Broadening the Gains

from Generative AI: The

Role of Fiscal Policies
Prepared by Fernanda Brollo, Era Dabla-Norris, Ruud de Mooij,
Daniel Garcia-Macia, Tibor Hanappi, Li Liu, and Anh Dinh Minh
Nguyen

SDN/2024/002

IMF Staff Discussion Notes (SDNs) showcase

policy-related analysis and research being
developed by IMF staff members and are
published to elicit comments and to encourage
debate. The views expressed in Staff Discussion
Notes are those of the author(s) and do not
necessarily represent the views of the IMF,
its Executive Board, or IMF management.

2024
JUNE
STAFF DISCUSSION NOTES Broadening the Gains from Generative AI: The Role of Fiscal Policies

©2024 International Monetary Fund SDN/2024/002

IMF Staff Discussion Notes

Fiscal Affairs Department

Broadening the Gains from Generative AI: The Role of Fiscal Policies

Prepared by Fernanda Brollo, Era Dabla-Norris, Ruud de Mooij, Daniel Garcia-Macia, Tibor Hanappi,
Li Liu, and Anh Dinh Minh Nguyen *

Authorized for distribution by Vitor Gaspar

June 2024

IMF Staff Discussion Notes (SDNs) showcase policy-related analysis and research being
developed by IMF staff members and are published to elicit comments and to encourage debate.
The views expressed in Staff Discussion Notes are those of the author(s) and do not necessarily
represent the views of the IMF, its Executive Board, or IMF management.

ABSTRACT: Generative artificial intelligence (gen AI) holds immense potential to boost productivity growth and
advance public service delivery, but it also raises profound concerns about massive labor disruptions and rising
inequality. This note discusses how fiscal policies can be employed to steer the technology and its deployment
in ways that serve humanity best while cushioning the negative labor market and distributional effects to
broaden the gains. Given the vast uncertainty about the nature, impact, and speed of developments in gen AI,
governments should take an agile approach that prepares them for both business as usual and highly
disruptive scenarios.

Recommended Citation: Brollo and Others. 2024. “Broadening the Gains from Generative AI: The Role of
Fiscal Policies.” IMF Staff Discussion Note SDN2024/002, International Monetary Fund, Washington, DC.

ISBN: 979-8-40027-717-7

JEL Classification Numbers: E24, H2, J24, O33

Artificial intelligence, AI, technological change, labor market, fiscal

Keywords:
policy, social protection systems, tax systems

[email protected]; [email protected]; [email protected];

Author’s E-Mail Address: [email protected]; [email protected]; [email protected];
[email protected].

* Preparation of this SDN was led by Era Dabla-Norris and Ruud de Mooij. The authors thank Gita Gopinath and Vitor

Gaspar for feedback and guidance and many IMF colleagues for useful comments. The views expressed herein are
those of the authors and should not be attributed to the IMF, its Executive Board, or its management.
STAFF DISCUSSION NOTES Broadening the Gains from Generative AI: The Role of Fiscal Policies

Contents
Executive Summary _______________________________________________________________________________ 1
Introduction ______________________________________________________________________________________ 3
Upgrading Social Protection Systems _____________________________________________________________ 6
Upgrading Tax Systems __________________________________________________________________________ 14
Funding for AI Innovation and Deployment______________________________________________________ 23
Conclusions and Policy Implications _____________________________________________________________ 24

REFERENCES ____________________________________________________________________________________ 37

FIGURES
1. Effect of Robots on Employment and Wages in the U.S. Local Labor Market: The Role of
Unemployment Insurance_________________________________________________________________________ 7
2. Effect of Robots on Wages in the U.S. Local Labor Market: The Role of Unemployment Insurance
by Levels of Education ____________________________________________________________________________ 7
3. Effect of Robots on Poverty: The Role of Social Assistance _____________________________________ 8
4. Baseline Automation Shocks __________________________________________________________________ 10
5. Different UI Programs _________________________________________________________________________ 11
6. Transition Dynamics ___________________________________________________________________________ 11
7. Illustrative Scenarios for EMDEs _______________________________________________________________ 12
8 Maximum Benefit Duration ____________________________________________________________________ 13
9. Corporate Tax Biases for Labor-Saving Assets _________________________________________________ 16
10. Corporate Tax Bias for Labor-Saving Assets by Country ______________________________________ 16
11. Changing Corporate Tax Bias ________________________________________________________________ 17
12. Taxing Automation versus Taxing Labor Income to Finance Unemployment Support ________19
13. Concentration of Capital Income Among Top Earners: Cross-Country Evidence ______________20
14. Average Tax Rates on Labor and Capital Income, 5-year Moving Average ___________________21
15. Income Share by Source and Income Groups in the US _____________________________________ 22

ANNEXES
Annex 1. Model Framework ______________________________________________________________________ 28
Annex 2. Effective Tax Rates _____________________________________________________________________ 32
Annex 3. Corporate Taxes and Labor Income Share ______________________________________________ 35

INTERNATIONAL MONETARY FUND 1

STAFF DISCUSSION NOTES Broadening the Gains from Generative AI: The Role of Fiscal Policies

Executive Summary
Rapid advances in generative artificial intelligence (gen AI) hold immense potential to transform
production processes and significantly accelerate productivity growth. Gen AI has the power to
revolutionize information availability and utilization, enabling governments to advance revenue
mobilization and deliver more efficient public services across sectors, including health care, education,
public procurement, and social transfers.

Alongside its promise, gen AI also presents challenges. A critical distinction between gen AI and past
disruptive technologies (such as the steam engine, electricity, and early computers) lies in its potential for
rapid diffusion. The sheer scale and speed of the transformation pose risks to labor markets. While
automation and robots have already displaced low- and middle-skill jobs involving routine tasks, gen AI’s
capabilities extend to more intelligent automation, potentially amplifying job losses in cognitive
occupations. Consequently, the labor income share in national income may further decline, exacerbating
income and wealth inequality. Dominant firms in increasingly concentrated markets could reinforce their
market power and enjoy monopoly rents.

This note provides analysis and guidance for policymakers as they prepare for the transformative impact
of gen AI. Special taxes on AI to reduce the speed of AI investment are not recommended as they can be
hard to operationalize and hamper productivity growth. That said, corporate tax incentives that currently
encourage rapid labor displacement, prevalent in several advanced economies, should be reconsidered
because they magnify the social costs of excessive labor market dislocation. At the same time, corporate
tax distortions that hold up labor-saving investments, which are more prevalent in developing economies,
can also be costly and especially harmful in less-disruptive labor market scenarios. General taxes on
capital income, which have systematically declined across the world during past decades, should be
strengthened to protect the tax base against a further decline in the labor-income share and to offset
rising wealth inequality.

Fiscal policies can also cushion the negative labor market and distributional effects of gen AI and help
distribute the gains more evenly. This calls for adequate social protection systems (social insurance, labor
market programs, social assistance), as new analysis in this note shows. Most countries have scope to
broaden the coverage and generosity of unemployment insurance, improve portability of entitlements,
and consider forms of wage insurance. Combined with active labor market policies, this can help workers
manage the transition while adapting to changing skill requirements. Innovative approaches that enhance
digital technologies can facilitate expanded coverage of social assistance programs, particularly for those
who suffer a prolonged impact from transitions or who work in the informal sector in emerging market and
developing economies. Education and training policies must adapt to new realities, help prepare workers
for the jobs of the future, and be better placed to offer lifelong learning. Sector-based training,
apprenticeship, and upskilling and re-skilling programs could play a greater role in helping workers
transition to new tasks and sectors.

Now that AI has matured to the commercial adoption phase, public funding should focus on areas less
likely to receive private sector investment—such as fundamental research, necessary infrastructure

INTERNATIONAL MONETARY FUND 1

STAFF DISCUSSION NOTES Broadening the Gains from Generative AI: The Role of Fiscal Policies

(particularly in emerging market and developing economies), and applications in the public sector
(education, health care, government administration).

Given uncertainty surrounding the transformative nature, impact, and pace of gen AI, policymakers must
remain agile. Policy should bring about conditions that steer innovation and deployment in ways that
harness the benefits of gen AI and serve collective human interests, and it should be ready to cushion the
transition costs for workers and households and prevent rising inequality. Fiscal policies thus need to
prepare for both businesses as usual and highly disruptive scenarios.

INTERNATIONAL MONETARY FUND 2

STAFF DISCUSSION NOTES Broadening the Gains from Generative AI: The Role of Fiscal Policies

“It is coming inhumanely fast, and it will seem unbelievably unfair.”

Richard Baldwin (2019)

Introduction
Rise of “cognitive” automation. 1 Artificial intelligence (AI) holds the potential to transform the nature of
production processes and lift productivity and growth. Recent rapid advances in AI, particularly the advent
of generative AI (gen AI) based on large language models that can produce new content, are greatly
expanding the set of activities that can be performed more efficiently by computers than by humans. As a
result, AI’s potential applications in the workplace are broadening. Unlike past industrial transformations
driven by general purpose technologies such as the steam engine, electricity, and early computers—gen
AI can proliferate much faster than previous disruptive technologies. Even as businesses are working to
figure out the best way to deploy the current generation of AI, advances are happening at breakneck
speed, and the potential impacts are uncertain.

Untold benefits. Gen AI could revolutionize business and public sector operations. For firms and
industries, it holds the potential for new revenue, cost savings, and improved products and processes.
For governments, it portends improvements in public service delivery, making them more efficient and
effective. For instance, fiscal operations, procurement, and revenue collection could benefit from
enhanced fraud detection and automated audit and assurance processes. Personalized interactive
learning, augmented reality, and remote patient monitoring could radically transform public education and
health care, helping services reach people more quickly and equitably. Gen AI–fueled advances in public
sector operations could subsequently inform more effective policy design and help transform regulatory
operations.

