuss further below and Eloundou et al. (2023) themselves note, their exposure measure

often captures the possibility that generative AI and related digital technologies can perform some of the

subtasks in an occupation, providing more time for workers to focus on other activities, and thus contains

both automation and task complementarity elements.

time frames.

I take Eloundou et al.’s estimates of tasks that are exposed to AI (without distinguishing

automation vs. task complementarities). I then aggregate this to the occupational level and

weight the importance of each occupation by its wage bill share in the US economy. This

calculation implies that 19.9% of US labor tasks are exposed to AI. I then use Svanberg et

al.’s estimate for computer vision tasks that, among all exposed tasks, 23% can be proﬁtably

performed by AI (for the rest, the authors estimate that the costs would exceed the beneﬁts).

I take the average labor cost savings to be 27%—the average of the estimates in Noy and

Zhang (2023) and Brynjolfsson et al. (2023)—and turn this into total cost savings using

industry labor shares, which imply an average total cost savings of 15.4%.

This calculation implies that total factor productivity (TFP) eﬀects within the next 10

years should be no more than 0.71% in total—or approximately a 0.07% increase in TFP

growth annually. If we add bigger productivity gains from Peng et al. (2023), which are less

likely to be broadly applicable, or incorporate further declines in GPU costs, this number

still remains around 1%.

To turn these numbers into GDP estimates, we need to know how much the capital stock

will increase due to AI. I start with the benchmark of a rise in the capital stock proportional

to the increase in TFP. This benchmark is consistent with the fact that generative AI does

not seem to require huge investments by users (beyond those made by designers and trainers

of the models). With these investment eﬀects incorporated, GDP is also estimated to grow

by around 1.1% over the next 10 years. When I assume that investments will be similar

to those for earlier automation technologies and use the full framework from Acemoglu and

Restrepo (2022) to estimate the increase in the capital stock, the upper bound on GDP

eﬀects rises to 1.6 − 1.8%. Nevertheless, my framework also clariﬁes that what is relevant

for consumer welfare is TFP, rather than GDP, since the additional investment comes out

of consumption.

I then argue that the numbers above may be overestimates of the aggregate productivity

beneﬁts from AI, because existing estimates of productivity gains and cost savings are in

tasks that are “easy-to-learn”, which then makes them easy for AI. In contrast, some of

the future eﬀects will come from “hard-to-learn” and hard for AI tasks, where there are

many context-dependent factors aﬀecting decision-making, and most learning is based on the

behavior of human agents performing similar tasks (rather than objective outcome measures).

For example, if AI models continue to increase their energy requirements, this would contribute to

measured GDP, but would not be a beneﬁcial change for welfare.

Productivity gains in these hard tasks will be less—though, of course, it is challenging to

determine exactly how much less. Using a range of (speculative) assumptions, I estimate

an upper bound of 74% easy tasks among Eloundou et al.’s exposed tasks. I suppose that

productivity gains in hard tasks will be approximately one quarter of the easy ones. This

leads to an updated, more modest increase in TFP and GDP in the next 10 years that can

be upper bounded by 0.55% and 0.90%, respectively.

New tasks created with AI can more signiﬁcantly boost productivity. However, some of

the new AI-generated tasks are manipulative and may have negative social value, such as

deepfakes, misleading digital advertisements, addictive social media or AI-powered malicious

computer attacks. While it is diﬃcult to put numbers on good and bad new tasks, based

on recent research I suggest that the negative eﬀects from new bad tasks could be sizable.

I make a very speculative attempt using numbers on the negative welfare eﬀects of social

media from a recent paper by Bursztyn et al. (2023). These authors ﬁnd that consumers have

positive willingness to pay for using social media (in particular Instagram and TikTok) when

others are using it, but they would prefer that neither themselves nor others use it. Roughly

speaking, their estimates imply that revenue can increase by about $53 per user-month,

but this has a negative impact on total GDP/welfare equivalent to $19 per user-month.

Combining these numbers with an estimate of the fraction of activities that may generate

negative social value (in practice, revenues from social media and spending on attack-defense

arms races in IT security), I suggest that with more intensive use of AI, it is possible to have

nontrivial increases in GDP. For example, AI may appear to increase GDP by 2%, while in

reality reducing welfare by −0.72%.

I also explore AI’s wage and inequality eﬀects. My framework implies that productivity

gains from AI are unlikely to lead to sizable wage rises. Moreover, even if AI improves

the productivity of low- and middle-performing workers (or workers with limited expertise

in complex tasks), I argue that this may not translate into lower inequality. In fact, I

show by means of a simple example how an increase in the productivity of low-skill workers

in certain tasks can lead to higher rather than lower inequality. Adapting the general

equilibrium estimates from Acemoglu and Restrepo (2022) to the setting of AI, I ﬁnd that

the more intensive use of AI is unlikely to lead to substantial wage declines for aﬀected

groups, because AI-exposed tasks are more evenly distributed across demographic groups

than were the tasks exposed to earlier waves of automation. Nevertheless, I estimate that

AI will not reduce inequality and is likely to have a negative eﬀect on the real earnings of

low-education women (especially white, native-born women). My ﬁndings also suggest that

AI will further expand the gap between capital and labor income as a whole.

Finally, I argue that as originally suggested in Acemoglu and Restrepo (2018), more

favorable wage and inequality eﬀects, as well as more sizable productivity beneﬁts, will

likely depend on the creation of new tasks for workers in general and especially for middle-

and low-pay workers. While this is feasible in theory and I have argued elsewhere how it

could be achieved (Acemoglu, 2021 and Acemoglu et al., 2023), I also discuss why this does

not seem to be the focus of artiﬁcial intelligence research at the moment.

In sum, it should be clear that forecasting AI’s eﬀects on the macroeconomy is extremely

diﬃcult and will have to be based on a number of speculative assumptions. Nevertheless,

the gist of this paper is that a simple framework can discipline our thinking and forecasts,

and if we take this framework and existing estimates seriously, it is diﬃcult to arrive at very

large macroeconomic gains.

The rest of the paper is organized as follows. The next section outlines the concep-

tual framework I use throughout the paper and derives a number of theoretical insights

on aggregate productivity gains, investment responses, and wage and inequality eﬀects. It

also discusses the crucial distinction between easy-to-learn and hard-to-learn tasks and their

productivity implications, and introduces the contrast between good and bad new tasks.

Section 3 provides a preliminary quantitative analysis of new AI breakthroughs within this

framework. It ﬁrst presents a baseline (upper bound) estimate on the basis of the fraction

of existing tasks that are likely to be impacted by AI within the next 10 years and existing

estimates of cost savings (productivity improvements) from AI. It then reﬁnes this estimate

by introducing the distinction between easy-to-learn and hard-to-learn tasks and undertakes

a preliminary classiﬁcation of AI-exposed tasks into the easy and hard categories. I then

make an even more speculative attempt at incorporating the macroeconomic implications of

bad new tasks into this framework. Finally, I report estimates on the wage and inequality

implications of recent AI advances. Section 4 concludes with a general discussion, while the

Appendix, which is available upon request, includes additional information on how tasks are

classiﬁed into exposed and non-exposed and easy-to-learn and hard-to-learn categories.

2 Conceptual Framework

The model here builds on Acemoglu and Autor (2011) and Acemoglu and Restrepo (2018,

2019b, 2022), and I focus on the main elements of the framework, referring the reader to

these papers for further details and reﬁnements. The economy is static and involves the

production of a unique ﬁnal good, and all markets are competitive.

The production of a unique ﬁnal good takes place by combining a set of tasks, with

measure N, using the following production function

Y = B(N)



y(z)

σ−1



σ−1

, (1)

where Y (z) denotes the output of task z for z ∈ [0, N], σ ≥ 0 is the elasticity of substitution

between tasks and the parameter B(N) depends on N to capture the possible system-wide

eﬀects of new tasks, though in what follows I will suppress this dependence to simplify the

notation. For now, the elasticity σ can take any value, but it is reasonable to presume σ ≤ 1,

so that tasks are gross complements. I later set the elasticity of substitution between tasks

to σ ' 0.5, as estimated by Humlum (2023) and also imposed in Acemoglu and Restrepo

(2022).

Tasks can be produced using capital or labor according to the production function

y(z) = A

(z)l(z) + A

(z)k(z) for any z ∈ [0, N],

where A

and A

are labor-augmenting and capital-augmenting productivity terms, γ

(z)

and γ

(z) are labor’s and capital’s task-speciﬁc productivity schedules, and l(z) and k(z)

denote labor and capital allocated to performing task z. This task production function

implies that capital and labor have diﬀerent productivities in diﬀerent tasks, but within a

task they are perfect substitutes.

Throughout, I assume that γ

(z)/γ

(z) is increasing in z, so that labor has a comparative

advantage in higher-indexed tasks. This implies that there exists some threshold I such that

tasks z ≤ I are produced with capital and those above this threshold are produced with

labor.

Acemoglu and Restrepo (2018) provide a dynamic version of this economy with capital accumulation

and endogenous technological choices, while Acemoglu and Restrepo (2022) provide a generalization with

multiple types of labor and multiple sectors, and Acemoglu and Restrepo (2023) consider a non-competitive

version of this economy. Extending the framework in any of these directions has no eﬀect on the results and

implications I explore here.

One important simpliﬁcation is to assume that tasks assigned to labor do not require any capital or

tools, which is clearly unrealistic. The online Appendix of Acemoglu and Restrepo (2018) shows that the

results are very similar if the task production function is modiﬁed such that:

y(z) = A

(z)



l(z)

1−κ

(z)



+ A

(z)k(z),

where κ ∈ (0, 1) and k

(z) is labor-complementary capital in task z (while k(z) denotes capital used for

automating task z). Because κ < 1, tasks assigned to labor are still less intensive in capital than are

fully-automated tasks.

I normalize the total population to 1 and assume that diﬀerent workers have diﬀerent

units of eﬀective labor. To simplify the discussion, I assume that there are two types of labor,

high-skill and low-skill, and there is no comparative advantage diﬀerence between these two

types of labor (I return to comparative advantage later). The only diﬀerence is that high-skill

labor, which makes up a fraction φ

of the population, has λ

units of eﬀective labor, while

the remaining φ

= 1 − φ

low-skill labor has only λ

< λ

units of eﬀective labor. This

speciﬁcation ensures that both high-skill and low-skill workers could be performing some of

the same tasks. It also implies that wage inequality is pinned down by λ

/λ

—a feature I

relax later.

I also assume that all labor is supplied inelastically, so I write the total supply of labor

+ φ

= L.

The labor market-clearing condition is

L =

l(z)dz, (2)

and I denote the wage rate by w.

Capital is specialized for the tasks in which it is used, and I assume that capital of type

z is produced linearly from the ﬁnal good with unit cost

R(z) = R(K)ρ(z), (3)

where

K =

k(z)dz

is the overall capital stock of the economy. All ﬁrms take the cost of capital for task z, R(z),

as given. The ﬁrst term in (3) implies that the required rate of return on overall capital can

increase when the capital stock of the economy is larger and the second term is task-speciﬁc,

representing the possibility that diﬀerent types of capital could have diﬀerent costs. For

tasks that are not yet technologically automated—meaning that they cannot be produced

by capital—we can either set γ

(z) = 0 or take ρ(z) to be very large.

Finally, I assume that there exists a (non-satiated) representative household that con-

sumes the ﬁnal good (net of capital expenditures).

2.1 Equilibrium

I focus on a competitive equilibrium, which satisﬁes the following usual conditions:

• The allocation of tasks z ∈ [0, N] is cost-minimizing. That is, task z ∈ [0, N] is

produced by labor if and only if

(z)

R(z)

(z)

• The amount of capital k(z) is chosen to maximize Y − R(z)k(z), where Y is given as

in (1) and the overall capital stock of the economy K is taken as given.

• The labor market clears. That is, (2) holds.

Notice that the ﬁrst condition imposes an innocuous tie-breaking rule that when indiﬀer-

ent, ﬁrms use capital for performing a task. Given this tie-breaking rule, all tasks z > I will

be performed by labor (i.e., l(z) = 0 for all z ≤ I and k(z) = 0 for all z > I). Whether this

is high- or low-skill labor is indeterminate in the baseline model, so I focus on the overall

amount of eﬀective labor units.