Impact on labor markets and inequality. Adoption of gen AI will likely be uneven, and the size and
rapid speed of transformation risk disrupting labor markets. Many jobs involving routine tasks have
already been eliminated through laborsaving automation, which has driven down average wages and
increasingly polarized wages and employment, even with the advent of new forms of work. 2 Evidence
suggests that while earlier automation waves displaced mostly blue-collar (lower-skilled) workers, white-
collar (high-skilled) workers are most exposed to AI. 3 But AI is also capable of powering more intelligent
robots and could lead to further automation of blue-collar jobs. As such, laborsaving automation could
amplify job losses in both low-skill and cognitive occupations, further reducing the labor share and wages
relative to capital and exacerbating income and wealth inequality. Although AI could increase the
productivity of firms and thus raise the demand for labor in nonautomated tasks—and generate new

1 This note defines automation as technological advances that replace human labor with machine labor. Progress in AI may lead to
greater automation, and the speed of this change may significantly increase after the advent of generative artificial intelligence (gen
AI).
2
Automation is hardly a novel phenomenon. Traditional sectors such as agriculture and manufacturing have already experienced
large substitutions of labor with machine capital. Historically, jobs displaced by automation have been offset by the creation of new
jobs, but computerization of white-collar services in many advanced economies has accelerated in recent years (Acemoglu and
Restrepo 2018).
3
Webb (2020) finds that AI is directed mainly at high-skill tasks and will affect highly educated and older workers. Felten, Raj, and
Seamans (2023) obtain analogous findings when restricting their measure to machine learning applications more related to gen AI,
such as language models and image generation. See also Eloundou and others (2023); Pizzinelli and others (2023); Autor (2024);
and Brussevich, Dabla-Norris, and Khalid (forthcoming).

INTERNATIONAL MONETARY FUND 3

STAFF DISCUSSION NOTES Broadening the Gains from Generative AI: The Role of Fiscal Policies

jobs—the transition could be costly. Gen AI could also reinforce the more unequal distribution of capital
income seen in recent decades: rising market power and economic rents enjoyed by dominant firms in
increasingly concentrated and winner-take-all markets (Kehrig and Vincent 2021; Acemoglu and Johnson
2023).

Alleviating costly transitions and broadening gains. Adequately distributing the gains and
opportunities is necessary not just to protect the vulnerable and ensure social cohesion, but also to fully
harness the opportunities presented by gen AI. By offering financial support during unemployment,
promoting new skills acquisition, and creating a safety net, social protection systems can help individuals
adapt to job market changes. Traditionally, social protection systems have encompassed payroll-based
insurance components (for example, unemployment benefits), lifelong education and training initiatives,
and cash transfers and other forms of noncontributory social assistance programs. These all provide
financial support to low-income households during long periods of unemployment. A largely unexplored
question is whether and how social policies will have to be reimagined in the face of disruptive
technological changes from AI.

Taxation of investment in AI: a growing debate. Should investment in AI be taxed? Taxation,

alongside regulation, can mitigate the disruptive labor market implications of rapid job displacements by
discouraging and slowing the deployment of automation. The downside of such taxes is, however, that
they distort productivity-enhancing investment and reduce economic growth. While the net balance for
social welfare is unclear and may vary between scenarios, an important question is how such a tax could
be implemented in practice.

Increasing importance of taxing capital income. Gen AI, like other types of innovation, can lead to
higher top-income inequality. A long-held view is that progressive income taxes can help address rising
inequality, including through the taxation of capital income, while balancing the trade-off with efficiency.
Also, higher investment in education and social spending to broaden the gains from AI require higher
public revenue, while labor substitution can do the opposite if capital income is taxed less than labor
income. Developing economies specializing in labor-intensive sectors and exposed to “reshoring” are
particularly at risk of losing tax revenue (Korinek, Schindler, and Stiglitz 2022).

Steering innovation in AI. Fiscal policies may also influence the path of innovation and deployment of
AI. Some have argued that policies could also favor applications that expand, rather than substitute for,
human capabilities and could lead to new occupational tasks (Acemoglu, Autor, and Johnson 2023).
While it is unclear what this entails in practice, development opportunities in emerging markets from AI
deployment could be significant.

This note. This note focuses on the role of fiscal policies in supporting a more equal distribution of gains
and opportunities from gen AI. A key question is how countries can use spending policies and reform tax
systems to mitigate labor disruptions during the transition and offset adverse distributional impacts of
innovation while preserving AI-driven productivity growth. Specifically, the note addresses the following
four questions:

INTERNATIONAL MONETARY FUND 4

STAFF DISCUSSION NOTES Broadening the Gains from Generative AI: The Role of Fiscal Policies

 How have social protection systems helped reduce labor market disruption during past episodes of
automation?

 Looking ahead, how can countries strengthen social spending during rapid technological transitions?

 Have tax systems provided excessive incentives for automation? Should automation be taxed to
mitigate labor market disruptions and pay for its effects on workers?

 How should governments design redistributive taxation in the face of inequality and winner-take-all
dynamics from gen AI—especially taxes on capital income?

Previous work and contribution to literature. Recent studies have focused on the productivity and
labor market impacts of AI (Korinek, Schindler, and Stiglitz 2022; Korinek 2023a; Baily, Brynjolfsson, and
Korinek 2023; Brynjolfsson, Li, and Raymond 2023; Noy and Zhang 2023; Cazzaniga and others 2024). 4
Fiscal policies aimed at spreading the gains and mitigating the risks from AI are less well studied. This
note builds on some pioneering studies in this field, including Berg and others (2021); Beraja and Zorzi
(2024); Costinot and Werning (2023); Guerreiro, Rebelo, and Teles (2022); and Thuemmel (2023). Our
contributions focus on social protection and tax policies. In particular, this note offers new empirical
analysis of the role of social protection systems during past automation waves and discusses the
desirable characteristics of social spending in the face of disruptive technological developments. Model
simulations illustrate the impact of spending and tax policies on labor market outcomes and welfare. The
note also presents a novel discussion of how current tax systems affect firms’ decisions to invest in labor-
displacing capital assets, examines the case for taxing AI, and elaborates on recommendations to
enhance the taxation of capital income. Finally, the note briefly touches on whether fiscal policies should
promote innovation and deployment of gen AI.

Caveats. How gen AI technologies will evolve and transform economies is highly uncertain. Different
scenarios are plausible regarding (1) the speed of improvement in the capabilities of AI, (2) the degree of
adoption of the newest technologies across countries and firms and how they will be used, (3) the extent
to which AI will replace or complement different types of workers, (4) how people will adapt to the new
realities of work, (5) the policy responses of governments, and (6) the implications of these factors for
productivity growth and economic well-being. Fiscal policies must adapt to changing conditions and
prepare for both business as usual and highly disruptive scenarios.

4
See also Aghion and others (2022), Autor (2022), and Comunale and Manera (2024) for comprehensive literature reviews.

INTERNATIONAL MONETARY FUND 5

STAFF DISCUSSION NOTES Broadening the Gains from Generative AI: The Role of Fiscal Policies

Upgrading Social Protection Systems

Alleviating labor market disruptions. How can social protection systems deliver stable employment and
productivity growth (efficiency) while providing adequate protection to workers (equity) in a world with gen
AI? Social insurance, such as unemployment insurance (UI), can enhance individual and social welfare
by smoothing consumption in the presence of credit and insurance market failures. It enables the
unemployed to look for better jobs that match their skills, thereby improving the quality of job matches
(Marimon and Zilibotti 1999; Chetty 2008). Active labor market policies (ALMPs) complement UI and can
shorten periods of unemployment (unemployment “spells”) by improving workers’ skills (through retraining
programs) and reducing information gaps between job seekers and job providers. Social assistance
programs, such as cash transfers, provide financial support to low-income households during long
unemployment spells. This section looks at the role of UI, ALMPs, and social assistance programs in
alleviating adverse labor market outcomes in the past and assesses whether these are fit for purpose in
the future. Although gen AI has broader implications for high-skilled workers in cognitive tasks, lessons
from previous waves of automation offer valuable insight into how social protection systems can help
cushion the negative impact on labor markets.

Lessons from Past Automation Waves

Labor force displacement from automation. The impact of automation on labor markets depends on
whether the technology is substitutable for or complementary to various types of tasks performed by
workers. Mounting evidence indicates that automation in recent decades displaced workers in routine
tasks, pushing down average wages and intensifying job polarization. Increased use of industrial robots in
the United States hurt local labor markets. It drove down employment and wages, especially for manual
and routine jobs (Acemoglu and Restrepo 2020; Restrepo 2023), with displaced workers moving into
lower-paying occupations (Braxton and Taska 2023). Displacement of lower-skilled workers is also
observed in Europe (for example, Graetz and Michaels 2018; Acemoglu, Lelarge, and Restrepo 2020),
although workers adapted over time. Dauth and others (2021) find that lost manufacturing jobs in
Germany were replaced by new positions in the service sector, and young workers adjusted their
education choices, favoring college and university over vocational training.

Conceptual approach. Did social protection programs reduce the negative impact of automation on labor
market outcomes? To answer this question, the analysis that follows provides new evidence on how
social protection can mitigate the long-term effects of industrial robots on employment and wages at the
level of commuting zones in the United States. As in Acemoglu and Restrepo (2020), the underlying
intuition is that increased robotization in a commuting zone can reduce employment and wages relative to
other commuting zones. This reflects both the direct effects of robot adoption on employment and wages
and the spillover effects on the nontradables (service) sector resulting from the decline in local demand. 5
The novel analysis then exploits differences in the generosity of unemployment insurance and social

5
Specifically, exposure to robots is an adjusted Bartik-type measure that combines industry-level variation in the use of robots and
industry employment shares at the commuting zone level, adjusting for overall expansion of each industry’s output.

INTERNATIONAL MONETARY FUND 6

STAFF DISCUSSION NOTES Broadening the Gains from Generative AI: The Role of Fiscal Policies

assistance across states to examine whether these programs helped attenuate the adverse labor market
outcomes (see Brollo 2024 for details).
Figure 1. Effect of Robots on Employment and Wages in the
Cushioning effect of unemployment US Local Labor Market: The Role of Unemployment
Insurance (Log differences, 2000s)
insurance. The empirical results
reported in Figure 1 indicate that the
impact of robotization on employment
does not depend on UI generosity.
This is not surprising: UI benefits are
temporary and thus unlikely to
generate long-term effects (neither
positive income effects that boost local
labor demand nor negative effects on
labor supply because workers are
discouraged from job search). In
Source: Acemoglu and Restrepo (2020) and IMF staff calculations.
contrast, states with more generous UI Note: The figure presents estimates from instrumental variable regressions of
employment and wages on robot adoption using cross-sectional data on US
benefits saw a smaller decline in commuting zones. The dependent variables are change in employment to
wages as a result of robotization— population ratio or change in log of average hourly wage in a commuting
zone over the sample period 2000–07. Right side variables include (1) a
about two-thirds smaller than other measure of exposure to robots, which combines industry-level variation in the
use of robots and baseline employment shares at the commuting zone level,
states. This finding suggests that more adjusting for overall expansion of each industry’s output; (2) a dummy
generous UI allows displaced workers variable capturing high (low) unemployment insurance (UI) generosity, which
is equal to one if UI generosity is above (below) the median across US states;
to find jobs that better match their (3) the interaction between (1) and (2). Regressions also include controls for
commuting zone demographic characteristics, the share of employment in
skills, which contributes to more manufacturing, exposure to Chinese imports, and the share of employment in
efficient labor allocation. This effect is routine jobs. The bars show the effect of robot adoption for commuting zones
in states with high and low UI generosity, separately. The generosity of UI
particularly pronounced for workers benefits at the state level is measured as the product of the maximum legal
without a college degree (Figure 2), benefit amount and its duration. Whiskers indicate 95 percent confidence
intervals. Differences between high and low UI states are statistically
possibly because these workers rely significant at the 1 percent level. See Brollo (2024) for details.
relatively more on unemployment
Figure 2. Effect of Robots on Wages in the US Local Labor
insurance benefits when unemployed. Market: The Role of Unemployment Insurance by Levels of
These findings suggest that UI Education (Log differences, 2000s)
programs can effectively lessen the
adverse effects of industrial robots on
wages because they facilitate intensive
job search and allow more time for new
skills acquisition, potentially leading to
better job matching and increased
worker productivity.

Source: Acemoglu and Restrepo (2020) and IMF staff calculations.

Note: The figure presents estimates from instrumental variable regressions
of wages for workers with different educational attainment on robot adoption
using cross-sectional data on US commuting zones. See the Figure 1 note
for regression specification. Differences between high and low UI states for
workers with less than a college degree are statistically significant at the 1
percent level. See Brollo (2024) for details.

INTERNATIONAL MONETARY FUND 7

STAFF DISCUSSION NOTES Broadening the Gains from Generative AI: The Role of Fiscal Policies

Impact on poverty. In addition to its effects on Figure 3. Effect of Robots on Poverty: The Role of
labor markets, robotization may also contribute Social Assistance (Log differences, 2000s)
to increasing poverty, especially if the negative
impact of robotization is more pronounced for
workers at the bottom of the wage distribution.
These workers are at higher risk of falling into
poverty, because it is harder for them to find
new jobs with similar pay. Social assistance
programs can play a key attenuating role in this
regard. Overall, the analysis shows that
robotization resulted in a small long-term
increase in poverty: one additional robot per
thousand workers increased the poverty rate by
0.3 percentage point (3 percent increase).
Source: Acemoglu and Restrepo (2020) and IMF staff
However, most of the increase in poverty as a calculations.
result of robotization is attenuated in Note: The figure presents estimates from instrumental variable
regressions of the poverty rate on robot adoption. The dependent
commuting zones where social assistance is variables are the change in the poverty ratio in a commuting zone
relatively more generous (Figure 3). 6 over the sample period 2000–07. Right side variables are the
same as noted in Figure 1. Whiskers indicate 95 percent
confidence intervals. Differences between high and low SA states
Lessons from the past. Overall, these findings are statistically significant at the 1 percent level. See Brollo (2024)
for details. SA = social assistance.
suggest that the design of social protection
systems played a role in ameliorating adverse labor market and poverty impacts in the past. Although
robotization can lead to displacement of workers in routine and manual tasks, the impact of gen AI could
potentially be more widespread, replacing a broader spectrum of both routine and high-skill nonroutine
tasks. This calls for more fundamental changes in education and training systems and policy frameworks
to mitigate potential broader societal implications. The extent to which existing systems will need to be
upgraded in a world of rapid technological change and potentially more significant labor market
displacement is discussed in the next section.

Strengthening Social Spending during Rapid Technological Transitions

Costly transitions from disruptive technological advances. AI could generate significant long-term
productivity and growth dividends, but the transition may be very costly owing to labor market mismatches
and long periods of unemployment as a result of skill specificity. For instance, labor market adjustment
can be slower if gen AI benefits mainly production activities requiring specific skills that differ from those
used in the rest of the economy (Adão, Beraja, and Pandalai-Nayar 2024). 7 Workers may also face
barriers to mobility and go through long unemployment or retraining spells before finding a new job.
Indeed, technology-induced labor displacement often proceeds over a generation, with older workers

6
The generosity of social assistance at the state level is based on the generosity of Temporary Assistance for Needy Families
(TANF) benefits, the largest cash assistance program in the US. The generosity of TANF benefits in each state is measured by the
maximum monthly benefit for a family of three with no income in 1999, the year before the period covered in the analysis.
7 Adão, Beraja, and Pandalai-Nayar (2024) show that when skill specificity is stronger, as in the case of information and
communication technology (ICT), adjustment of labor markets is driven more by the gradual entry of younger generations than by
reallocation of older incumbent workers.

INTERNATIONAL MONETARY FUND 8

STAFF DISCUSSION NOTES Broadening the Gains from Generative AI: The Role of Fiscal Policies

leaving the workforce and fewer younger workers entering such jobs (Bürgisser 2023). The advent of AI
could aggravate adjustment costs if it entails broader substitution of nonroutine tasks (Acemoglu 2021)
and affects younger workers for whom early retirement is not an option. This section sheds light on
pertinent design features of UI, and how UI and ALMPs could be optimally combined to address potential
labor market disruptions from gen AI.

Model-based analysis and relevant channels. A model-based analysis is used to identify optimal
characteristics of social spending policy in response to disruptive technological advances (see Annex 1
for details). The analysis extends a tractable HANK-DGSE model with labor market frictions developed by
Ravn and Sterk (2021). The model features a potential for automation and two sectors of production,
allowing for a discussion of the asymmetric sectoral impacts (see also McKinsey 2023) and policies
supporting sectoral mobility. Each sector employs labor, traditional capital, and automated capital that can
substitute for labor, following the approach of Berg, Buffie, and Zanna (2018). Matching workers and firms
is costly (for example, as a result of search and matching frictions). Unemployed workers can search for
jobs in both sectors, but changing sectors means a period of unemployment, which can be lengthy if there
is skill mismatch. The model features UI support and ALMPs designed to facilitate sectoral mobility. The
expenses associated with these policies are funded with labor income taxes, ensuring budget neutrality
each period. Taxes on automated capital are considered in the next section.

Accounting for trade-offs and costly transitions. To understand gen AI’s potential impact on income,
productivity, and labor markets, the model simulates a sizable acceleration in the productivity of
automated capital in one sector. The baseline scenario is calibrated to a representative advanced
economy and assumes no policy changes. The increase in automation results in a gradual reduction in
labor demand in the sector, leading to a 20 percent decline in employment in the new steady state
(approached in about 15 years). 8 At the same time, the new steady state is characterized by an increase
in wages of about 15 percent, as automation increases aggregate productivity (Figure 4). 9 The long-term
effects of AI are subject to high uncertainty. For instance, Korinek (2023b) and Korinek and Suh (2024)
show that output could double in 15 years with a gradual advance in AI. Cazzaniga and others (2024) find
a smaller increase of 10–16 percent, and Acemoglu (2024) suggests even smaller gains, with an increase
in GDP of about 1–2 percent over the subsequent 10 years. 10 Irrespective of the magnitude of potential
long-term benefits, the transition entails short-term costs. Unemployment rises temporarily because of the
cost of relocating workers across sectors, which reduces the number of job vacancies posted by firms.
Furthermore, unemployment hurts the most vulnerable groups, with a substantial fall in consumption of

8
This simulation is based on a projection by McKinsey (2023) that automation could replace the time spent on work activities by 20
to 30 percent by 2030. Note however that the projection includes both the extensive and intensive margins, whereas the model
captures only the extensive margin.
9
The unemployment rate decreases in the new steady state, primarily because of the rise in labor demand in the sector that is not
affected by automation. This increase in demand more than compensates for the job losses in the sector that faces automation
shocks, as a result of complementarity between sectors in producing the final products. Over the long term, wages also increase,
which, in turn, increases unemployment benefits as a proportion of these wages. Consequently, this enhancement in support
improves the consumption levels of unemployed workers relative to their initial consumption figures.
10
Cazzaniga and others (2024) assume that the AI shocks would reduce the labor share by 5.5 percentage points based on the
historical change observed in the United Kingdom between 1980 and 2014. Annex Figure 1.1 shows that assuming a similar size of
the shock in our model leads to a similar increase in output.

INTERNATIONAL MONETARY FUND 9

STAFF DISCUSSION NOTES Broadening the Gains from Generative AI: The Role of Fiscal Policies

unemployed workers. 11 Incomplete insurance gives rise to a precautionary saving motive that further
propagates and amplifies aggregate shocks (Ravn and Sterk 2021). Such transition costs highlight the
need for policies to support affected workers and households.

Figure 4. Baseline Automation Shocks

Source: IMF staff calculations.

Note: The figure presents the baseline automation scenario without policies, with changes between the initial and final steady states
in panel 1 and deviations from the initial steady state over time in panel 2. The model is calibrated to a representative advanced
economy. See Annex 1 for details. ppts = percentage points.