In a competitive equilibrium, all tasks performed by labor must have

σ−1

(z)

σ−1

l(z)

−

= w. (4)

This implies that for any two tasks z > I and z

> I,

l(z)

l(z

)

(z)

σ−1

)

σ−1

. (5)

Notice that when σ < 1, less labor is allocated to tasks in which labor’s productivity is

higher—a feature whose implications I will emphasize later. Equation (5), combined with

the labor market-clearing condition (2), implies

l(z) =

(z)

σ−1

(z)

σ−1

L. (6)

Moreover, with a similar reasoning for any task z < I, only capital is used, and the

ﬁrst-order condition for capital is simply

σ−1

(z)

σ−1

k(z)

−

= R(K)ρ(z). (7)

Combining (6) and (7) with (1), GDP or total output can be written as

Y =









(z)

σ−1



(BA

σ−1

1 −





(z)

R(K)ρ(z)



σ−1



σ−1

−1







σ−1

. (8)

The denominator here is due to the roundabout nature of production, and I assume that



(z)

R(K)ρ(z)



σ−1

−1

< 1 (9)

to ensure that output is ﬁnite in this economy. (Otherwise, because output linearly produces

machines, which then produce output, overall output can reach inﬁnity). With an identical

argument to that in Acemoglu and Restrepo (2022), an equilibrium exists and is unique,

provided that (9) is satisﬁed.

2.2 How AI Could Aﬀect Production

Before completing the characterization of equilibrium, I discuss how AI could aﬀect produc-

tion in this economy.

1. AI enables further (extensive-margin) automation, increasing I. Such automation could

be triggered either because AI reduces the cost of capital for some marginal tasks

(i.e., tasks slightly above I) or increases the eﬀectiveness of machinery or algorithms

performing some marginal tasks, thus raising γ

(z) for some z above I. Obvious

examples of this type of automation include generative AI tools such as large language

models (LLMs) taking over simple writing, translation and classiﬁcation tasks as well

as somewhat more complex tasks related to customer service and information provision,

or computer vision technologies taking over image recognition and classiﬁcation tasks.

2. AI can generate new task complementarities, raising the productivity of labor in tasks

it is performing. For example, AI could provide better information to workers, directly

increasing their productivity. This possibility could be modeled as AI reducing the cost

of complementary capital k

(z) in some tasks z > I in the more general formulation in

footnote 8. Alternatively, AI could automate some subtasks (such as providing ready-

made subroutines to computer programmers) and simultaneously enable humans to

specialize in other subtasks, where their performance improves. This channel would

require the explicit modeling of the range of subtasks making up each task. In this case,

new AI technologies would perform some of these subtasks and do so with suﬃciently

high productivity, so that the subtask-level displacement would be weaker than the

productivity gains, expanding the demand for labor and the marginal productivity

of labor in these tasks. The logic of the productivity eﬀect being larger than the

displacement eﬀect is the same as in the basic models of automation, as exposited

in Acemoglu and Restrepo (2018, 2019b). Even more interestingly, AI may enable

workers to specialize in the non-automated subtasks and raise their expertise in these

activities (e.g., when humans spend less time in writing standard subroutines, they

can become better at other parts of programming). I represent task complementarities

by an increase in γ

(z) in some tasks z ≤ I, or when they happen in all tasks, by an

increase in A

3. AI could induce deepening of automation—meaning improving performance, γ

(z), or

reducing costs, ρ(z), in some previously capital-intensive tasks (tasks z ≤ I). Examples

include IT security, automated control of inventories, and better automated quality

control (see Acemoglu and Restrepo, 2019a).

4. AI can generate new labor-intensive products or tasks, which corresponds to an increase

in N. As argued in Acemoglu and Restrepo (2020b), Acemoglu (2021) and Acemoglu

et al. (2023), there are many pathways for such new tasks. Later I discuss the case

where some of these new products and tasks can be manipulative and have negative

social value.

The eﬀects of new AI tools will depend on the extent of each one of these eﬀects, and I

will try to provide more speciﬁcity on these possibilities later. In the rest of this section, I

will derive the consequences of diﬀerent eﬀects of AI.

2.3 Equilibrium Wages and Comparative Statics

As a ﬁrst step, let us combine (4) and (6), so that the equilibrium wage can be expressed as

w =





(BA

)

σ−1



(z)

σ−1



. (10)

This equation is intuitive. The ﬁrst term shows that the wage is proportional to labor

productivity (raised to the power 1/σ), and the second term captures the contribution to

the marginal productivity of labor coming from Hicks-neutral and labor-augmenting tech-

nologies, while the third term represents the contribution of the allocation of tasks to the

marginal productivity of labor. The eﬀect of any small technological change (potentially

altering multiple dimensions of the production technology, such as B; A

and A

; γ

(z) and

(z); and I and N) can then be written as:

d ln w =

d ln





σ − 1

(d ln B + d ln A

) +

d ln



(z)

σ−1



. (11)

The eﬀect of an extensive-margin automation—an increase in I—is given by

d ln w

d ln Y

−

(I)

σ−1



(z)

σ−1



In general, this expression has ambiguous sign, so automation can reduce wages. More specif-

ically, there are two opposing eﬀects (Acemoglu and Restrepo, 2018, 2019b): (a) automation

always produces a positive eﬀect on wages (and labor demand) because it increases produc-

tivity (or equivalently, reduces costs). This positive productivity eﬀect is represented by

the ﬁrst term; (b) simultaneously, automation displaces workers from the tasks they used to

perform. The negative displacement eﬀect is represented by the second term. In the special

case where R(K) is constant, it can be veriﬁed that automation increases wages. This is not

the case, in general, when R(K) is increasing, as shown in Acemoglu and Restrepo (2018),

because the displacement eﬀect can be larger than the productivity gains.

Overall, the impact of (extensive-margin) automation on the equilibrium wage is closely

tied to its productivity eﬀect, to which I next turn.

Before doing so, I also note that the eﬀects of task complementarities may be a little more

complex than typically assumed. Even though an increase in γ

(z) increases the marginal

physical product of labor, the equilibrium wage is determined by the value of the marginal

product of labor, which depends on the adjustment of task prices. As tasks produced by

labor become more abundant/easier to perform, these task prices are reduced, and in the

empirically relevant case where σ < 1 (as noted above), these task prices decline more than

the increase in physical productivity. The equilibrium wage may still increase because of

productivity gains, but the beneﬁts to labor may be limited overall. For example, holding

I and N constant, an increase in A

will leave wages constant when σ = s

where s

denotes the capital share in national income (Acemoglu and Restrepo, 2018). When σ < s

higher A

can actually reduce real wages. Since s

' 0.4 currently in the US economy,

Humlum’s estimates of σ mentioned above, of about 0.5, imply that task complementarities

or labor-augmenting technological improvements will not raise wages much. Even when they

increase wages, these technological shifts reduce the labor share, just like automation does

In particular, for any diﬀerentiable constant returns to scale production function F (K, L), Euler’s the-

orem implies: F (K, L) = wL + RK. Suppose R is ﬁxed. Then any change in technology that increases

F (K, L) at the initial factor supplies must increase w =

F (K,L)−RK

. This is not the case when K is ﬁxed,

because R can increase more than F (K, L). In Acemoglu and Restrepo (2018, 2019b), the capital stock is

taken as given, so a negative impact is possible.

In neoclassical growth models with exponential discounting and time-separable preferences, R is ﬁxed in

the long run. However, Acemoglu et al. (2024) show that with more general preferences, R tends to increase

following automation and other technological changes, so a negative wage eﬀect is possible.

(Acemoglu and Restrepo, 2018, 2019b).

Finally, I note that deepening of automation (which is less likely to be relevant in the

case of AI for the reasons discussed in the Introduction) and new tasks always increase wages

and the latter always increases the labor share of national income as well (and thus tends

to narrow the gap between capital and labor income). I return to a discussion of new tasks

later.

2.4 Hulten’s Theorem

The nexus of the many eﬀects of AI on the economy is its impact on productivity. To

gauge the extent of this impact, we can appeal to Hulten’s theorem, which provides a simple

formula for competitive economies with constant returns to scale, specifying how micro-

level productivity improvements translate into macro changes. Since the economy here is

competitive, this theorem applies and disciplines the productivity eﬀects. I now explain this

theorem and for ease of exposition, I ﬁrst take the capital stock of the economy, K, as given.

Consider an arbitrary small change in all/any aspects of technologies (B; A

and A

; γ

(z)

and γ

(z); and I and N), so that the potential eﬀects of AI based both on automation

and complementarities in existing tasks are taken into account. Recall that GDP can be

equivalently expressed as the price-weighted aggregate of task outputs:

Y =

p(z)y(z)dz.

Since I am considering small changes in technology and the competitive equilibrium is eﬃ-

cient, the impact of all reallocations of factors across tasks and indirect eﬀects via prices will

be second-order and can be ignored in computing GDP changes.

Then, denoting changes

by ∆, this can be written as

∆Y =

p(z)∆y(z)dz, and thus

∆Y

p(z)y(z)

∆y(z)

y(z)

dz.

Hence, focusing on small changes, deﬁning χ(z) = p(z)y(z)/Y as the GDP share of task z,

and recalling that since capital and labor are ﬁxed, this is also the total factor productivity

Namely, change in real GDP can be expressed as the change in the value of the objective function

of the social planner (e.g., the utility function of the representative household if one exists). Then the

Lagrange multipliers on task-level resource constraints are equivalent to the prices and can be taken as given

in evaluating the eﬀects of small changes in parameters/technology on the maximized value of this objective

function, by the envelope theorem. See Hulten (1978).

(TFP) change, I write

d ln TFP = d ln Y |

χ(z)π

(z)dz, (12)

where π

(z) = d ln y(z) is the productivity improvement or cost savings in task z driven by

AI. I emphasize once more that these cost savings could be rooted in automation or task

complementarities. Total eﬀect on GDP is given by

d ln Y = d ln TFP + s

d ln K, (13)

where, again, s

is the capital share of GDP.

This derivation also clariﬁes that (12) and

(13) apply regardless of the task-level production functions, and thus are more general than

the wage and inequality results reported above.

If the main impact of AI on productivity is via automation, then π

(z)’s will come from

cost savings from automation. These should not be huge. In the case of robots, expert

reports put cost savings to about 30% of labor costs (see Acemoglu and Restrepo, 2020a),

though there are reasons to expect that cost savings from AI could be smaller than this,

as I discuss below. Estimates from Noy and Zhang (2023) and Brynjolfsson et al. (2023),

which I review in greater detail below, suggest numbers in the same ballpark—about 27% of

labor costs.

Although it is not clear whether these come from extensive-margin automation

or task complementarities, this distinction is less important for the productivity eﬀects, as

noted above. In practice, labor cost savings need to be translated into overall cost savings,

taking into account what fraction of costs in an industry are due to labor. For example, if

the share of labor in industry i is α

, then labor cost savings of π

will translate into an

overall cost savings of π

= α

. Taking the average of these across industries, and denoting

this economy-wide cost savings by ¯π, I can write

d ln TFP = ¯π × GDP share of tasks impacted by AI. (14)

The contribution of capital to output growth can also be derived in the general case, just like Hulten’s

theorem. In particular, for any constant returns to scale production function F (K, L|A), where A is a

technology shifter, we have d ln Y/dA = d ln F (K, L|A)/dA + (d ln F (K, L)/dK) · dK/dA. The ﬁrst term

d ln F (K, L|A)/dA is d lnTFP. Competitive factor markets imply d ln Y = (RK/Y )d ln K, and thus d ln Y/ =

dTFP+s

d ln K. Moreover, d ln s

= d ln K − d ln Y , and hence d ln Y = (d lnTFP+s

d ln s

)/(1 − s

When the capital-output ratio (or equivalently the capital share) remains constant in response to changes

in technology, then d ln Y = d lnTFP/(1 − s

), which is the ﬁrst approximation I use. Later, I estimate by

how much the capital-output ratio is predicted to increase due to automation and update this estimate.

Peng et al. (2023) estimate even larger eﬀects on how quickly certain programming tasks can be com-

pleted. But these refer to a very narrow set of tasks—subroutines that GitHub Copilot can write in some

common programming languages—and are thus less broadly applicable, and I consider them in robustness

checks.

I discuss below how this share can be estimated from recent studies on which occupations

and tasks will be impacted by AI.