Balancing UI benefit adequacy and work incentives. By providing income insurance for unemployed
workers, UI can help cushion the consumption loss and mitigate the adverse effects of higher
unemployment. This is particularly important in the case of unemployment driven by technological
transformation, when workers need more time to re-skill and search for jobs in other sectors, which
suggests a need for relatively longer duration of unemployment benefits. Two illustrative designs of UI are
considered: (1) a permanent increase in the UI replacement ratio by 1 percentage point; and (2) a
temporary asymmetric adjustment rule that increases the replacement ratio in proportion to the previous
quarter’s unemployment gap by a factor of 0.6 once the unemployment rate rises by more than 1
percentage point relative to its steady-state level (Figure 5). Both options effectively mitigate the drop in
consumption by unemployed workers compared with the baseline. Nevertheless, a permanent increase in
UI could discourage job searches and lead to an increase in the unemployment rate (Blanchard and
Tirole 2008), thereby lowering overall welfare. 12 In addition, higher unemployment and benefit generosity
can be fiscally costly, especially if AI destroys higher-wage jobs, requiring distortive tax hikes or public
debt accumulation that further weigh on economic activity. Conversely, scaling up unemployment
generosity during the transition could help manage fiscal costs and mitigate negative job search
incentives while still providing sufficient income support—along the lines of the evidence presented in the

11
In the model, the only source of income for unemployed workers is unemployment income support. The consumption fall would be
mitigated to the extent affected workers are wealthier and hold larger liquid savings to smooth consumption. Yet liquid wealth is very
low for many households in the United States (Challe and others 2017).
12A smaller permanent increase in the replacement ratio would mitigate the job search disincentive, but at the cost of smaller
consumption smoothing by the unemployed.

INTERNATIONAL MONETARY FUND 10

STAFF DISCUSSION NOTES Broadening the Gains from Generative AI: The Role of Fiscal Policies

previous section. Comparing the two schemes, the Figure 5. Different UI Programs
temporary UI adjustment aligned with unemployment
levels appears to yield the highest welfare benefits.
Overall, UI will need to be carefully designed to
facilitate mobility and adjustment while minimizing the
adverse effect on efficiency, but its efficacy will also
depend on how quickly advances in AI materialize
and associated transition costs. For instance, if
advances in AI accelerate even further in coming
years and all cognitive work can be performed by
machines over a relatively short period, disruption to
the workforce would be the most severe. Preparing
for such a scenario will require rethinking the design
of unemployment insurance programs; for example,
specifying benefits that depend on the duration of
unemployment spells and linking them better with
training and re-skilling programs.

Combining UI and ALMPs. ALMPs in the form of

training and skill development programs can enhance
the employability of workers and improve the quality
of matching between workers and employers,
particularly in response to asymmetric shocks
affecting specific skill sets. An adequate mix of Source: IMF staff calculations.
Note: The figure shows the baseline simulation and two
income and employment support can effectively unemployment income support programs: a permanent
increase in the replacement rate of 1 percentage point and
mitigate the surge in unemployment, consequently a temporary program aligned with the state of the
helping to limit the fall in wages (Figure 6). As a economy. These policies are funded with labor income
taxes, ensuring budget neutrality each period.
result, combining UI and ALMPs can help reduce

Figure 6. Transition Dynamics

Source: IMF staff calculations.

Note: The figure presents the baseline scenario and two policy options: temporary unemployment income support (UI) and a
combination of UI and active labor market policies (ALMPs). These policies are funded with labor income taxes, ensuring budget
neutrality each period.

INTERNATIONAL MONETARY FUND 11

STAFF DISCUSSION NOTES Broadening the Gains from Generative AI: The Role of Fiscal Policies

the transition costs while at the same time accelerating labor reallocation. As before, the policy mix and
its efficacy in ameliorating labor market impacts will depend on the pace of automation and size of
transition costs. For instance, the value of social protection could be smaller if gen AI displaces older,
skilled workers with larger savings. Similarly, skill specificities will imply different types of ALMPs than
have been considered in the past and could require broader access to effective training programs.

Distinct challenges facing emerging Figure 7. Illustrative Scenarios for Emerging Market and
market and developing economies. Developing Economies
While most studies on AI exposure focus
on advanced economies, some evidence
suggests that emerging market and
developing economies have a lower share
of high-skill occupations and, therefore,
are less exposed to AI (Pizzinelli and
others 2023). Nevertheless, emerging
market and developing economies, on
average, are also less prepared to adopt
AI (Cazzaniga and others 2024). For
instance, labor market policies and social
protection systems in many emerging
market and developing economies are
Source: IMF staff simulation.
more limited in scope and scale because Note: The figure presents the baseline scenario and two policy
of their large informal sectors, budgetary scenarios. Reflecting the narratives discussed in the text, the exposure
of emerging market and developing economies to automation shocks is
constraints, and less-developed calibrated smaller and more gradual than for the advanced economy
institutional capacity. Furthermore, these counterparts, but their policy space is more constrained. Both policy
scenarios include unemployment insurance and active labor market
economies have a larger share of young policies (ALMPs), but one features a larger response and faster
deployment (2.5 years after the initial shock), while the other is smaller,
people who are not in employment, with delayed implementation of ALMPs (five years after the initial shock).
education, or training (ILO 2022), which The values of the unemployment rate and consumption change are
averaged over 60 quarters. Bubble size is proportional to the amount of
raises concern about their ability to adjust time the unemployment rate is 1 percentage point above its initial steady
state.
to technological transitions (Adão, Beraja,
and Pandalai-Nayar 2024). At the same time, larger insurance and credit market imperfections and less
personal wealth available for consumption smoothing imply that the welfare gains from UI are potentially
greater than in advanced economies (Chetty and Looney 2006). 13 A model-based illustration (Figure 7)
suggests that countries with more fiscal space and capacity to effectively scale up income support and
ALMPs can significantly mitigate the impacts of the shock and more quickly realize the higher productivity
benefits.

13
For emerging market and developing economies with large informal sectors, providing income support and facilitating the
transition by training or retraining also help prevent workers from dropping out of the formal sector in response to automation
disruptions.

INTERNATIONAL MONETARY FUND 12

STAFF DISCUSSION NOTES Broadening the Gains from Generative AI: The Role of Fiscal Policies

Upgrading unemployment insurance

Figure 8. Maximum Benefit Duration
systems. The extent to which UI systems
cushion the negative effects of AI will likely
depend on the exposure to automation and
coverage, generosity, and design features.
Many advanced economies have established
generous unemployment benefit programs,
but the extent to which workers are eligible
for such programs varies. In addition, the
maximum benefit duration is usually less
than 12 months (Figure 8), and the support
declines over the unemployment spell. 14
Access to basic UI is not universal but
largely fragmented along occupational lines, Source: OECD (2023b). Sign “//” indicates a duration of more than 45
months.
and benefits are not portable. Minimum Note: Data labels in the figure use International Organization for
Standardization (ISO) country codes.
contribution durations typically exclude
temporary workers and labor market (re-) entrants, and maximum benefit receipt durations exclude the
long-term unemployed in most countries. Policymakers could consider extending the scope, portability,
and flexibility of unemployment insurance to allow sufficient time to allow for retraining. There is also a
growing debate over wage insurance that entails a temporary subsidy for displaced workers who find
reemployment at a lower wage, which would shorten the duration of unemployment and support workers
for whom job training may be less effective (Hyman and others 2021). 15

Enhancing social assistance. Social assistance, which contributes to equity in societies, could be
strengthened by adopting comprehensive programs that support workers directly or indirectly affected by
technological shifts, such as those facing long-term unemployment or reduced local labor demand as a
result of industry closures or automation. At one end of the spectrum are enhanced means-tested
guaranteed minimum income programs—which distribute cash or other assistance to households, with
benefits gradually declining as income rises—with other forms of support (for example, systematically
investing in training and job transition services, as discussed later). At the other end is providing
unconditional benefits to all, independent of income or employment status. The latter approach, by
design, would cover higher-income groups likely to be hurt by AI, potentially generating significant fiscal
costs (IMF 2017b). Such an approach may not be desirable at this point since the negative effects on
labor markets are not widespread and existing social safety nets offer more protection at lower cost in
most advanced economies. That said, the appropriate design, coverage, and eligibility of social
assistance programs must be carefully assessed in the face of potentially widespread disruptive
technological changes.

14 The maximum US benefit duration is on the low side of Organisation for Economic Co-operation and Development countries:
most states provide a maximum of 26 weeks.
15
As part of the broader Trade Adjustment Assistance in the United States, there is a wage insurance program in place for workers
ages 50 and older in particular industries that are severely affected by import competition (Frey 2019).

INTERNATIONAL MONETARY FUND 13

STAFF DISCUSSION NOTES Broadening the Gains from Generative AI: The Role of Fiscal Policies

Strengthening social protection in emerging market and developing economies. While almost half
of unemployed workers receive unemployment benefits in advanced economies, unemployment income
support programs in most emerging market and developing economies cover only a very limited number
of people, which reflects large informal sectors and limited administrative capacity. As a result, in most
emerging market and developing economies, income support to the unemployed is provided mostly
through social assistance programs (Brollo and others 2024a). Innovative approaches relying on digital
technologies can enable rapid expansion of coverage, but this can involve important trade-offs between
achieving high coverage and containing the associated fiscal cost (Brollo and others 2024b).

Upskilling and training workers. To facilitate transitions for those already in the labor force, some
countries have taken steps to promote preemptive acquisition of new skills (“lifelong learning”) while
overcoming credit constraints. For example, Singapore offers unconditional grants to all adults for training
throughout their working lives. Less stable employment relationships in the future also put a premium on
educational and training opportunities centered on workers rather than jobs. Finally, employer-provided
training may need replacement or substitution by other programs, with implications for ALMPs. Evidence
suggests that sector-based training programs in the United States that focused on workers in particular
industries (for example, manufacturing, health care, transportation, ICT) led to earnings gains of 14–38
percent in the year following training completion (Katz and others 2022), with persistent gains. Other
studies have found that upskilling can be more beneficial than on-the-job training programs in the case of
workers displaced by “offshoring” (Humlum, Munch, and Rasmussen 2023). These findings shed light on
the nature of policies needed in the face of job loss or downsizing as a result of AI and automation.
Looking ahead, the viability of re-skilling and retraining programs as opposed to alternatives such as early
retirement should be assessed to consider the challenges older workers might face in adapting to new
technologies.