2.5 Easy Tasks and Hard Tasks

As noted above, the proof-of-concept studies Peng et al. (2023), Noy and Zhang (2023), and

Brynjolfsson et al. (2023) have focused on tasks where AI could have clear beneﬁts and where

there were already sophisticated applications (such as GitHub Copilot for the ﬁrst study)

or where some businesses were already making an eﬀort to use generative AI (as in the

customer service application of the third study). I now argue that estimates from tasks that

are relatively easy for current AI technologies, even if reliable within their chosen context,

cannot be directly extrapolated to the rest of the economy. More generally, it is useful to

distinguish between “easy-to-learn tasks” where outside learning (and thus learning by AI

models) is easy from those where gaining expertise on the basis of outside observation is

hard. Easy-to-learn tasks, which are relatively straightforward for (generative) AI to learn

and implement, are deﬁned by two characteristics:

• there is a simple (low-dimensional) mapping between action and the (perfect) outcome

measure, and

• there is a reliable, observable outcome metric.

How to boil an egg (or providing instructions for boiling an egg), the veriﬁcation of the

identity of somebody locked out of a system or the composition of some well-known program-

ming subroutines are easy tasks. The desired outcome—an egg that is boiled to the desired

level, allowing only authorized people to access the system, or whether the subroutine works

or not—is clear. In none of these cases do the successful outcomes depend on the complex

interaction of many dimensions of actions. With reliable, objective measures of success (well

boiled egg, no security breach given the ground truth of authorized people, or a subroutine

that does not crash), AI models can learn to perform well in a relatively straightforward

manner. Beyond this, AI models can also learned quite well from human actions, because

there are expert humans who perform well in these tasks, such as expert programmers, and

because objective measures of success are available, experts can be identiﬁed.

The hard vs. easy distinction is diﬀerent from the routine vs. non-routine distinction introduced in

Autor et al. (2003). Routine tasks are those that involve repeated performance of the same activities (such

as knitting or switchboard operations) in stable, predictable environments. The use of digital technologies in

these tasks involves step-by-step programming of relevant software and hardware. A signiﬁcant number of

“Hard tasks” typically do not have a simple mapping between action and desired outcome.

In hard problems, what leads to the desired outcome in a given problem is typically not

known and strongly depends on contextual factors, or the number of relevant contexts may

be vast, or new problem-solving may be required. Additionally, there is typically not enough

information for the AI system to learn or it is unclear exactly what needs to be learned.

Diagnosing the cause of a persistent cough and proposing a course of treatment is a hard

problem. There are many complex interactions between past events that may be the cause

of the lingering cough and many rare conditions that should be considered. Moreover,

there is no large, well-curated data set of successful diagnoses and cures. In hard tasks, AI

models can still learn from human decision-makers, but because there is no clear metric of

success, identifying and learning from workers with the highest level of expertise will not be

straightforward either. As a result, there will be a tendency for the performance of AI models

to be similar to the average performance of human decision-makers, limiting the potential

for large productivity improvements and cost savings.

AI productivity gains observed so far are from easy tasks. It is reasonable to expect

that productivity gains in hard tasks are more limited, at least at ﬁrst. Productivity gains

in easy tasks result from AI models performing these tasks more or less at the same level

as expert workers and/or at lower cost than humans. For example, in the Noy and Zhang

(2023) study, expert workers are those that are able to summarize and write reasonably well.

The desired outcome in this case is straightforward to determine and does not require new

problem-solving eﬀorts (this may have been diﬀerent if the writing tasks were more complex

or required “creativity”). Generative AI models, such as GPT-4, are trained on vast amounts

of this type of writing, making this a clear example of an easy task. Productivity gains in

this case come from the fact that lower-expertise or less-skilled workers can be brought up to

the level of more expert workers at little cost and the majority of the workers are helped to

perform their assigned tasks more quickly. In easy problems, there may even be additional

productivity gains from AI models discovering action combinations that were not typically

routine tasks have already been automated, whereas AI’s promise is its ability to perform non-routine tasks,

as also emphasized by Susskind (2020). Some non-routine tasks such as parts of computer programming or

writing of simple text are easy-to-learn, while others that require more context-speciﬁc decisions and where

metrics of successful performance are scarce would be much harder to learn by outside observation.

Crucially, “easy tasks” in this sense may not be easy for humans. In fact, cost savings from easy tasks will

be most pronounced when they are expensive for humans to perform. Conversely, some hard tasks, such as

those that are based on intuition, experience and judgment, may be relatively straightforward for humans

because they do not learn to perform these tasks on the basis of outside learning alone. This is the reason

why cost savings relative to human performance are likely to be limited in hard tasks.

tried or known by expert humans.

There are several barriers to productivity gains in hard tasks. First, the lack of a simple

mapping between action and desired outcome will make it much more diﬃcult to train AI

models, and early automation and even human-complementary use of AI may be hampered

by this slow learning, delaying any productivity gains. Second, human complementarity may

be much harder to achieve. When there is no good information on what the desired outcome

is (e.g., what was the right diagnosis for the persistent cough?), most of the training data

of AI models will come from how humans act in similar circumstances. As a result, learning

from humans will not lead to better than human performance and is unlikely to generate

new complementarities and reveal diﬀerent insights than what humans are already doing.

If the premises of this section—that productivity gains will be slower in hard tasks than

in easy tasks in the next 10 years and tasks impacted by AI in the near future will be harder

than those today—are true, then (14) may need to be adjusted in an obvious way to account

for which fraction of tasks are easy and hard as deﬁned here. Speciﬁcally, denoting the

fraction of easy tasks among those that are impacted by AI as µ

, and the average cost

savings for easy and hard tasks, respectively, as ¯π

and ¯π

, then (14) becomes

d ln TFP =



¯π

+ (1 − µ

)¯π



× GDP share of tasks impacted by AI. (15)

I also attempt to provide estimates using this updated TFP equation below, which will lead

to gains that are about 25% less than those that do not take the distinction between easy

For instance, suppose humans typically start boiling the water with the egg inside, and it reduces

variability if the egg is placed inside the saucepan after the water has boiled, then this is something some

AI models may discover. AlphaZero’s discovery of new eﬀective chess moves can be viewed as an example

of the same phenomenon, even if it is in the context of a much more complex and interesting problem.

Looked at it this way, AlphaFold’s big success in protein folding can be considered to have both easy

and hard aspects. On the hard side, protein folding is a highly multidimensional problem, with no simple

mapping between action and desired outcome. On the easy side, however, the desired outcome is observed,

so with large amount of computing power, it is possible for an AI model to do even better than humans, as

AlphaFold managed to do.

Mathematically, the general learning problem can be formulated as follows. There is a mapping for

outcomes speciﬁed as f : X ×Z → Y , where X is the space of actions and Z denotes the space of “contexts”,

and Y is the set of possible outcomes. Observed outcomes are noisy versions of the true outcomes in Y , with

some noise process which I do not need to specify here. The mapping f is unknown.

In principle, there are two ways in which AI can be trained in such a problem. The ﬁrst one is that with

suﬃcient observations on Y and combinations of inputs from X × Z, the AI model may learn the entire

function f or some restricted version thereof. Or if Y is not observable, then the AI model may be trained

from human choices, which can be summarized by a correspondence G : Z ⇒ X, specifying the possible

actions that humans take when confronted with context Z. Either problem can be challenging if X × Z is

very high dimensional. However, the latter type of learning is less likely to lead to something diﬀerent and

better than what humans were doing already. It can sometimes ensure performance at the level of expert

humans, if these can be identiﬁed, but lack of reliable outcome data may also make it diﬃcult to identify

expertise.

and hard tasks into account.

2.6 Investment Responses

To go from TFP responses to total GDP responses, we need to see how much capital increases.

This is also relevant for understanding what magnitude of an investment boom generative

AI may trigger.

With this aim, consider a change in technology, for example due to AI. Its eﬀects on

capital and investment can be gleaned from (7), which implies

k(z) =

σ−1

(z)

σ−1

(R(K)ρ(z))

and hence

d ln k(z) = (σ−1)d ln B +(σ−1)d ln A

+(σ − 1) d ln γ

(z)+d ln Y −σd ln R(K)−σd ln ρ(z).

(16)

In what follows, I suppose that B and A

do not change. If AI signiﬁcantly improves the

practice of science and invention and/or creates new high-productivity tasks, such changes

may occur in the future. But within a 10-year horizon, it is reasonable to assume that they

are constant.

First, consider the use of AI in some task z < I, meaning that AI adds to and improves

the performance of existing capital equipment in tasks already performed by capital.

Then:

d ln k(z) = (σ − 1) d ln [A

· γ

(z)] + d ln Y − σd ln R(K).

Notice, however, that when σ < 1, the ﬁrst term is negative—an increase in the productivity

of AI-augmented capital in capital-intensive tasks reduces investment and the equilibrium

capital stock. As noted above, I take σ = 0.5, so this ﬁrst term is negative and non-trivial.

The last term is also nonpositive, provided that total investment increases. Hence, a natural

upper bound for the proportional increase in the capital stock of tasks already performed by

capital is the proportional increase in output.

Additionally, if I increases, investment will jump from zero to a positive amount in the

newly-automated tasks.

Because in the estimates I report below, only a modest fraction

of tasks will be automated with AI, this increase may be small as well. Hence, I start with a

The expression for the case in which labor-intensive tasks also use capital, as in footnote 8, is similar

and leads to the same conclusion.

Around I, this amount is given by k(I) =

σ−1

(I)

σ−1

(R(K)ρ(I))

. If labor-intensive tasks were previously

using capital as well, then the jump would be smaller but still positive.

ﬁrst estimate that takes the proportional increase in the capital stock to be the same as the

increase in aggregate productivity. I then provide another estimate incorporating the full

structure of between-industry and between-task substitution, based on the framework and

results of Acemoglu and Restrepo (2022). This will enable me to estimate implied changes

in the capital share and then use the expressions in footnote 11 to derive updated investment

response and GDP estimates.

I ﬁnally note that consumer welfare in this economy is proportional to TFP, and in

particular, GDP increases overstate improvements in consumer welfare, because investment

comes at the expense of consumption (Acemoglu and Restrepo, 2022).

2.7 New Good Tasks

Adding new tasks will expand productivity, and this eﬀect could be in principle larger than

the cost savings due to automation and task complementarities. In addition, new tasks will

tend to increase wages. In particular, using the same steps as before,

d ln w

d ln Y

(N)

σ−1



(z)

σ−1



σ − 1

(N)

B(N )

. (17)

This is always positive and could be large. Note also that the wage and productivity

impact of new tasks can be potentially larger than cost savings in existing tasks, and this

is particularly likely to be the case when new tasks improve the entire production process

(as captured by the B term), or when they add new sources of cost improvements or com-

plementary functions. Despite new tasks’ central role in wage and productivity growth and

in reducing labor income inequality (see Acemoglu and Restrepo, 2018, and Autor et al.,

2024), I will not focus on new good tasks generated by AI for the reasons discussed in detail

in the Conclusion.

2.8 New Bad Tasks

AI may generate new tasks that increase revenue and not consumer utility, such as through

addiction or manipulation, or in the context of production, via security attacks by malicious

actors. What this implies is that even when AI-related new tasks increase GDP, they may

not have the same positive eﬀect on welfare. To capture this, let us suppose that welfare

is given by W = ln C − ln E, where the second term is an externality, for example from

misinformation or manipulative activities, as I discuss below. In that case, the welfare

eﬀects of new tasks generated by AI will be

d ln Y

d ln E

. (18)

Although it is very diﬃcult to ascertain the magnitude of such negative welfare eﬀects, I

argue, based on some recent studies, that they may be nontrivial. Speciﬁcally, I use estimates

on the relative magnitudes of the two terms in (18) and proxy for the magnitude of the ﬁrst

term by the revenues of tasks where AI can produce new bad tasks or socially harmful

activities.

2.9 Wage and Inequality Implications of AI

As discussed in the Introduction, a number of commentators and experts are cautiously

optimistic that advances in generative AI could be beneﬁcial for labor or at the very least

not impact workers as adversely as previous waves of digital technologies, such as robotics

and software systems, which were predominantly used for automation. There are three

potential pathways via which such optimism may be realized.

1. AI can enable productivity increases in tasks currently produced by labor. This is the

task-complementarities channel and can be captured either by an increase in A

or an

increase in γ

(z) for a subset of the tasks that are automated or by increases in λ

and λ

(which, recall, are the productivities of unskilled and highly-skilled workers).

However, recall that when σ < 1, these types of productivity improvements will reduce

the labor share, and thus inequality between capital and labor will increase.

2. If AI generates very large productivity gains, it may increase wages even though it

reduces the labor share (Acemoglu and Restrepo, 2018, 2019b). This channel thus

critically hinges on the magnitude of the productivity eﬀects discussed above, but in

any case, always increases inequality between capital and labor.

3. As already discussed, some early studies show that within narrow occupations, lower-

performing or lower-expertise workers are the ones beneﬁting from generative AI. This

raises the possibility that AI could be more complementary to lower-skill workers and

may reduce labor income inequality. In my framework, this would be captured by an

increase in λ

relative to λ

. However, even in this case, inequality between capital

and labor will rise (provided that σ < 1).

4. If AI created new (good) tasks, these would reduce inequality between capital and

labor, and if enough new tasks were targeting lower-skill workers, this could also reduce

labor income inequality (Acemoglu and Restrepo, 2018).

I now argue theoretically that there are several reasons why 1 and 2 listed here are unlikely

to be major sources of wage growth or signiﬁcant limits on inequality. First consider a 1%

increase in A

(or an equivalent increase in γ

(z) for labor-intensive tasks). As explained

above, this may not increase wages at all, or may lead to only small wage increases. In

fact, from equation (11), abstracting from the productivity eﬀect, the direct impact on the

equilibrium wage will be a (σ − 1)/σ% change. When σ < 1, which is the plausible case

as discussed above, this is negative. The overall impact may still be an increase in wages

because of the productivity eﬀect, but as already noted, when σ is approximately equal to the

share of capital in national income, the overall impact will be essentially zero. In conclusion,

without the creation of a suﬃcient number of new tasks, inequality between capital and

labor will increase and wage rises may be limited.

What about a reduction in inequality because lower-skill workers beneﬁt more? In the

model here, the earnings of high-skill workers relative to low-skill workers is always pinned

at λ

/λ

. So if new technologies reduce this ratio, they will reduce the gap between high-

skill and low-skill workers. But even this conclusion needs to be qualiﬁed. Acemoglu and

Restrepo (2022) show that in more general settings, with multiple skill groups, there will be

ripple eﬀects whereby impacted demographic groups can then compete for tasks previously

performed by other groups. In such a situation, an overall increase in labor productivity

of both high-skill and low-skill workers in some tasks can lead to their displacement from

these tasks, and then the ripple eﬀects can, in principle, aﬀect low-skill workers even more

adversely than high-skill workers.

While such adverse eﬀects on low-skill workers are a general possibility in the framework

of Acemoglu and Restrepo (2022), I am not aware of worked-out examples where an increase

in the productivity of low-skill workers increases inequality. I now provide such an example.

2.10 How Greater Low-Skill Productivity Can Lead to Higher In-

equality

Let me illustrate this possibility with a simple example, by relaxing the assumption that there

are no comparative advantage diﬀerences between high-skill and low-skill workers. Suppose

that the economy starts from an equilibrium in which tasks below some I are performed

by capital, and tasks between I and N are performed by a combination of high-skill and

low-skill workers. In particular, suppose that low-skill workers have a comparative advantage

in tasks between I and I

∗

, and denote their (constant) productivity in these tasks by λ

while the productivity of high-skill workers in these tasks is also constant and equal to λ

Suppose also that the relative productivity of high-skill workers in tasks between I

∗

and

N is ω > λ

/λ

—indicating that high-skill workers have a comparative advantage in higher-

indexed tasks. However, because there is no strict comparative advantage, the equilibrium

may involve both types of workers performing some of the same tasks.

Assume also that

the elasticity of substitution between tasks is σ < 1.

Let us start with an equilibrium in which both high-skill and low-skill workers perform

tasks z ∈ (I, I

∗

], while only high-skill workers perform tasks z ∈ (I

∗

, N]. In this initial equi-

librium, the relative wage of skilled workers will be pinned down by the relative productivities

of the two types of workers in the tasks they are both performing—i.e., z ∈ (I, I

∗

]—and is

given by λ

/λ

Suppose that labor productivity in z ∈ (I, I

∗

] increases due to advances in AI, and this

also is more helpful for lower-skilled workers, so λ

/λ

declines. I now show that these

advances could boost inequality. Suppose that after this increase in productivity, because

σ < 1, the prices of tasks z ∈ (I, I

∗

] will decline and there will be less labor assigned

to these tasks. If this eﬀect is signiﬁcant, all high-skill workers may be allocated away

from these tasks, and the amount of labor demanded in these tasks may fall short of the

supply of low-skill workers. In this case, the post-AI allocation may involve only low-skill

workers performing tasks z ∈ (I, I

∗

], while both low-skill and high-skill workers perform

tasks z ∈ (I

∗

, N]. Then, regardless of how much λ

/λ

declines, the relative wage of skilled

workers will be determined by the tasks that both types of workers are performing, which

are now those above I

∗

, and thus will be equal to ω > λ

/λ

. Hence, inequality increases

following the rise in the productivity of low-skill workers.

Therefore, I have just proven that, in this general scenario, even a reduced productivity

gap between low-skill and high-skill workers in some tasks could lead to greater inequality.

Hence, the overall inequality implications of AI cannot be directly deduced from its eﬀects

on the performance of workers of diﬀerent skills in a given set of tasks and requires a fuller

empirical exploration. In the next section, I investigate AI’s inequality eﬀects by adapting

This structure makes the model similar to one that has a ﬁnite number of tasks—in this instance, three

tasks. But I use the setting with a continuum of tasks for continuity with the rest of the paper. The

“paradoxical” result I am highlighting here does not depend on this stark structure or on having just two

types of labor.

the framework and estimates of Acemoglu and Restrepo (2022) to the current setting.

3 A Preliminary Quantitative Evaluation

In this section, I provide a preliminary quantitative evaluation of the possible eﬀects of

recent breakthroughs in AI over a horizon of 10 years. The centerpiece will be the use

of Hulten’s theorem and recent estimates of which tasks can be automated using AI and

computer vision technologies and the cost savings thereof. Once I obtain these estimates

from Hulten’s theorem for TFP growth, I convert them to GDP growth estimates using

another series of assumptions on how the capital stock of the economy will respond. I also

make an attempt to think about how AI could aﬀect new tasks—including the distinction

between good and bad new tasks and their implications. Finally, I combine these numbers

with the more detailed framework in Acemoglu and Restrepo (2022) to obtain even more

speculative estimates for wage and inequality implications. Before moving to the estimates,

I ﬁrst describe the sources I use, further discuss existing productivity estimates and motivate

various parameter choices.

3.1 Data Sources and Parameter Choices

The centerpiece of the estimates in this paper is equation (14), and its later reﬁnement to

(15). To implement this equation, two pieces of information are needed:

1. GDP share of tasks that are impacted by AI (inclusive of computer vision) within the

next 10 years.

2. Average cost savings in these tasks due to AI, ¯π.

It is impossible to have accurate estimates of either of these two quantities, and hence

my repeated caution that the numbers here—and for that matter, other estimates in the

literature and the public debate—should be interpreted with great caution as suggestive

numbers. Nevertheless, there are studies that have already shed light on these quantities. I

now discuss what these are and how I use them.

GDP Share of Tasks Impacted by AI

The most careful estimates of which tasks are exposed to recent AI and computer vision

advances come from Eloundou et al. (2023). These authors use two related methodologies

for classifying which tasks are exposed to AI and computer vision. Both of those start from

O*NET task and Detailed Work Activity (DWA) descriptions. The authors ask GPT-4 to

classify all 19,265 tasks and 2,087 DWAs. They also develop a coarser index by manually

classifying DWAs and then cross-validate their GPT-based measure with this “human” cod-

ing. Here, I focus on the GPT-based measure which allows greater granularity, as I explain

next. In addition, Eloundou et al. (2023) distinguish between a direct exposure (α) measure,

which is based on their assessment of what large language models (LLMs) can achieve now.

They then develop a second, more aggressive measure (their so-called β measure), allowing

“indirect” exposure to a hypothetical LLM+. This includes tasks that will be (possibly)

exposed to new software and other advances building on current LLM and computer vision

technologies.

The index Eloundou et al. report in the paper is based on a binary coding, while they

also construct an “automation” index, and in this case their β measure includes granular

information about what fraction of the activities might be impacted by LLM+ (ranging

across 0%, 25%, 50%, 75% and 100%). In what follows, I take their automation index to be

able to use this granular information. This index still contains both automation and task

complementarities, even if it emphasizes automation a little more than their other exposure

indices (because they code the granular information on the basis of information about what

activities can be automated). In particular, as the authors themselves note, the impact

of generative AI may often involve automation of some subtasks, allowing human workers

to focus on other activities. The granular information contained in this index is especially

useful for my purposes, because it provides an assessment of which tasks/occupations are

less likely to be impacted by AI. I set all tasks that the authors classify as having not more

than 50% of activities impacted by AI and computer vision to zero, and I refer to the rest

as “AI exposed tasks”.

There are several problems that need to be tackled to turn these estimates into the

quantities I need.

• Although Eloundou et al.’s automation index emphasizes automation, it still includes

elements of “augmentation” or task complementarity. This motivates my interpretation

that combining their measure with equation (14) will capture cost savings from both

automation and task complementarities.

They also report a third, more aggressive measure, which they refer to as ζ. This measure is more spec-

ulative and focuses on what can be ultimately performed by LLM+. In line with the authors’ interpretation,

this is unlikely to be the case in the near future, and I ignore this third measure.

• Eloundou et al.’s data need to be converted into GDP shares. To do this, I combine

tasks into occupations, and then I aggregate across occupations using their wage bills,

computed from the U.S. Bureau of Labor Statistics National Occupational Employment

and Wage Estimates pooled across the years 2019-2022. This procedure yields a wage

bill-weighted share of exposed occupations equal to 19.9%. I interpret this number to

be the same as the GDP share of tasks exposed to AI.

• Eloundou et al.’s approach is to determine tasks that can be ultimately performed

by generative AI and computer vision technologies (such as the technology already

incorporated in Dall-E). Two things are missing from the information they provide.

The ﬁrst is how much of the task impact is likely to be realized within the next 10 years.

The second is whether, in all of these cases, it is proﬁtable to use AI (e.g., whether

automating using AI is cost-eﬀective). Svanberg et al. (2024) make an attempt to

provide answers to these questions in the case of computer vision technologies, which

are a subset of the technologies Eloundou et al. consider. I take Svanberg et al.’s

estimates and extrapolate them to all of the tasks Eloundou et al. consider.

Namely,

Svanberg et al.’s base estimates imply that among computer vision-exposed tasks, 23%

can be feasibly (and proﬁtably) automated within 10 years.

Applying this number

to Eloundou et al.’s estimates, I arrive at the GDP share of tasks impacted by AI as

0.23 × 0.199 = 4.6% of all tasks (or occupations) will be impacted by AI and computer

vision technologies within the next 10 years.

Cost Savings from AI

I base my estimate of ¯π on the experimental studies that have already provided “proof-of-

concept” estimates of productivity improvements or labor cost reductions due to AI. Three

studies, which I have already mentioned, are particularly notable here, and I now describe

each one of them.

This is not a trivial step. Svanberg et al. focus on automation using computer vision, while Eloundou

et al.’s exposed technologies include task complementarities as noted above and go beyond computer vision

tasks. One could imagine that the share of tasks that can be proﬁtably used with AI within 10 years is

diﬀerent between these two categories. Unfortunately, I do not have another source other than Svanberg et

al. for such an estimate for non-computer vision AI.

Svanberg et al. also make further extrapolations assuming cost declines in computer vision technologies.

They base these on estimates from Besiroglu and Hobbhahn (2022), which project a 22% yearly decrease in

computation costs attributed to the expansion of compute power from GPUs. However, a doubling of GPU

capacity will not necessarily translate into a 50% decline in costs in general due to diminishing returns and

bottlenecks created by other inputs, and because of the limitation of the current architecture. I therefore do

not include these in the baseline but return to them in the robustness discussion.

• Peng et al. (2023) design an experiment where freelance computer programmers are

given access to and encouraged to use GitHub Copilot, which (at the time of the

experiment) was powered by OpenAI Codex (GPT-3). Participants were given the

task of implementing a HTTP server on JavaScript, a popular language for which

resources and subroutines are readily available, and GPT-3 was already trained on these

resources. They compare the experimental treatment group to a control group that is

not given access to GitHub Copilot. They ﬁnd that the treatment group performed

the assigned tasks on average 55.8% faster than the control group. As an aside, like

the other studies I mention here, Peng et al. ﬁnd that these improvements come from

otherwise less well-performing subjects.