Upgrading infrastructure for social assistance systems. To effectively provide sufficient coverage and
benefits, social assistance programs in both advanced and emerging market and developing economies
will require robust and universal information systems for beneficiary identification and verification. These
systems must be integrated across various social protection programs and must have efficient delivery
systems and strong institutional frameworks. To foster beneficiaries’ prospects of finding productive
employment, ALMPs must be integrated with social assistance programs for which, for example,
continued eligibility for benefits is conditional on participation in programs that offer job search support
and counseling services or skills training (IMF 2022).

Upgrading Tax Systems

AI and taxation. Whereas the previous section shows how social protection systems can cushion the
distributional and labor market implications of gen AI, this section first explores the extent to which current
tax systems already direct investment allocations toward or away from AI—for example, through capital
allowances and incentives. Building on recent literature that explores the role of taxation on robots and
other labor-saving and skill-biased technological change, it then analyzes the potential role of special
taxes on gen AI. Finally, the redistributive role of taxation in the context of gen AI—notably the taxation of
capital income—is discussed. This is especially important to the extent that gen AI exacerbates the

INTERNATIONAL MONETARY FUND 14

STAFF DISCUSSION NOTES Broadening the Gains from Generative AI: The Role of Fiscal Policies

concentration of wealth as a result of economic rents from rising market power. Moreover, if gen AI
reduces the labor income share and capital income is taxed less, requisite tax revenue will fall even as
countries must pay for upgrades to social protection systems.

Do Current Tax Systems Favor Labor-Displacing Investments?

Existing tax systems already differentiate between investment in broad asset categories. These
categories might include equipment (for example, machinery and computers), structures (for example,
offices), inventory, and intellectual property (for example, software, patents). Different tax treatments for
these asset categories result from deliberate policies, such as accelerated tax depreciation, investment
tax credits, and reduced tax rates for particular assets (for example, intellectual property). The incentive to
invest in a particular asset as provided by the tax system can be summarized by the marginal effective tax
rate (METR), which measures the extent to which taxation increases the pretax rate of return needed by
investors to break even (the cost of capital). A neutral tax system equalizes METRs across assets, while
variations in METRs reflect incentives and disincentives for private investment. For example, a higher
METR for machinery, relative to buildings, discourages investment in machinery while favoring that in
buildings. Investments in machinery and equipment (including tangible ICT) are often incentivized
because they are deemed to have a higher social return and stronger impact on economic growth. 16
Further, investments in research and development and intellectual property are often incentivized to
internalize positive externalities from innovation. However, differential taxation can also inadvertently
result in misallocation of capital assets and reduce productivity (IMF 2017a; Fatica 2017; Liu 2011).

Complementarity with labor. Variations in METRs can provide incentives for investments in labor-
displacing versus labor-augmenting technologies. Indeed, different asset categories vary in their
complementarity with labor. For example, investments in nonresidential structures enhance the
productivity of labor, whereas certain forms of purchased software and patents are likely to be labor
saving—although the degree of complementarity with labor may differ vastly within asset classes.
Evidence for advanced economies suggests the following:

• Machinery and equipment tend to be labor complements (Aum and Shin 2022; Jerbashian 2022).

• Computer hardware complements high-skilled labor but substitutes for low-skilled labor (Berman,
Bound, and Griliches 1994; Berndt and Morrison 1995; Autor, Katz, and Krueger 1998).

• Software is generally found to replace labor, with an estimated elasticity of substitution of 1.7 (Aum
and Shin 2022). Acquired intellectual property represents a broader asset category that could have
similarly high substitution elasticities; for example, in the case of firm-specific gen AI tools. However,
gen AI may encompass both labor-saving and labor-complementing assets, such as purchased
software for AI algorithm development and robust data infrastructure, high-performance computing
hardware, and acquired intellectual property, as well as investment in AI researchers and employee

16
For example, see Ohrn (2019), Zwick and Mahon (2017), and House and Shapiro (2008) for recent evidence in the United States;
Schaller (2006) for Canada; and Maffini, Xing, and Devereux (2019) and Bond and Xing (2015) for Europe, among others.

INTERNATIONAL MONETARY FUND 15

STAFF DISCUSSION NOTES Broadening the Gains from Generative AI: The Role of Fiscal Policies

training programs. Figure 9. Corporate Tax Biases for Labor-Saving Assets

(Percent)
Some tax systems currently favor
labor-displacing investments. Tax-
induced (dis)incentives for investment
in software, computer hardware, and
intellectual property can be measured
by their respective METRs in
comparison to that for buildings, for
example (a positive difference
indicating a higher tax and thus a
disincentive for these investments).
Figure 9 shows that the tax advantage
for intellectual property (that is,
acquired patents, utility models,
Sources: OECD 2023d; and IMF staff calculations.
trademarks) is almost universal across Note: The figure shows the extent of corporate tax biases for labor-saving assets,
countries, probably to capture positive measured as the METR for acquired software and computer hardware relative to the
METR for buildings in 2017–22; the METR for intellectual property is measured
spillovers from innovation. However, relative to the METR for buildings in 2017–20. A positive value denotes a higher tax
burden on the asset relative to buildings. METRs are based on the assumption that
most economies impose a higher tax the investment is financed by equity. AEs = advanced economies; EMDEs =
on acquired software and computer emerging market and developing economies; METR = marginal effective tax rate.
hardware than on
Figure 10. Corporate Tax Bias for Labor-Saving Assets by Economy: Top 10
buildings. Yet there
(Percent)
is significant
heterogeneity
across economies.
For instance, the
tax systems of
Germany, the
United States, The
Netherlands, New
Zealand,
Singapore, and
Hong Kong SAR
tend to favor
acquired software
and computer
Sources: OECD 2023d; ZEW 2021; and IMF staff calculations.
hardware (Figure Note: The figure shows the 10 economies with the largest corporate tax bias favoring labor-saving assets. The
10) and thus bias is measured as the METR for each asset type relative to the METR for buildings and is based on a
sample of 85 economies for 2022 for acquired software and computer hardware; for intellectual property the
encourage sample covers only the EU-27 for 2020. A negative value denotes a lower METR on the asset relative to
buildings. METR = marginal effective tax rate.
automation. This
can be undesirable, especially in a disruptive AI scenario. Most other countries, especially developing
economies, discourage investment in these labor-saving assets (see Annex Figure 2.2). This can also be

INTERNATIONAL MONETARY FUND 16

STAFF DISCUSSION NOTES Broadening the Gains from Generative AI: The Role of Fiscal Policies

distortive, especially in a scenario with more modest labor market implications, and could stymie AI
deployment.

Changes over time. Figure 11. Changing Corporate Tax Bias

Some countries have
scaled up tax
incentives to invest in
labor-saving assets
over time (Figure
11). For example, the
US Tax Cuts and
Jobs Act (TCJA)
became effective in
2018, allowing
companies to fully
expense capital
Source: IMF staff calculations based on ZEW 2021 and OECD 2023 data.
expenditures on Note: The figure plots the METR differentials for investments in intellectual property in 1998 against
acquired software those in 2020, such that the diagonal corresponds to the case where the tax treatment relative to
nonresidential structures has not changed over time. METR differentials for acquired software and
and computer computer hardware compare 2017 to 2022. Data labels in the figure use International Organization
for Standardization (ISO) country codes.
hardware. 17
Germany implemented a similar approach in 2021, and India reduced its statutory corporate income tax
(CIT) rate in 2020. which disproportionately increased the value of tax depreciation for acquired software
and computer hardware. Hungary, the Slovak Republic, and the Czech Republic introduced accelerated
tax depreciation for acquired intellectual property assets in the early 2000s. 18 For the median country, the
tax treatment of software and hardware became more generous by about 0.8 and 3 percentage points
(relative to buildings) over the years 2017–22. 19 These reforms may have incentivized automation.

Should AI Be Taxed?

Taxing automation. The implications of automation for labor markets and income inequality have
sparked debate over whether such investments should be taxed—for example, through a robot tax. This
subsection reviews these arguments and discusses the case for taxing AI.

Argument for not taxing AI: production efficiency. A natural starting point for the analysis of AI and
taxation is the principle that the tax system should not distort firms’ production decisions, thus maintaining
production efficiency. Diamond and Mirrlees (1971) show that even if nondistortionary lump-sum taxes
are unavailable (that is, in a “second-best” world), taxes should be such that production decisions remain
efficient. To satisfy this condition it is usually argued that a capital income tax should be neutral; that is, all

17
The full expensing will expire by 2026. Proponents of a long-term stable policy of full expensing have argued for this policy to be
made permanent.
18
These differ from incentives for in-house innovation through R&D tax credits or intellectual property (IP) regimes. Under such IP
regimes, or ‘patent boxes’, income from the exploitation of IP benefits from beneficial tax treatment such as, e.g., a lower rate than
the standard statutory tax rate.
19
Based on 85 countries covered in OECD (2023d) and excluding countries where there were no tax reforms.

INTERNATIONAL MONETARY FUND 17

STAFF DISCUSSION NOTES Broadening the Gains from Generative AI: The Role of Fiscal Policies

returns to capital should face the same METR. In other words, capital income taxes should not be
differentiated across different sectors or economic activities. The direct policy implication is that there
should be no special tax on gen AI, robots, or other forms of labor-replacing technology. Yet there can be
both efficiency and equity reasons to deviate from this principle. 20

Efficiency considerations: welfare losses from excessive job dislocation. One argument for taxing
automation is to mitigate excessive job displacement, which comes at a social cost when technological
change unfolds quickly while labor market adaptation is slow because of labor market frictions. Acemoglu,
Manera, and Restrepo (2020) develop a task-based model with labor market frictions. In this framework, a
robot tax can be desirable to discourage automation at the margin (that is, where automation yields the
smallest productivity gains). 21 Credit constraints provide further arguments for taxing automation to
internalize the external costs of job displacement. Beraja and Zorzi (2024) show that while this implies an
efficiency-based case for taxing automation during a transitional phase, this argument does not hold up in
the long run.