• Noy and Zhang (2023) design an online experiment where individuals in a range of

white-collar occupations are recruited and presented with simple writing tasks (in

particular, tasks like writing press releases, short reports and analyses that are designed

to take 20 to 30 minutes and resemble real life tasks in the participants’ occupations).

The treatment group is given access to and encouraged to use ChatGPT-3.5, while

the control group is not. They verify that there is very low usage of ChatGPT in

the control group. They estimate that access to ChatGPT enables, on average, 40%

faster completion of the task at hand. They also estimate an 18% improvement in

quality scores, as judged by peers and ChatGPT-based scoring of the output. In their

case, too, the gains come mostly from subjects that performed less well before the

experiment.

• Brynjolfsson et al. (2023) is the only study that I am aware of that looks at the use of

generative AI tools in a real business setting with a careful experimental design. The

business they focus on is a customer service provider, which uses a custom generative

AI tool to help customer service associates. The rollout took the form of a treatment

group getting access to this tool, while the control group did not. Brynjolfsson, Li

and Raymond evaluate the impact of the rollout on cost savings, by focusing on how

quickly tasks (open customer tickets) are resolved. They also look at self-reported

customer satisfaction. They ﬁnd a signiﬁcant improvement in the speed with which

tasks are completed by customer service associates—an eﬀect of about 14% on average.

However, they additionally estimate a slight and statistically insigniﬁcant decline in

quality, as judged by the users themselves. Like the other studies, Brynjolfsson, Li and

Raymond also ﬁnd that the results are predominantly among the lower-performing, less

expert employees. In fact, they estimate that the top quintile of associates experience

no improvement at all.

I interpret all three studies as providing labor cost savings from AI, broadly construed—in

particular, meaning that all three studies include both automation and task complementarity

elements. For instance, for the GitHub Copilot users in Peng et al., the authors’ interpreta-

tion is that some of the subtasks previously performed by programmers, such as the writing

of common routines, are now done by the Copilot. Along the lines of the discussion in

Section 2.2, suppose that programming in JavaScript involves N subtasks, which need to

be completed for a successful program. These include initial planning (which approach to

adopt, how to organize the program, etc.), composition of subroutines, putting the subrou-

tines together, debugging the subroutines, debugging the master program and then assessing

whether the program achieves the planned aims. When all of these subtasks are performed,

then the task at hand—the writing of a speciﬁc computer program—is completed. We can

think of generative AI as taking over a subset of the composition of subroutine tasks. Then

in line with the framework in Acemoglu and Restrepo (2018, 2019b), the use of this new

technology will create displacement and productivity eﬀects. Overall performance will in-

crease because of productivity eﬀects, provided that this technology is better than/faster

than humans at composing some subroutines (but humans are still needed for the other sub-

tasks, including planning, debugging and checking). Demand for human labor in this task

may increase or decrease depending on the magnitude of the displacement and productivity

eﬀects and the elasticity of substitution between this task and other tasks (as well as the

demand elasticity for the product that is ultimately being produced for the market).

In addition, when simple subroutines are taken over by the Copilot, this may enable hu-

man workers to specialize in higher-level subtasks, potentially generating task complemen-

tarities and further productivity gains. Importantly, however, in this experimental setting,

we do not see the displacement eﬀects, because each of the treatment subjects are given

the task and there is no possibility of reducing the number of workers performing this task.

Hence, from these experiments, we can only learn about the productivity gains, inclusive

of any task complementarities—and not about the displacement eﬀects. This interpretation

clariﬁes why I am comfortable bundling automation and task complementarities together for

the purposes of estimating productivity gains from AI.

What I have just described for programming also applies to the other two studies. Chat-

GPT is doing some of the drafting, which human subjects can then take and incorporate into

their writing with some veriﬁcation and modiﬁcation. The same applies for the customer

service associates, who are allowed to copy and paste text suggested by the generative AI

tool in the setting studied by Brynjolfsson, Li and Raymond.

I also assess that the tasks in these three studies are broadly comparable to the exposed

tasks considered in Eloundou et al. (2023), even though I argue later that they are more

likely to be in the easy-to-learn category. Hence, these studies are likely to be informative

about cost savings (productivity improvements) for the exposed tasks in Eloundou et al.

As a baseline, I ignore the quality eﬀects (which, as noted, are not uniform between the

three studies) and focus on the average increase in speed and interpret this as average cost

savings. I return to which types of workers beneﬁt more and the inequality implications

below. Finally, as a baseline, I use the average of the estimates from Noy and Zhang (2023)

and Brynjolfsson et al. (2023), and turn to the average of the three studies as a robustness

check. The reason for this is that Peng et al.’s setting is less likely to be relevant to other

tasks and occupations, since the task in question is a very well-deﬁned one for which Github

Copilot was extensively trained, and this has no direct equivalent in the other tasks we are

focusing on. Under these assumptions, the average labor cost savings are 27% (= 0.27).

Recall that these numbers refer to declines in labor costs in occupations, where some tasks

involve combination of capital and labor, and what is relevant in equation (14) is overall cost

savings. To convert these into overall cost savings, I use the industry-level estimates from

Eloundou et al. and combine these with industry labor shares from the BEA, as described

in the Appendix.

This gives an average (AI exposure-adjusted) labor share of 0.57, and

thus the average (overall) cost savings from AI are about 0.27 × 0.57 = 0.154.

If we add the numbers from Peng et al. to the mix, the average labor cost savings become

0.36, and thus the average overall cost savings come to 0.205.

3.2 Aggregate Productivity Gains: A First Pass

The ﬁrst-pass estimate of productivity (TFP) gains at the aggregate level can be obtained

simply by combining the numbers derived in the previous subsection with equation (14).

Brieﬂy, I follow Eloundou et al. and map exposed tasks to occupations, and then map occupations

to14 NIPA industries using the fractions of workers in each occupation employed across industries. I then

compute the labor share of exposed tasks using industry labor shares weighted by the value-added shares of

these industries in gross national income. The industry labor shares include self-employment income and are

for the years 2019-2022 from the U.S. Bureau of Bureau of Economic Analysis National Income and Product

Accounts (NIPA) data.

For reference, the average (value-added-weighted) labor share, without the exposure adjustment, is 0.60

between 2019-2022. This suggests that exposed tasks are, on average, in industries with slightly lower labor

shares than the national average.

This implies:

TFP gains over 10 years = GDP share impacted by AI over the next 10 years

× average cost savings of impact tasks.

Therefore, on the basis of the numbers from the previous subsection, I can approximate:

TFP gains over the next 10 years = 0.046 × 0.154

= 0.0071.

In other words, according to this basic estimation strategy, TFP gains over the next 10 years

from AI are about 0.71%—meaning that relative to the baseline without the current suite

of AI and computer vision advances, TFP will be higher by 0.71 percentage points in 10

years, or annual TFP growth will be higher by about 0.07%. This is a nontrivial, but modest

eﬀect, and certainly much less than both the revolutionary changes some are predicting and

the less hyperbolic but still substantial improvements forecast by Goldman Sachs and the

McKinsey Global Institute, which I discussed in the Introduction.

If we were to consider the higher productivity numbers from Peng et al., 0.154 would be

replaced by 0.205, and the 10-year TFP gains would be 0.94%, instead of 0.71%.

The only modiﬁcation that would make a sizable diﬀerence to these numbers is to increase

the fraction of tasks that will be impacted over the next 10 years. One way of doing this is to

inﬂate the numbers from Svanberg et al.. This could be because either the fraction of tasks

that can be feasibly automated will be diﬀerent for generative AI than for computer vision,

or because within 10 years this fraction will increase signiﬁcantly. For example, in Svanberg

et al.’s scenarios where costs for computer vision decline very rapidly, such as 10% a year, the

fraction of tasks that are feasibly automated may increase from 23% to approximately 30%.

This would raise the GDP share impacted by AI to approximately 6%, and correspondingly

increase the TFP gains over the next 10 years to about 1%. Note that even this number is

quite modest, and moreover, 10% cost declines for computer vision tasks is quite aggressive

(since, as already noted above, even if GPU costs were to decrease by 10% or even 20%,

this would not lead to a 10% decline in costs of performing computer vision tasks, given the

presence of other inputs, such as programming and data, as well as the inherent limitations

of the current generative AI architecture).

Finally, I note three important considerations missing from these computations.

1. These adoption numbers ignore the fact that there is still very little investment in AI

in the US corporate sector. Acemoglu et al. (2022) estimate that less than 1.5% of

US businesses had any investment in AI in 2019, and this is particularly true beyond

the very large companies in manufacturing, information services and business services.

Since many of the tasks considered in Eloundou et al. (2023) are performed in small

and medium-sized enterprises, this is unlikely to change quickly. If generative AI tools

become monopolized in the hands of a few companies, this might also slow down their

adoption by small and medium-sized ﬁrms. These considerations suggest that even the

0.045% number for the share of GDP impacted by AI may be a big overestimate, and

the true numbers could be much smaller.

2. Any major technology creates adjustment costs when adopted at large scale, because

other organizational aspects need to evolve as well and this is typically quite costly

and slow. In the context of digital technologies, Greenwood and Yorukoglu (1997) and

Brynjolfsson et al. (2021), among others, have argued that productivity gains will take

a J-shaped pattern, and the former paper predicts that the ﬂat part of the J-curve lasts

no less than 20 years for digital technologies. If so, the 15.4% overall cost reductions

may be a signiﬁcant overestimate for the next 10 years.

3. As already discussed above, some of the tasks in the list of Eloundou et al. (2023)

are hard-to-learn tasks, where productivity gains may be signiﬁcantly less than those

based on the experimental studies that have focused on the easy-to-learn tasks.

In the next subsection, I make a preliminary attempt at incorporating the third possi-

bility, but I will ignore the ﬁrst two. Nevertheless, these considerations make me conclude

that even the 0.71% increase in TFP within the next 10 years due to AI is likely to be an

upper bound on this technology’s medium-run eﬀects.

3.3 Aggregate Productivity Gains: Incorporating Hard Tasks

In this subsection, I reﬁne the estimates from the previous subsection by switching to equa-

tion (15), which can be rewritten as:

TFP gains = GDP share of impacted easy tasks × average cost savings from easy tasks

+GDP share of impacted hard tasks × average cost savings from hard tasks.

I take the 27% labor cost savings, from Noy and Zhang (2023) and Brynjolfsson et al.

(2023), to apply to easy tasks. The discussion in the previous section suggests that most

hard tasks have not been impacted or automated yet, and hence it is impossible to know

what their cost savings will be. Here I take those productivity gains to be 7%. My reasoning

is as follows. I consider the tasks involved in the Peng et al. (2023) study to be very easy

for generative AI for reasons explained above. Those studied in Noy and Zhang (2023) are

also on the easy side, and led to cost savings of about 40%, which is about two thirds of the

cost savings in Peng et al., while the customer service tasks in Brynjolfsson et al. (2023) are

already moving towards somewhat more complex tasks, and these had cost savings of only

14%. I imagine that many of the hard-to-learn tasks are more challenging for AI models than

the simpler end of the customer service tasks to which Brynjolfsson et al.’s numbers refer.

This motivates my choice of half of the cost savings of their study, 7% (which is also about

a quarter of the baseline 27% cost reduction estimate I used in the previous subsection).

The cost-saving numbers for both easy and hard tasks are again multiplied with 0.57 to

convert them into overall cost savings, which give 0.154 and 0.040 for easy and hard tasks,

respectively. To obtain the shares of easy and hard tasks, I start from Eloundou et al.’s data

and methodology, and then develop a procedure, implemented using GPT-4 like they do, for

sorting these into easy and hard tasks. The key characteristic for easy tasks is the presence

of a well-observed outcome and a straightforward rule that links actions/recommendations

to characteristics of the problem at hand.

To implement this procedure, I start from the 4089 exposed tasks determined from the

above procedure. Each one of these tasks has a statement on O*NET that includes at least

one verb. Each task also belongs to the higher level of aggregation of Detailed Work Activites

(DWAs) and the (even coarser) 332 Intermediate Work Activities (IWAs). The procedure

proceeds in four steps:

1. Classiﬁcation of verbs: Each task statement includes at least one verb which is

located at the beginning and can be easily identiﬁed. These primary verbs describe

much of what is “hard” or “easy” that needs to be learned about the task. We classiﬁed

tasks manually into easy and hard categories.

The full list of tasks is provided in the

Appendix. As an illustration, verbs for easy tasks include, among others: compute,

resolve, count, draft, grade, transcribe, classify, standardize, write, and record. Many

of these verbs are associated with simple actions that follow a clear set of steps and also

implicitly have a well-deﬁned metric for success (such as accounting or grading). In

contrast, verbs for hard tasks include: participate, advise, instruct, diagnose, educate,

hire, represent, testify, and care. The latter set of verbs describes more open-ended

This step and other manual coding steps in this subsection were carried out by Can Ye¸sildere.

activities for which there is less likely to be a clear metric of success. Yet other verbs,

such as analyze, maintain or inspect, do not fall into either of these categories, and are

coded as “uncertain”.

2. Classiﬁcation of IWAs: IWAs provide additional context for verbs, especially for

actions that verbs alone lack. We manually classify the 332 IWAs into easy and hard

tasks. The full list here is also given in the Appendix. Some of the easy IWAs are:

Evaluate project feasibility; Maintain sales or ﬁnancial records; Explain regulations,

policies, or procedures; Issue documentation; and Teach safety procedures or standards

to others. Some of the hard ones include: Monitor health conditions of humans or

animals; Evaluate scholarly work; Evaluate the quality or accuracy of data; Maintain

safety or security; and Train animals. These examples highlight the same principle

mentioned in the discussion of verbs—easy-to-learn activities are those for which there

is a clear metric of success and simple rules that can achieve this successful outcome,

while these are absent in hard-to-learn activities.

3. Latent Dirichlet Allocation (LDA) topic modeling: While tasks may share

IWAs and verbs, each task is also worded uniquely to capture the subject and the

more detailed description of activities. Consider the verb “review”. While “writing an

audit report” and “writing a letter of recommendation” are both associated with the

same verb and the same IWA, the subject matters of these tasks are quite diﬀerent.

To better disambiguate the contexts of these tasks, we use the LDA to allocate tasks

into clusters. LDA procedure is unsupervised and clusters tasks using the co-occurence

matrices of words extracted from each task statement. It assigns a probability to each

task belonging to a topic cluster. We feed all 19281 tasks into the algorithm and use

LDA to identify 100 clusters and the probability that each task belongs to one of those

100 clusters.

4. Final assignment: The ﬁnal step of our procedure is to derive a probability that

each exposed task is easy or hard. We ﬁrst manually classify a random sample of 500

exposed tasks as easy or hard. We then use the classiﬁcation of verbs, classiﬁcation of

IWAs, and the LDA-derived probabilities to train a gradient-boosted tree to match the

manual classiﬁcations of the training sample of tasks. We ﬁnally obtain a probabilistic

Of course, there is considerable ambiguity in some cases. For example, “Maintain safety and security”

also includes IT security activities, such as making sure that a new password can be issued to an authorized

person, and this would be an easy-for-AI task. The third step is aimed at dealing with such ambiguities.

assignment of each one of the 4089 exposed tasks into easy or hard task category from

this algorithm.

The end result is that 74% of wage bill-weighted exposed tasks in Eloundou et al. are easy.

This means that easy exposed tasks comprise 3.3% of GDP, while the remaining exposed

tasks (comprising 1.2% of GDP) are hard. Incorporating this information together with the

assumptions on cost savings yields the following tighter upper bound on TFP gains in the

next 10 years:

TFP gains over the next 10 years = 0.033 × 0.154 + (0.045 − 0.033) × 0.040

= .0055.

Unsurprisingly, this is smaller than the estimate in the previous subsection. Since automating

hard tasks—and even more so, introducing task complementarities for them—will be more

challenging, I view this estimate to be more reasonable. Either way, the TFP gains within

the next 10 years appear quite modest.

3.4 From TFP to GDP

Given the TFP eﬀects of AI over the 10 year horizon, it is straightforward to compute the

total GDP eﬀects of AI over the same horizon by using equation (13). Namely, this equation

can be written as:

GDP gains over the next 10 years = TFP gains over the next 10 years

+ capital share × proportional increase in capital stock.

As a simple benchmark, I start by assuming that the capital stock will grow proportion-

ately with TFP, which implies:

GDP gains over the next 10 years = TFP gains over the next 10 years /(1− capital share).

Using the capital share for the entire private business sector, 0.40, this implies that GDP

gains will be equal to the TFP gains multiplied by 1.66 (= 1/(1 − 0.4)). Hence taking the

As examples, the algorithm assigns a probability of 0.98 that “Maintaining equipment service records”

is an easy task, while only a probability of 0.098 that “Interviewing credit card applicants by telephone or

in person” is an easy task. Because writing and drafting are classiﬁed as verbs associated with easy tasks,

activities that involve writing are classiﬁed generally as easy. Some of this may understate the extent of the

diﬃculty of some writing-related tasks. For example, “Writing reports or academic papers to communicate

ﬁndings of climate-related studies” is classiﬁed to be an easy task with probability 0.67, which is likely to

be an underestimate of how hard such a task will continue to be for AI in the foreseeable future.

baseline estimate of an increase in TFP of 0.71%, I obtain a ﬁrst estimate for GDP growth

due to AI of 1.1% over 10 years, or taking the presence of hard tasks into account, a lower

estimate of 0.92%. Recall, however, that it is the TFP gains that are the relevant numbers

for consumer welfare, since the additional investment comes out of consumption and may

involve additional heavy energy use, as remarked in footnote 6.

I update the GDP eﬀects of AI advances in Section 3.6 when I model the between-task

and between-industry substitution patterns.

3.5 Consequences of New Bad Tasks

The calculations so far leave out the eﬀects of new tasks introduced thanks to AI (and equiv-

alently, the system-wide adjustments that AI may enable in some businesses, as emphasized

by Bresnahan (2019) and Agrawal et al. (2023). It is even more challenging to put numbers

on the eﬀects of new tasks. If AI helps create new tasks that increase productivity and espe-

cially contributes to the reinstatement of workers of diﬀerent skill levels into the production

process, its consequences can be much more positive.

Here my purpose is simply to point out that in the case of AI, because some of the new

tasks may be of the “bad” type, there may be an overstatement of the welfare gains from

AI when we look at GDP numbers. As an extremely preliminary attempt to argue that this

could be important, I will draw on two sources of data.

The ﬁrst is the recent study by Bursztyn et al. (2023), which provides suggestive esti-

mates for the extent of this problem in the case of social media. Bursztyn et al. (2023) run

an experiment to evaluate the welfare eﬀects of social media for college students, the demo-

graphic group most engaged with social media. They ask TikTok and Instagram users how

much they would need to be paid to not use the platform for a month. Users of these two

platforms are willing to pay up to $59 and $47 a month, respectively, to continue to use social

media—or on average $53 per user-month. However, they ﬁnd that users are also willing

to pay $28 and $10 every month to get everyone from their social network oﬀ TikTok and

Instagram, respectively—or an average $19 per user-month. If non-users are also included in

the analysis, the willingness to pay to stop everybody using social media increases to $47 for

TikTok and $13 for Instagram. This suggests that while companies can proﬁtably market

AI-based social media, the net eﬀect on welfare may be negative. Quantitatively, I ignore

the non-users (since the population in the study is college students, who are more likely to

be impacted by social media activity even when they do not use it than the adult population

in the United States). I also start with a benchmark in which AI-powered platforms can

capture the full (average) $53 willingness to pay per user (for example, because they are

eﬀectively price discriminating by the intensity with which they are collecting and monetiz-

ing users’ data). Then taking the average between the two platforms, I conclude that for

every $53 of revenue, there is a net negative eﬀect on users of $19. Put diﬀerently, the total

eﬀect from this class of new bad tasks, the equivalent of (18) in theory, is −19/53 ' −0.36

in proportionate terms. Note also that if I assume instead that social media companies can

capture less than the full $53 per user-month, then the denominator in this expression will

be smaller, and thus the proportionate damage per dollar of revenue would be higher. Hence,

the negative eﬀects of new bad tasks here should be interpreted as lower bounds.

I combine the number −0.36 with estimates of (i) revenues from social media and digital

ads, and (ii) spending on malicious IT attacks and IT security against these attacks to

compute apparent and real gains from new bad tasks. Speciﬁcally, the total revenues in 2022

of Meta (Facebook plus Instagram), Alphabet (Google and Youtube), Snapchat, TikTok and

Twitter (“X”) comes to 460 billion, or about 1.64% of US GDP,

while a lower bound on IT

security is 78 billion or 0.28% of US GDP.

Assuming that spending by malicious actors on

IT attacks is at least one third of this and combining these three estimates, I arrive to 2% of

US GDP. This number could be a signiﬁcant underestimate of other manipulative activities

enabled by new AI technologies. On the other hand, it could also overstate the problem, since

only a fraction of digital ad revenues will come from such manipulative activities (and this

may be particularly true for digital ad revenues from Google search). These considerations

suggest that the numbers here should be taken as merely suggestive.

Taking the 2% of US GDP as revenues and using the numbers from Bursztyn et al. (2023)

suggests that the total negative eﬀects of these manipulative activities is 0.02×0.36 = 0.072%

of GDP. This number points to a sizable negative impact on welfare, and could be larger

if manipulative uses of generative AI becomes more widespread. In contrast, if one were to

simply count the revenue coming from these new tasks, one might conclude that they would

increase GDP by 2%. This discussion thus suggests caution in interpreting all increases in

GDP coming from the use of generative AI as a positive eﬀect on consumer welfare.

The numbers for Meta ($130 billion), Alphabet ($307 billion) and Snapchat ($4.6 billion) are from these

companies’ 2023 10K ﬁlings. Fortune reported on December 12, 2023 that X’s annual revenue was $2.5

billion in 2023, while Financial Times reported on March 15, 2024 that revenues from TikTok’s US business

had reached $16 billion (see Fortune and Reuters).

See Statista

3.6 Wage and Inequality Implications

Finally, in this subsection I evaluate the wage and inequality consequences of generative

AI advances. To do this, the present framework needs to be extended to include multiple

demographic groups that have diﬀerent comparative advantages across diﬀerent tasks, as in

Acemoglu and Restrepo (2022). To save space, I do not introduce this generalization and

refer the reader to that paper. Instead, I start with the following equation from their paper,

which is a generalization of (11) above to a setting with multiple sectors and multiple demo-

graphic groups (but, for simplicity, without the task complementary and labor-augmenting

changes):

d ln w

= Θ



d ln y +

d ln ζ −

d ln Γ

auto



Here, g refers to demographic group g, and following their paper, I will focus on 500 demo-

graphic groups, deﬁned by education, age group, gender, ethnicity and native vs. foreign-

born status. In addition, d ln y is the change in GDP resulting from the technology change

(capturing the productivity eﬀect), σ is the elasticity of substitution between tasks, and

ζ is the vector of induced industry shifts (e.g., because automation of tasks in one indus-

try aﬀects prices and causes a reallocation of spending across sectors). Most importantly,

d ln Γ

auto

is the vector of demographic group-level displacement caused by the technology

shock—namely, it is a column vector of 500 entries, and Θ

is the gth-row vector of the

propagation matrix, which summarizes how the displacement impacting the other demo-

graphic groups aﬀects demographic group g (this is the reason why it is pre-multiplying the

eﬀects of industry shifts and displacements for all groups).

The propagation matrix represents the full “ripple eﬀects”—the impact of the displace-

ment of one demographic group on others, as they leave the tasks they were previously

performing and compete with other groups to be employed in other tasks. Such realloca-

tions are the key channel via which direct productivity gains for a group may end up harming

it at the end (as my example in the previous section illustrated). They are also the mecha-

nism via which the displacement of a demographic group may end up being more damaging

to another demographic group.

The ripple eﬀects estimated in Acemoglu and Restrepo (2022) may be context-speciﬁc—

meaning that the magnitudes of these eﬀects could be quite diﬀerent for the tasks impacted

by AI and automation technologies that their paper focuses on. Nevertheless, since it is

impossible to estimate these ripple eﬀects for the future impact of generative AI technologies,

Figure 1: Distribution of AI exposure across the wage distribution

Feasible AI exposure (%)

$10 $13.3 $17.8 $23.7 $31.6 $42 $57 $75

Hourly wage in 2018-2022 (logarithmic scale, 2020 dollars)

Less than high school High school

Some college College

Postgraduate

Notes: This ﬁgure depicts the AI exposure measure (both easy and hard tasks, combined) across 500 demographic groups.