Equity considerations: mitigating wage inequality. Another argument for taxing automation is to
mitigate wage inequality arising from technological change. If governments do not have access to other
tax instruments for redistribution, distorting the adoption of new technologies to influence the wage
distribution may be optimal. The underlying idea is that reduced automation increases demand for low-
skilled labor and reduces demand for high-skilled workers, which then compresses relative wages—so-
called predistribution (Guerreiro, Rebelo, and Teles 2022; Costinot and Werning 2023; Thuemmel 2023).
Costinot and Werning (2023) estimate the optimal tax rate on robots at between 1 and 3.7 percent of the
price of the robots. The more disruptive the new technology, the higher the optimal tax rate. 22 Yet this
inequality-based argument may have limited salience in the face of gen AI. For instance, occupations that
are most affected by automation through robots are not filled by unskilled workers but rather by middle-
skilled routine jobs. In this case, a tax on automation will increase the relative wages for these middle-
skilled workers and thus decrease inequality at the top but raise inequality at the bottom. It can then be
optimal to either impose a tax or a subsidy on automation (Thuemmel 2023). The implications of a tax on
gen AI may have even more variable implications across the skill distribution, making the case for a tax or
subsidy ambiguous.

20
Another reason could be the carbon footprint from AI servers, which require vast amounts of electricity (de Vries 2023). A carbon
tax would be the most efficient way to internalize these external costs into the price of the technology. In its absence, however, a tax
on AI (or energy used by AI) provides a crude and admittedly less efficient alternative to doing so.
21
The automation tax is optimal if there are limits to changing general taxes on capital and labor. Without these restrictions, high
capital taxes relative to labor taxes are sufficient to efficiently mitigate labor substitution (Box 1 discusses the implications of general
taxes for labor and capital from developments in the labor income share).
22
Other studies look at tax reform, rather than optimal taxation, to illustrate the trade-off between distributional benefits and
efficiency costs of robot taxes (Berg and others 2021). Prettner and Strulik (2020) emphasize that robot taxes are not the best tool to
redistribute income because of their large cost to growth compared with other policies.

INTERNATIONAL MONETARY FUND 18

STAFF DISCUSSION NOTES Broadening the Gains from Generative AI: The Role of Fiscal Policies

Model-based illustration. Model Figure 12. Taxing Automation versus Taxing Labor Income to
simulations show that, if the transition Finance Unemployment Support
costs from labor displacement are
high (such as under a highly
disruptive AI scenario), a temporary
automation tax could improve welfare.
The macroeconomic and welfare
impacts of taxing automation are
analyzed with the same HANK-DGSE
model described previously. Figure 12
shows the implications of temporarily
taxing automation to finance
unemployment insurance benefits
during the transition (compared with Source: IMF staff calculations.
Note: The figure compares the effects of financing social policies with a
using labor taxes as a financing
temporary automation tax relative to doing so with a temporary labor income
source and at varying transition tax. Each bubble shows how shifting to an automation tax changes the
response of average wages and the unemployment rate (averaged over 60
costs). 23 An automation tax increases quarters). The different bubbles show how the results change with varying
the cost of using automated capital, transition costs, with larger bubbles corresponding to larger costs. Red
bubbles indicate that taxing automation implies a welfare loss relative to taxing
thereby discouraging firms from labor, while the opposite holds for green bubbles. ppts = percentage points.
substituting labor with capital. This
mitigates the surge in unemployment and yields a short-term welfare gain. However, it comes at the cost
of lower wages (averaging over the short-to-medium term) because of lower productivity. The welfare
gain comes from internalizing labor market and credit frictions (efficiency grounds) as well as from
redistributing income from capital owners to unemployed workers (equity grounds). We find that welfare is
likely to fall if transition costs are modest. However, if the transition cost is substantial, taxing automation
can improve welfare even if the government has no preference for redistribution (as shown in the largest
bubble in Figure 12). Hence, if the pace and depth of the disruptions from gen AI are larger than what we
have seen with past automation, the case for taxing gen AI is strengthened.

Practical considerations. Notwithstanding theoretical arguments, there are practical hurdles in

implementing taxes on automation, robots, and gen AI. Implementing such taxes requires that
governments identify the technologies (or the capital assets in which they are embodied) that risk
displacing labor. However, to codify such characteristics into actual policy is a tall order. Tax systems
usually differentiate capital assets by their useful lifespan and other characteristics, but it is not obvious
how these distinctions translate into different job tasks and whether they are substitutes for or
complements to the new technology. This makes it difficult to define the tax base of a specific technology
in practice. Differences in tax rates across assets that are similar can also incentivize relabeling of assets
to avoid taxes. From an international perspective, the location of AI assets can be highly mobile so that a

23
Insofar as a collective unemployment benefit system and ALMPs are needed to address labor market disruptions, laying off a
worker imposes a financial cost on society that firms do not internalize. This externality provides a case for “layoff” taxes (Blanchard
and Tirole 2008). Yet these taxes can also affect hiring rates and can reduce firm-level productivity, making them less desirable in
the long run (Autor, Kerr, and Kugler 2007).

INTERNATIONAL MONETARY FUND 19

STAFF DISCUSSION NOTES Broadening the Gains from Generative AI: The Role of Fiscal Policies

tax can be easily avoided by relocating or producing the AI abroad. A specific tax on gen AI is therefore
not recommended.

What can be done? As the previous section shows, higher METRs for labor-displacing assets because
of corporate tax incentives can in fact mimic an automation tax—which is the case in most countries.
Admittedly, such policies could be more distortionary in a scenario where labor disruptions are limited and
if applied to broad asset categories that include both labor-saving and labor-augmenting assets. Some
countries have corporate tax systems that mimic the opposite of an automation tax and give preferential
tax treatment to asset classes that are overall labor-displacing. These regimes could be reconsidered to
mitigate excessive labor displacement, which would be especially costly in a more disruptive labor market
scenario. Apart from taxes, governments can also include labor market considerations in AI regulation.
For instance, the EU AI Act approved by the European Parliament in March 2024 will require employers
to notify employees and workers’ representatives before implementing “high-risk AI systems.”

Broadening the Gains from AI through Taxation

AI makes capital income taxes more important. The Figure 13. Concentration of Capital Income
among Top Earners: Cross-Country
taxation of capital income is a controversial issue in
Evidence
public finance. Today’s predominant view is that
capital income should be taxed to serve both efficiency
and equity purposes. 24 AI reinforces these arguments
in light of both rising inequality and erosion of the
income tax base.

Mitigating rising inequality. Capital income is

considerably more concentrated at the top of the
income distribution compared with labor income
(Figure 13). Hence, a rising capital income share will
likely increase inequality. Past innovations have
indeed led to rising income and wealth inequality
(Aghion and others 2019). Moreover, these
innovations gave rise to significant (quasi) rents, which Sources: Luxembourg Income Survey; and IMF staff
calculations.
tend to be highly concentrated among a small group of Note: The figure shows the income share held by the top
top income earners. If the wider use of gen AI further 1 percent and the top 10 percent in labor and capital
income in European countries. For example, while the
amplifies these trends, effectively taxing capital income top 10 percent in labor income in European countries
to mitigate rising inequality will become more earned between 20 and 30 percent of the total, the top 10
percent in capital income earned between 60 and 95
important. percent of the total. The farther a country is to the left of
the 45-degree line, the more concentrated capital income
versus labor income.

24
A classic result in public finance is that the optimal tax on capital income is zero under certain conditions (Atkinson and Stiglitz
(1972, 1976); Chamley 1986; Judd 1985). The plausibility of these conditions has recently been challenged, however, and a positive
capital income tax is likely to be efficient under alternative, more realistic assumptions (see, for example, Banks and Diamond 2011;
Straub and Werning 2020).

INTERNATIONAL MONETARY FUND 20

STAFF DISCUSSION NOTES Broadening the Gains from Generative AI: The Role of Fiscal Policies

Mitigating base erosion. Lower effective taxation of capital relative to labor in many countries risks
eroding the income tax base. For instance, since it is difficult to distinguish between the source of income
earned by self-employed entrepreneurs, a lower tax on capital income relative to labor induces
entrepreneurs to label their income as capital income to minimize their tax burden (de Mooij and
Nicodème 2008; Devereux, Liu, and Loretz 2014). Such base erosion will be amplified when the labor
income share would drop further as a result of gen AI. This makes effective taxation of capital income at a
similar rate as labor income increasingly relevant to mitigate base erosion and sustain public revenue,
especially in light of the increasing needs to finance expenditures for upgraded social protection systems.

Declining tax burden on capital. The average tax burden on capital income—measured by corporate
and personal taxes on capital as a share of capital income—has consistently declined in advanced
economies since the 1980s (Figure 14). At the same time, the average tax burden on labor—measured
by personal taxes on labor and social security contributions as a ratio of labor income—has steadily
increased. Whereas capital and labor income were taxed at similar average rates in the early 1980s, the
gap had grown to almost 10 percentage points in 2018. However, there is significant cross-country
heterogeneity (Annex Figure 2.3). For instance, in Canada, Germany, Japan, the United Kingdom, and
the United States, the decline in the average capital tax was significant until 2018. In other countries,
such as France and Italy, the gap has also increased, but mainly as a result of the rise in the average tax
on labor; the average capital tax has been more stable. In emerging market and developing economies,
average tax rates on both labor and capital are generally much lower than in advanced economies.
Capital income is generally taxed more than labor income—reflecting the often-smaller coverage of the
personal income tax and the relative importance of the corporate income tax in these countries (see panel
2 in Figure 14).

Figure 14. Average Tax Rates on Labor and Capital Income, Five-Year Moving Average
(Percent)

Sources: Bachas and others 2022; and IMF staff calculations.

Note: ATRs are constructed by relating historical data on taxes paid to a measure of the tax base, based on global
macroeconomic data covering 1965–2018 (https://globaltaxation.world/). See Annex 2. ATR = average tax rate.

INTERNATIONAL MONETARY FUND 21

STAFF DISCUSSION NOTES Broadening the Gains from Generative AI: The Role of Fiscal Policies

Policy measures. To reverse the declining trend in capital income taxation, countries can consider
various measures to strengthen their systems, including the following: 25

• Strengthening the corporate income tax. The CIT is an effective withholding mechanism for the
taxation of capital income but has come under severe pressure from international profit shifting and
tax competition. The global minimum tax, as agreed on by the members of the Inclusive Framework,
will reduce pressures of tax competition for all countries and forestall a race to the bottom in CIT
(Hebous and Keen 2023). Thus, it might enable countries to reverse past reductions in effective tax
rates.

• Effective taxation of economic rents. Rising profits as a result of gen AI may call for a
supplemental tax on excess profits from monopoly rents (see, for example, Hebous, Prihardini, and
Vernon 2022). These do not need to be targeted to AI companies but could apply more generally. A
common approach between countries would mitigate the risk of cross-border profit shifting, which
could be significant in light of the intangible nature of gen AI assets.

• Improving enforcement. Capital income taxes have been under pressure from tax evasion,
especially to offshore low-tax countries or jurisdictions with strict secrecy standards. Recent global
developments facilitating exchange of information for tax purposes between countries, especially
automatic exchange of information (AEOI) under the initiative of the Group of Twenty and the US
Foreign Account Tax Compliance Act (FATCA), have already helped effectively counter offshore tax
evasion (EUTAX 2024). Gen AI can further enhance this enforcement through more effective use of
information to counter tax fraud. Such
improved enforcement will also enable Figure 15. Income Share by Source and Income
Groups in the United States
countries to design better systems of capital
income taxation, with higher effective rates.

• Enhancing capital gains taxation. Capital

gains are often taxed less than other forms
of capital income, primarily because they
are taxed only when realized, allowing for
deferral of tax liabilities. In some countries,
capital gains are partially or fully exempt
from personal income tax 26 or taxed less if
an asset is held for more than a specific
period. Moreover, certain types of returns
Source: US Internal Revenue Service.
(for example, on government bonds) and Note: The figure shows the sources of income among different
certain investors (for example, institutional income groups, from the overall average to the top percent of the
income distribution.
investors) receive preferential treatment. At

25
See Hebous and others (2024) for a comprehensive discussion on how to effectively tax capital income and/or wealth, including
through taxes on wealth and wealth transfers.
26
In the US, assets transferred at death receive a step up to their market value, implying that gains up to that point become
effectively exempt.

INTERNATIONAL MONETARY FUND 22

STAFF DISCUSSION NOTES Broadening the Gains from Generative AI: The Role of Fiscal Policies

the same time, capital gains are highly concentrated among top income earners (Figure 15).
Enhancing the taxation of capital gains thus presents an opportunity to mitigate inequality while
bolstering revenue to support additional redistributive measures.

Should labor be taxed less? Whereas across-the-board tax cuts on labor might be too expensive,
targeted income tax credits for workers or job credits for employers could alleviate the relative tax burden
on certain types of labor compared with capital. Income tax credits—such as the Earned Income Tax
Credit (EITC) in the United States—have been effectively targeted to low-income earners. To be effective
in reducing job displacement, they should at least partially reduce labor costs for employers—which tends
to be the case in the United States (Nichols and Rothstein 2015). Targeting specific sectors prone to
excessive job displacement could be more challenging given the potentially pervasive and variable impact
of the new technologies on labor. Moreover, practical obstacles may arise in relation to employment
credits if firms could relabel new positions to gain eligibility, as was recently observed in the United States
(Gurmu, Sjoquist, and Wheeler 2021; Chirinko and Wilson 2023).

Funding for AI Innovation and Deployment

Fiscal policies as drivers of AI innovation. Current advances in AI are the fruit of decades-long
investment in fundamental research, often benefiting from publicly funded programs (for example, AI Next
and AI Institutes in the United States and the European Union’s partnership on AI, data, and robotics),
subsidies, and research and development grants to firms. As discussed earlier, tax policies have also
encouraged investment in AI to different degrees across countries. Should fiscal policies continue to be
used to finance AI innovation and deployment?

Funding for AI innovation. Global corporate investment in AI has soared more than tenfold in the past
decade. After decades of public and private research, AI technology has matured to the commercial
adoption phase, suggesting that fiscal support for overall innovation in AI is now less of a priority in
advanced economies. Governments should instead focus on areas where social returns are greater than
private returns and lead to insufficient private investment. These include funding fundamental research
with broader applications, providing the necessary infrastructure (for example, digital connectivity,
electricity grids), particularly in emerging market and developing economies, and promoting AI
applications in the public sector (education, health, government administration), where productivity has
faltered and consumer costs have surged in past decades. Favoring applications that expand, rather than
substitute for, human capabilities (Acemoglu, Autor, and Johnson 2023) is desirable in theory but may not
be feasible with existing levels of administrative capacity, even in advanced economies. Instead, the
focus can be on funding innovative upskilling and re-skilling programs and strengthening social protection
systems more broadly.

Considerable scope for enhancing AI deployment in emerging market and developing economies.
From high-quality education and learning, through precisely targeted and individually customized human
capital investments, to improved access to financial services, AI could be deployed to improve
development outcomes in emerging market and developing economies. AI could also help provide cost-
effective solutions to deliver social services to those who need them most, including remote communities.

INTERNATIONAL MONETARY FUND 23

STAFF DISCUSSION NOTES Broadening the Gains from Generative AI: The Role of Fiscal Policies

Other opportunities include risk management—disease prevention, natural disaster management, and
humanitarian crisis management—which tends to be weaker in emerging market and developing
economies. In addition to weaknesses in the digital infrastructure, critical constraints on adoption of AI
solutions include lack of a developed digital economy and a supporting entrepreneurial ecosystem and
scarce local AI expertise—highlighting a role for governments to close gaps.

Need to upgrade administrative and governance capacity. Effectively directing AI innovation can be a
tall order (IMF 2024), not least because AI technology is undergoing rapid change and applications of
specific technologies are hard to predict. Governments must upgrade their capabilities, investing in
expertise to be able to select and vet funding to projects and update regulation as the speed, autonomy,
and opacity of AI systems challenge traditional models of regulation. National investment in talent, data,
and computer resources, as well as in national procurement capacity, will complement these upgrades. A
dedicated agency, inspired by the model of the US National Institutes of Health, that mobilizes the private
sector, academia, and other stakeholders could track AI developments and use. Such an institution could
help develop a broader accountability framework for companies that build, deploy, and control AI, as well
as for downstream users. The centrality of data governance suggests that AI governance cannot be
divorced from the governance of data and the promotion of data commons. Given the global reach of AI
technologies and high cross-border spillovers, policy initiatives should seek international collaboration. To
this end, the European Organization for Nuclear Research (CERN), which operates the largest particle
physics laboratory in the world, and similar international scientific collaborations may offer useful lessons
(United Nations 2023). A “distributed CERN,” reimagined for AI, could expand opportunities for
international cooperation on innovation, use, and regulation.

Conclusions and Policy Implications

Forward-thinking policies. Many questions remain about what gen AI will mean for labor markets and
economies at large in the coming decades. Will lower-skilled workers gain on higher-skilled workers?
What kinds of new jobs will emerge? Will opportunity and wage inequality increase or decrease? Given
the tremendous uncertainty about the speed of AI progress and its economic impact in the short, medium,
and long term, policymakers must be prepared for alternative scenarios (Korinek and Suh 2024). If AI
supports workers by raising labor productivity and creating new job opportunities, policy responses will
differ markedly from a scenario in which AI rapidly displaces jobs, reduces wages, and increases
inequality. And not all countries are likely to be affected equally. While the answers are not yet certain,
countries will need to assess whether their social protection, education, and tax systems are fit for
purpose and flexible enough to cope with a wide range of potential scenarios.

The future of social protection. AI-induced labor market transformations have the potential to redefine
employment and reshape the skills demanded by employers, which will require a comprehensive
reassessment of labor policies and social protection mechanisms. If AI acts like general-purpose
technologies (for example, steam engines, electricity), then a range of work tasks will be automated but
new opportunities will arise as well. In this case, current social protection systems provide a solid
foundation, particularly in advanced economies. Unemployment insurance eligibility rules should be
robust and resilient to radical uncertainty and coverage should be broadened to encompass self-

INTERNATIONAL MONETARY FUND 24

STAFF DISCUSSION NOTES Broadening the Gains from Generative AI: The Role of Fiscal Policies

employed workers and those with atypical employment contracts. ALMPs should aim to improve skill
acquisition for those capable of adapting to new market requirements, while social assistance benefits
should target those permanently displaced or indirectly affected by labor market disruptions. Sector-
based training, apprenticeships, and upskilling and re-skilling programs could play a greater role in
helping workers move to new tasks and sectors. A future scenario in which the nature of work changes
dramatically (for example, if tasks became increasingly unnecessary) would also require gains to be
shared more widely, using unconditional transfers, which suggest a need to consider the design and
infrastructure required for such policies. AI itself could be leveraged to radically improve the efficiency and
quality of social protection systems in conjunction with traditional systems to reduce data privacy risks.

Educational systems. Education and training policies should be geared to upskilling workers to cope
with structural changes in the workplace and to matching the skill and task demands of new technologies
(OECD 2023c). This is also essential to ensure societies can harness the full potential of AI. While
spending matters, the quality and adaptability of education will make the difference in preparing workers
for change. Given the high uncertainty about which skills are needed at any point in time, educational
systems need to be flexible in responding to market demands, keeping equity and access in mind.
Educational systems themselves could take advantage of gen AI to foster higher-level skills such as
critical thinking, analysis, and strategy. But developments in gen AI and robotics will require that people
develop skills to work alongside AI systems and not just existing technologies.

Taxing AI. A special tax on gen AI to reduce its speed of adoption and prevent excessive labor
displacement will be hard to design and implement and would run the risk of hampering productivity
growth, including in areas where AI investment augments labor. Yet it is recommended that countries
reconsider the design of current corporate tax systems in how they incentivize investments in automation.
For instance, tax incentives in the form of capital allowances may need to be reconsidered in countries
where they are more generously applied to labor-displacing software or intangibles than to other assets.
At the same time, countries where corporate tax systems impose much higher tax burdens on AI may
hold up deployment and reduce productivity growth. Income tax credits and job credits could also be
considered to mitigate excessive labor displacement from automation, even if they cannot be targeted to
particular occupations. Finally, given the large amount of energy consumed by AI servers, taxing the
associated carbon emissions is a good way to reflect the external environmental costs in the price of the
technology.