The horizontal axis gives the average hourly wage of each demographic group between 2018 and 2022, computed from the ACS

5-year sample. Marker sizes are proportional to the average 2018-2022 employment level of each group and diﬀerent colors

indicate the education level of the group.

I will use their estimates.

An additional issue is that I have so far interpreted AI exposure to include both automa-

tion and task complementarities. In this subsection, I ignore the task complementarities and

presume that all of the AI exposure quantiﬁed so far will take the form of automation.

The methodology in Acemoglu and Restrepo (2022) starts from the d ln Γ

auto

vector,

which is at the demographics group level. To construct an equivalent of this measure, I take

the set of exposed occupations, and then use the wage bill shares of diﬀerent demographic

groups in these occupations to map the AI-generated displacement to the demographic group

level. For example, if for demographic group g, 5% of the wage bill share in 2019-2022

is in occupations that are fully exposed, then d ln Γ

auto

will be 0.05. I also assign these

occupations to industries using wage bill shares in order to compute industry-level impacts

If there were task complementarities as I have deﬁned above, this would induce some increases in the

productivity of exposed demographic groups in the remaining subtasks, and if so, it would also reduce

the automation-driven cost savings by a corresponding amount. If task complementarities and automation

aﬀected diﬀerent demographic groups within occupations symmetrically, this would have minor eﬀects on

the conclusions: although the wage eﬀects of productivity gains from task complementarities are a little

diﬀerent than the cost savings generated by automation (Acemoglu and Restrepo, 2022), the results would

remain broadly similar. If, on the other hand, some groups beneﬁted more from task complementarities

(for example, because they are overrepresented among middle-expertise workers that can most beneﬁt from

generative AI tools), the distributional consequences could be somewhat diﬀerent. Nevertheless, since I do

not have a way of distinguishing productivity eﬀects from test complementarities and automation, I am

unable to explore this issue further.

on costs and prices.

I compute the induced sectoral reallocations in the same way as

in Acemoglu and Restrepo (2022), using their parameterization of inter-sector elasticity of

substitution and the estimate of σ = 0.50 as in Humlum (2023) and take the elasticity of

substitution between sectors in consumption to be η = 0.2 as in Buera et al. (2022), which

was also imposed in Acemoglu and Restrepo (2022).

Figure 1 is the equivalent of Figure 5b in Acemoglu and Restrepo (2022) and presents

the distribution of AI exposure across demographic groups sorted by their hourly wage in

2018-22. It shows that AI exposure is much more equally distributed across demographic

groups than pre-AI automation (which was based on robotics, dedicated advanced machinery

and software systems).

Table 1 presents the main results. The seven rows are for the ﬁve education groups

(aggregating demographic groups according to their education level), for the average of

the workforce and for GDP. The ﬁrst column gives the AI exposure for each one of the

demographic groups, using our baseline exposure measure that does not distinguish easy

and hard tasks. The second column presents the direct impact of AI exposure on each one

of these groups, while the third column contains the full wage impact, taking into account

induced substitution between industries and the ripple eﬀects. The next four columns are

similar but now refer to the AI exposure measure that separates easy and hard tasks—column

4 gives exposure to easy tasks and column 5 is for exposure to hard tasks. Finally, column

8 shows the exposure measure from Acemoglu and Restrepo (2022) for comparison (but

recall that this measure refers to a 36-year period, rather than the 10-year timescale here).

The comparison of columns 1 and 4-5 to column 8 conﬁrms that AI exposure is much more

equally distributed across demographic groups than pre-AI automation exposure. Workers

with less than high school have the lowest exposure, while workers with college degrees and

those with associate degrees or some college have the highest exposure, and postgraduates

have the next lowest exposure.

This exercise uses the 49 BEA industries as in Acemoglu and Restrepo (2022) (but does not add the

public sector for which we have no exposure data).).

Baseline exposure Exposure adjusted for easy and hard task

(1) (2) (3) (4) (5) (6) (7) (8)

Baseline AI Direct eﬀect of Total wage eﬀect of Exposure to Exposure to Direct eﬀect of Total wage eﬀect of Direct task

exposure AI exposure AI exposure easy tasks hard tasks exposure to easy exposure to easy displacement

and hard tasks and hard tasks 1980-2016

Workers with less

than high school degree 0.0303 -0.0248 0.0201 0.0226 0.0076 -0.0281 0.0173 0.2690

Workers with

high school degree 0.0459 -0.0561 0.0124 0.0342 0.0117 -0.0594 0.0095 0.2706

Workers with

some college 0.0497 -0.0637 0.0143 0.0367 0.0130 -0.0670 0.0114 0.1886

Workers with

Bachelor’s degree 0.0525 -0.0693 0.0071 0.0382 0.0143 -0.0726 0.0049 0.0683

Workers with

postgraduate degree 0.0423 -0.0488 0.0108 0.0301 0.0122 -0.0522 0.0088 0.0343

Average

worker 0.0462 -0.0566 0.0117 0.0340 0.0122 -0.0599 0.0091 0.2107

GDP 0.0179 0.0162

Table 1: Exposure and wage eﬀects by education groups

Note: This table summarizes the eﬀects of AI exposure (with or without distinguishing easy and hard tasks) on the real wages of diﬀerent education groups and all workers. The

ﬁrst three columns are for the AI exposure measure that does not distinguish easy and hard tasks, while columns 4-7 is for the measure that introduces this distinction. Columns

1, 4 and 5 present levels of AI exposure for these diﬀerent cuts. Columns 2 and 6 provide the direct eﬀects of AI exposure, while columns 3 and 7 include the full equilibrium wage

impacts, incorporating the ripple eﬀects. Column 8, for comparison, includes the direct task displacement measure for 1980-2016 from Acemoglu and Restrepo (2022).

Figure 2: Decomposition of productivity eﬀects, industry shifts, direct displacement eﬀects

and ripple eﬀects

-.15

-.1

-.05

.05

.15

2.5 3 3.5 4 4.5

Hourly wage in 2018-2022

(log scale, 2020 dollars)

Less than high school

High school

Some college

College

Postgraduate

A. Productivity effect

-.15

-.1

-.05

.05

.15

2.5 3 3.5 4 4.5

Hourly wage in 2018-2022

(log scale, 2020 dollars)

B. Adding industry shifts

-.15

-.1

-.05

.05

.15

2.5 3 3.5 4 4.5

Hourly wage in 2018-2022

(log scale, 2020 dollars)

C. Adding direct task

displacement

-.15

-.1

-.05

.05

.15

2.5 3 3.5 4 4.5

Hourly wage in 2018-2022

(log scale, 2020 dollars)

D. Adding ripple effects

Notes: This ﬁgure is based on the estimates of the propagation matrix from Acemoglu and Restrepo (2022) and combines this

with the measure of exposure to easy and hard AI tasks in this paper. The ﬁrst panel includes just the productivity eﬀect.

The second panel adds the industry shifts induced by AI exposure. The third panel incorporates the direct displacement eﬀect,

while the ﬁnal panel adds the ripple eﬀects. The horizontal axis gives the average hourly wage for the relevant demographic

group between 2018 and 2022, computed from the ﬁve-year ACS sample. Marker sizes are proportional to the average 2018-2022

employment level of each group and diﬀerent colors indicate the education level of the group. See text for details.

Consequently, predicted wage impacts do not appear to increase inequality between edu-

cation groups. Focusing on the estimates in column 7, which incorporate easy and hard tasks,

there is a slightly higher wage growth for workers with less than a high school degree—about

1.2% within 10 years—but the gap between postgraduate versus high school and college

graduate workers also widens somewhat. In fact, the between-group standard deviation of

log wages (weighted by employment) increases slightly, from 0.35 to 0.36. The results are

quite similar in column 3 for the baseline AI exposure measure.

Finally, row 7 presents the estimate of the impact on GDP, taking into account the

equilibrium increasing the capital stock implied by the model. This is under the assumption

that all of the capital stock adjustment will take place within 10 years (while in practice

it may take longer) and that the required rate of return on capital investments does not

change (whereas with a sizable investment, we may expect an increase). This leads to an

upper bound GDP impact of 1.62% in column 7, when I distinguish between easy and hard

Figure 3: Total wage eﬀect of exposure to AI, by gender

Notes: This ﬁgure is based on the estimates of the propagation matrix from Acemoglu and Restrepo (2022) and combines

this with the measure of exposure to easy and hard AI tasks in this paper. Each panel includes wage eﬀect estimates for ﬁve

education groups. Reported estimates are weighted averages of the estimates for the more detailed subgroups (using average

employment 2018-2022 as weights). The upper left panel is for native-born white men, the lower left is for all other men, the

upper right is for native-born white women and the lower right is for all other women. See text for estimation details.

tasks. GDP therefore increases substantially more than average wages, and as a result, the

capital share of national income increases by about 0.38 percentage points. This conﬁrms

that inequality between capital and labor is likely to rise as a result of the rollout of AI.

Figure 2 performs the same exercise as Figure 7 in Acemoglu and Restrepo (2022),

focusing on the exposure measure that distinguishes between easy and hard tasks. As in

that paper, productivity eﬀects are (by construction) uniform across groups, and there is also

not much inequality generated by the cross-industry shifts shown in the second panel. The

third panel conﬁrms that AI’s direct eﬀects are more equally distributed across demographic

groups and throughout the wage distribution. In contrast to the ﬁndings in Acemoglu and

Restrepo (2022), ripple eﬀects do not change the inequality patterns by much, largely because

the direct eﬀects are already fairly equally distributed.

Figure 2 also indicates that there is a lot of variability in the experience of low-education

groups—and this is the reason why the between-group standard deviation of log wages

increases, as shown in Table 1. Figure 3 explores this issue further by depicting the real

wage changes of ﬁner groupings, distinguished by gender, by education and by white and

native-born status versus the rest. It reveals that low-education women, especially white,

native-born low-education women, are likely to experience declines in real wages as a result

of AI.

Overall, this exercise suggests that the inequality consequences of AI will not be as

adverse as pre-AI automation, because AI exposure is more equally distributed across de-

mographic groups. Nevertheless, there is no evidence that AI will reduce inequality between

demographic groups, as some are forecasting. Rather, my analysis suggests that it may have

a small positive eﬀect on overall (between-group) inequality and reduce the real earnings of

low-education women. It will also further widen the gap between capital and labor income.

4 Concluding Remarks and Discussion

Following its release on November 30, 2022, ChatGPT became the fastest spreading tech plat-

form in history, reaching approximately 100 million monthly users within just two months.

Its impressive features, and the greater capabilities of the newer version ChatGPT-4, released

in March 2023, soon captured imaginations, both among the general public and economic

commentators. Forecasts of large productivity gains have now become commonplace.

While there is no question that generative AI models, including ChatGPT, have impres-

sive achievements and have great potential for beneﬁcial economic eﬀects, the extent of their

macroeconomic consequences remains an open question.

There are four potential types of macro eﬀects that AI technologies can have in the

medium run:

1. They can quickly revolutionize every aspect of the economy, and lead to massive im-

provements in productivity, even taking us close to “singularity”. While this is a

possibility that cannot be completely ruled out, there is so far no evidence of such

revolutionary eﬀects, and these are not addressed in the current paper.

2. They can have more modest but still notable eﬀects on the macroeconomy by improving

productivity and reducing costs in a range of tasks. Some of the forecasts have focused

on these types of improvements and still produced relatively large numbers, such as a

1.5 − 3.4 percentage point per annum increase in economic growth within a 10 year

horizon.

3. They can impact wages and inequality because of their automation eﬀects, or con-

versely, lead to large wage increases, especially for lower-pay workers, as forecast by

The Economist (2023).

4. They can have macroeconomic eﬀects by producing deepfakes, misinformation, manip-

ulation and other “bads”.

In this paper, I use the task-based framework of Acemoglu and Restrepo (2018, 2019b,

2022) to evaluate the second and the third eﬀects, and I also take some steps to formalize

how the fourth set of eﬀects might work out in a task-based macro framework.

I base my approach on existing experimental studies that estimate productivity gains and

time savings from the use of generative AI tools in a number of settings. By building on these

studies, I am explicitly taking on board the idea that generative AI will lead to cost savings

and productivity improvements. Nevertheless, combining these numbers with estimates of

exposed tasks from Eloundou et al. (2023) and Svanberg et al. (2024) leads to much more

modest productivity eﬀects than most commentators and economists have claimed so far.

These numbers become even smaller once we take into account that many of the tasks for

which we have evidence of cost savings are relatively easy for AI, while in several other tasks

the integration of AI will face more formidable diﬃculties—mostly because these are likely

to involve more complex interactions between action and context and because they lack clear

metrics for success that are observable, and hence necessitate AI models to learn from the

(average) behavior of humans previously performing the same tasks.

Taking these considerations into account, I estimate that TFP eﬀects from AI advances

within the next 10 years will be modest—an upper bound that does not take into account

the distinction between hard and easy tasks would be about a 0.71% increase in total within

10 years, or about a 0.07% increase in annual TFP growth. When the presence of hard

tasks among those that will be exposed to AI is recognized, this upper bound drops to

about 0.55%. GDP eﬀects will be somewhat larger than this because automation and task

complementarities will also lead to greater investment. But my calculations suggest that the

GDP boost within the next 10 years should also be modest, in the range of 0.9% − 1.1% over

10 years in total, provided that the investment increase resulting from AI is modest, and in

the range of 1.6% − 1.8% in total, if there is a large investment boom.

If AI is used to create new tasks and products, these will also add to GDP and can

boost productivity growth. Nevertheless, when we incorporate the possibility that new

tasks generated by AI may be manipulative, the impact on welfare can be even smaller.

Based on numbers from Bursztyn et al. (2023), which pertain to the negative eﬀects of AI-

powered social media, I provide an illustrative calculation for social media, digital ads and

IT defense-attack spending. These could add to GDP by as much as 2%, but if we apply

the numbers from Bursztyn et al. (2023), their impact on welfare may be −0.72%. This

discussion suggests that it is important to consider the potential negative implications of

AI-generated new tasks and products on welfare.

Finally, I borrow heavily from the estimates of Acemoglu and Restrepo (2022) on the

economy-wide productivity, wage and inequality eﬀects of pre-AI automation technologies

to provide some guidance on what the impact of new AI advances will be. Because AI-

exposed tasks are more equally distributed within the population than tasks exposed to

pre-AI automation, I do not ﬁnd substantial negative wage eﬀects for any education group.

Nevertheless, the estimates do not point to signiﬁcant declines in inequality either and in

fact, my ﬁndings suggest that the real wages for low-education women may decline, overall

between-group inequality may increase slightly, and the gap between capital and labor income

is likely to widen further.

These results should not be interpreted as arguing that there are no major beneﬁts from

AI. First, an increase of about 0.55 − 0.71% in TFP within 10 years is modest but still far

from trivial. Second and more importantly, there may be other ways in which AI can be used

to generate much more notable beneﬁts. I have suggested in previous work (Acemoglu, 2021,

Acemoglu and Restrepo, 2020b) that if AI is used for generating new tasks for workers, it can

have more beneﬁcial productivity, wage and inequality consequences, and it can even increase

wages. This may be doubly true for generative AI, which could be used for providing better

information to workers and boosting their expertise, as argued in Acemoglu et al. (2023)

and explained brieﬂy here.

Many production workers today, including electricians, repair workers, plumbers, nurses,

educators, clerical workers, and increasingly many blue-collar workers in factories, are en-

gaged in problem-solving tasks. These tasks require real-time, context-dependent and reli-

able information. For instance, an electrician dealing with the malfunctioning of advanced

equipment or a short-circuit on the electricity grid will be hampered from solving these prob-

lems because he or she does not have suﬃcient expertise and the appropriate information for

troubleshooting. Reliable information that can be provided quickly by generative AI tools

can lead to signiﬁcant improvements in productivity. Similarly, generative AI in classrooms

can lead to a major reorganization of how teaching takes place, with much greater levels of

personalization, as these tools help teachers identify speciﬁc aspects of the curriculum with

which subgroups of students are having problems and propose new context-dependent teach-

ing strategies. Once again, reliability (as well as data privacy and protection) are key for

successfully creating such new tasks and delivering improvements in the quality of education.

Productivity improvements from new tasks are not incorporated into my estimates. This

is for three reasons. First and most parochially, this is much harder to measure and is not

included in the types of exposure considered in Eloundou et al. (2023) and Svanberg et al.

(2024). Second, and more importantly, I believe it is right not to include these in the likely

macroeconomic eﬀects, because these are not the areas receiving attention from the industry

at the moment, as also argued in Acemoglu (2021), Acemoglu and Restrepo (2020b) and

Acemoglu and Johnson (2023). Rather, areas of priority for the tech industry appear to be

around automation and online monetization, such as through search or social media digital

ads. Third, and relatedly, more beneﬁcial outcomes may require new institutions, policies

and regulations, as also suggested in Acemoglu and Johnson (2023) and Acemoglu et al.

(2023).

My assessment is that there are indeed much bigger gains to be had from generative

AI, which is a promising technology, but these gains will remain elusive unless there is a

fundamental reorientation of the industry, including perhaps a major change in the archi-

tecture of the most common generative AI models, such as the LLMs, in order to focus on

reliable information that can increase the marginal productivity of diﬀerent kinds of workers,

rather than prioritizing the development of general human-like conversational tools. This

major architectural change may be especially necessary because the new task beneﬁts will

crucially depend on providing reliable information to workers, and the very general-purpose

nature of the current approach to generative AI could be ill-suited for this purpose. To put

it simply, it remains an open question whether we need foundation models (or the current

kind of LLMs) that can engage in human-like conversations and write Shakespearean son-

nets if what we want is reliable information useful for educators, healthcare professionals,

electricians, plumbers and other craft workers.

References

Acemoglu, Daron, “AI’s Future Doesn’t Have to Be Dystopian.,” Boston Review, 2021.

https://www.bostonreview.net/forum/ais-future-doesnt-have-to-be-dystopian/.

and David Autor, “Skills, Tasks and Technologies: Implications for Employment and

Earnings,” in “Handbook of Labor Economics,” Vol. 4B, Elsevier, 2011, pp. 1043–1171.

edited by David Card and Orley Ashenfelter. Amsterdam: North-Holland.

and Pascual Restrepo, “The Race between Man and Machine: Implications of Tech-

nology for Growth, Factor Shares, and Employment,” American Economic Review, 2018,

108 (6), 1488–1542.

and , “Artiﬁcial Intelligence, Automation, and Work,” The Economics of Artiﬁcal

Intelligence: An Agenda, 2019, pp. 197–236. Edited by Ajay Agrawal, Joshua Gans and

Avi Goldfarb. University of Chicago Press.

and , “Automation and New Tasks: How Technology Displaces and Reinstates 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.

and , “The Wrong Kind of AI? Artiﬁcial Intelligence and the Future of Labour De-

mand,” Cambridge Journal of Regions, Economy and Society, 2020, 13 (1), 25–35.

and , “Tasks, Automation, and the Rise in US Wage Inequality,” Econometrica, 2022,

90 (5), 1973–2016.

and , “Automation and Rent Dissipation: Implications for Inequality, Productivity,

and Welfare,” 2023. Working Paper.

and Simon Johnson, Power and Progress: Our Thousand-Year Struggle Over Technol-

ogy and Prosperity, London, UK: Hachette, 2023.

, David Autor, and Simon Johnson, “Can We Have Pro-Worker AI?,” Technical

Report, MIT Shaping the Future of Work Initiative 2023.

, Gary W Anderson, David N Beede, Cathy Buﬃngton, Eric E Childress, Emin

Dinlersoz, Lucia S Foster, Nathan Goldschlag, John C Haltiwanger, Zachary

Kroﬀ et al., “Automation and the Workforce: A Firm-Level View from the 2019 Annual

Business Survey,” Technical Report, National Bureau of Economic Research 2022.

, Martin K. Jensen, and Pascual Restrepo, “Technology and Wages in the Long

Run,” 2024.

Agrawal, Ajay, Joshua S Gans, and Avi Goldfarb, “Artiﬁcial Intelligence Adoption

and System-Wide Change,” Journal of Economics & Management Strategy, 2023, pp. 1–

11.

Autor, David, “Applying AI to Rebuild Middle Class Jobs,” Technical Report, National

Bureau of Economic Research 2024.

, Caroline Chin, Anna Salomons, and Bryan Seegmiller, “New Frontiers: The

Origins and Content of New Work, 1940–2018,” The Quarterly Journal of Economics,

2024, p. Forthcoming.

Autor, David H, Frank Levy, and Richard J Murnane, “The Skill Content of Recent

Technological Change: An Empirical Exploration,” The Quarterly Journal of Economics,

2003, 118 (4), 1279–1333.

Besiroglu, Tamay and Marius Hobbhahn, “Trends in GPU Price-Performance,” Tech-

nical Report, EpochAI 2022. https://epochai.org/blog/trends-in-gpu-price-performance.

Bresnahan, Timothy, “Artiﬁcial Intelligence Technologies and Aggregate Growth

Prospects,” Prospects for Economic Growth in the United States, 2019, pp. 132–172.

Brynjolfsson, Erik, Daniel Rock, and Chad Syverson, “The Productivity J-curve:

How Intangibles Complement General Purpose Technologies,” American Economic Jour-

nal: Macroeconomics, 2021, 13 (1), 333–372.

, Danielle Li, and Lindsey R Raymond, “Generative AI at Work,” Technical Report,

National Bureau of Economic Research 2023.

Buera, Francisco J, Joseph P Kaboski, Richard Rogerson, and Juan I Vizcaino,

“Skill-Biased Structural Change,” The Review of Economic Studies, 2022, 89 (2), 592–625.

Bursztyn, Leonardo, Benjamin R Handel, Rafael Jimenez, and Christopher

Roth, “When Product Markets Become Collective Traps: The Case of Social Media,”

Technical Report, National Bureau of Economic Research 2023.

Chui, Michael, Eric Hazan, Roger Roberts, Alex Singla, Kate Smaje,

Alex Sukharevsky, Lareina Yee, and Rodney Zemmel, “The Economic Po-

tential of Generative AI: The Next Productivity Frontier,” McKinsey & Com-

pany, 2023. https://www.mckinsey.com/capabilities/mckinsey-digital/our-insights/the-

economic-potential-of-generative-AI-the-next-productivity-frontier#business-value.

Davidson, Tom, “Could Advanced AI Drive Explosive Economic Growth?,” Open Philan-

thropy, June 2021. https://www.openphilanthropy.org/research/could-advanced-ai-drive-

explosive-economic-growth/.

Eloundou, Tyna, Sam Manning, Pamela Mishkin, and Daniel Rock, “GPTs are

GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models,”

Technical Report, arXiv 2023.

Goldman Sachs, “Generative AI could raise global GDP by 7 percent,” 2023.

https://www.goldmansachs.com/intelligence/pages/generative-ai-could-raise-global-

gdp-by-7-percent.html.

Greenwood, Jeremy and Mehmet Yorukoglu, “1974,” in “Carnegie-Rochester confer-

ence series on public policy,” Vol. 46 Elsevier 1997, pp. 49–95.

Hulten, Charles R, “Growth Accounting with Intermediate Inputs,” The Review of Eco-

nomic Studies, 1978, 45 (3), 511–518.

Humlum, Anders, “Robot Adoption and Labor Market Dynamics,” 2023. Dissertation,

Princeton University.

Korinek, Anton and Donghyun Suh, “Scenarios for the Transition to AGI,” Technical

Report, National Bureau of Economic Research 2024.

Kurzweil, Ray, The Singularity Is Near: When Humans Transcend Biology, New York

City, USA: Viking Press, 2005.

Noy, Shakked and Whitney Zhang, “Experimental Evidence on the Productivity Eﬀects

of Generative Artiﬁcial Intelligence,” Science, 2023, 381 (6654), 187–192.

Peng, Sida, Eirini Kalliamvakou, Peter Cihon, and Mert Demirer, “The Impact of

AI on Developer Productivity: Evidence from Github Copilot,” Technical Report, arXiv

2023.

Susskind, Daniel, A World without Work: Technology, Automation and How We Should

Respond, London, UK: Penguin, 2020.

Svanberg, Maja, Wensu Li, Martin Fleming, Brian Goehring, and Neil Thomp-

son, “Beyond AI Exposure: Which Tasks are Cost-Eﬀective to Automate with Computer

Vision?,” 2024. Working Paper.

The Economist, “Blue-Collar Bonanza: Why Conventional Wisdom on Inequality is

Wrong,” December 2023. https://www.economist.com/weeklyedition/2023-12-02.