Capital taxation. The average tax on capital income has declined in advanced economies during the
past few decades. It is important that countries reverse this trend, especially under a disruptive AI
scenario. First, low taxation of capital compared with labor can contribute to excessive labor displacement
and exacerbate labor market frictions. Second, capital income taxes are essential to address the
increasing inequality associated with rising market power and economic rents enjoyed by dominant firms
in winner-take-all markets. Third, large labor displacements that reduce the labor income share will erode
the tax base and reduce public revenue. Enhancing capital income taxes will not only prevent this but will
boost revenue mobilization—which is necessary to finance higher education and social spending with the
arrival of automation. More effective taxation of capital income requires restoration of the corporate
income tax and calls for well-designed excess profit taxes, higher personal income taxes on capital

INTERNATIONAL MONETARY FUND 25

STAFF DISCUSSION NOTES Broadening the Gains from Generative AI: The Role of Fiscal Policies

through better enforcement of automatic information exchange between countries, and enhanced taxation
of capital gains.

Advancing tax systems through gen AI. Digital transformations in revenue administration have already
visibly reduced tax evasion around the world (Amaglobeli and others 2023). Gen AI has significant
potential to further advance tax administration practices to improve tax enforcement, given the critical role
of information and data. The AI-associated information revolution can ultimately enable tax system
redesign. Information constraints are at the heart of the theory of “second-best”; by transforming
information systems and management, gen AI will turn classic tax theory upside down and urge a rethink
of the old ways of doing things. It may, for instance, usher in the design of a personalized progressive
value-added tax, an income tax based on lifetime income, or a real-time market-value-based property tax.

AI that serves people. Fiscal policies can promote innovation and deployment of AI in applications with
greater social benefits, such as those that improve the quality of social services (education, health care,
government). Finding the right policy response in a highly uncertain world of AI will require upgrading
administrative and analytical capacity to monitor and evaluate trends in technological advances. Policies
to steer and cushion the implications of AI will depend on how future scenarios unfold. Ensuring that AI is
deployed for the common good, and that its benefits are distributed equitably, will require governmental
and intergovernmental action, with innovative ways to motivate participation by the private sector,
academia, and civil society. Innovation and deployment policies should thus work in tandem with social
protection, education, and tax and regulatory policies to broaden the gains of gen AI for all.

INTERNATIONAL MONETARY FUND 26

STAFF DISCUSSION NOTES Broadening the Gains from Generative AI: The Role of Fiscal Policies

Box 1. Taxation and the Decline of the Labor Income Share

The labor income share has fallen steadily since the 1980s in most advanced economies. Some have
argued that declining corporate income tax (CIT) rates explain part of this development. There are
several channels through which the CIT can impact the labor income share. For instance, a lower CIT
rate can boost investment so that the income share of labor decreases. This holds especially for
investment in labor-substituting capital, which reduces employment and wages (Kaymak and Schott
2023; Acemoglu and others 2020). Yet capital and labor can also be complementary, so that more
capital leads to higher labor productivity and higher wages, thus offsetting the decrease in the labor
income share (Fuest, Peichl, and Siegloch 2018; Garrett, Ohrn, and Suárez Serrato 2020). A lower CIT
rate can also induce self-employed entrepreneurs to report their income as profits instead of wages so
that the capital share artificially increases (de Mooij and Nicodème 2008; Devereux, Liu, and Loretz
2014). On balance, the impact of the CIT on the labor income share is an empirical question.

A panel regression of 42 advanced Box Figure 1.1. Declining Labor Share: Role of Taxation
and emerging market economies (Percentage points)
shows that, conditioning on other
macroeconomic determinants, the
CIT rate has a positive and
significant impact on the labor
income share (Annex 3). For each
percentage point reduction in the
statutory CIT rate, the labor share
falls 0.1 percent. In comparison, for
each percentage point reduction in
the top statutory PIT rate, the labor
share increases 0.11 percent. Thus,
the reduction in the average
statutory CIT rate in the past two
decades (from 27.7 to 23.9 percent
during 2005–18), combined with a
slight increase in the average top Sources: Bachas and others 2022; IMF, World Economic Outlook database;
PIT rate (from 41.9 to 43.7 percent), and IMF staff calculations.
Note: The figure shows the cumulative changes in the corporate and
is estimated to have reduced the
personal income tax rates, respectively, for 32 advanced economies during
labor share by 0.58 percentage point 2005–18 and their estimated impact on the labor share. CIT = corporate
in advanced economies over this income tax; PIT = personal income tax.
period.

INTERNATIONAL MONETARY FUND 27

STAFF DISCUSSION NOTES Broadening the Gains from Generative AI: The Role of Fiscal Policies

Annex 1. Model Framework

The analysis extends a tractable HANK-DGSE model with labor market frictions developed by Ravn and
Sterk (2021) to incorporate automation. The economy has two intermediate sectors that employ labor,
traditional capital, and automated capital. Automated capital differs from traditional capital in that it can
substitute for labor, following Berg, Buffie, and Zanna (2018). The two sectors are initially identical, but we
assume that only one is subject to an automation shock in order to capture cross-sector labor flows and
study policies supporting sectoral mobility. The output of the two intermediate sectors is combined to
produce a final good.

The labor market is subject to search and matching frictions following the Diamond-Mortensen-Pissarides
tradition. Three types of households are modeled: firm owners, employed workers, and unemployed
workers. Unemployed workers can search for jobs in either sector, but sectoral reallocation can be
costly—for instance, because of a potential mismatch between the skills of the unemployed workers and
those required by firms (Branch, Petrosky-Nadeau, and Rocheteau 2016; Di Pace and Hertweck 2019;
Walsh 2011). Because of these labor market frictions, job prospects are uncertain, exposing households
to idiosyncratic income risk. Employed workers are not fully insured against this risk and, therefore,
attempt to self-insure through precautionary saving.

Policy tools modeled include unemployment insurance and active labor market policies designed to
facilitate sectoral mobility. These policies are funded with labor income taxes, ensuring budget neutrality
each period. Other standard features include nominal rigidities and a Taylor-rule-based interest rate.

The following equations describe the production structure and labor market dynamics.

Sectoral production: Intermediate sector 𝑖𝑖 = {1,2} good is produced by combining traditional capital 𝑘𝑘𝑡𝑡𝑖𝑖 ,
automated capital 𝑚𝑚𝑡𝑡𝑖𝑖 , and labor 𝑛𝑛𝑡𝑡𝑖𝑖 with a constant elasticity of substitution (CES) function with elasticity
𝜂𝜂𝑖𝑖 and traditional capital share 𝛼𝛼𝑖𝑖 :
1/η𝑖𝑖 (𝜂𝜂𝑖𝑖 – 1)/𝜂𝜂𝑖𝑖 (𝜂𝜂𝑖𝑖 − 1)/𝜂𝜂𝑖𝑖 𝜂𝜂𝑖𝑖 /(𝜂𝜂𝑖𝑖 – 1)
𝑦𝑦𝑡𝑡𝑖𝑖 = �𝛼𝛼𝑖𝑖 �𝑘𝑘𝑡𝑡𝑖𝑖 – 1 � + (1 – 𝛼𝛼𝑖𝑖 )1/η𝑖𝑖 �𝑣𝑣𝑡𝑡𝑖𝑖 � � ,
in which 𝑣𝑣𝑡𝑡𝑖𝑖 represents the CES bundle between labor and automated capital, with elasticity 𝜂𝜂𝑖𝑖𝑖𝑖 and labor
share 𝑠𝑠𝑖𝑖 :
1/𝜂𝜂𝑖𝑖𝑖𝑖 (𝜂𝜂𝑖𝑖𝑖𝑖 – 1)/𝜂𝜂𝑖𝑖𝑖𝑖 (𝜂𝜂𝑖𝑖𝑖𝑖 − 1)/𝜂𝜂𝑖𝑖𝑖𝑖 𝜂𝜂𝑖𝑖𝑖𝑖 /(𝜂𝜂𝑖𝑖𝑖𝑖 – 1)
𝑣𝑣𝑡𝑡𝑖𝑖 = �𝑠𝑠𝑖𝑖 �𝑛𝑛𝑡𝑡𝑖𝑖 � + (1 – 𝑠𝑠𝑖𝑖 )1/𝜂𝜂𝑖𝑖𝑖𝑖 �𝐴𝐴𝑖𝑖𝑖𝑖 𝑖𝑖
𝑡𝑡 𝑚𝑚𝑡𝑡−1 � � ,
and 𝐴𝐴𝑖𝑖𝑖𝑖
𝑡𝑡 denotes the productivity of automated capital.

Final goods are produced using inputs from the two intermediate sectors with a Cobb-Douglas production
function 27:
𝑦𝑦𝑡𝑡 = 𝐴𝐴𝑡𝑡 (𝑦𝑦𝑡𝑡1 )𝜍𝜍 (𝑦𝑦𝑡𝑡2 )1−𝜍𝜍 ,
in which 𝐴𝐴𝑡𝑡 denotes aggregate total factor productivity.

27
Using a CES function with equal shares of the two sectors and an elasticity of 0.9 results in similar findings.

INTERNATIONAL MONETARY FUND 28

STAFF DISCUSSION NOTES Broadening the Gains from Generative AI: The Role of Fiscal Policies

Labor market dynamics: The value of being employed in sector 𝑖𝑖, 𝐸𝐸𝑡𝑡𝑖𝑖 , is a function of the Nash
bargained wage 𝑤𝑤𝑡𝑡𝑖𝑖 subject to the labor income tax rate 𝜏𝜏𝑡𝑡 , the beginning-of-period match destruction rate
𝜌𝜌, the probability of finding a job in the same quarter of job separation in each sector, and the transition
probability from employment to unemployment 𝑧𝑧𝑡𝑡 = 𝜌𝜌(1 − 𝑓𝑓𝑡𝑡 )28:
𝑗𝑗 𝑗𝑗
𝐸𝐸𝑡𝑡𝑖𝑖 = (1 − 𝜏𝜏𝑡𝑡 )𝑤𝑤𝑡𝑡𝑖𝑖 + 𝛽𝛽((1 − 𝜌𝜌)𝐸𝐸𝑡𝑡+1
𝑖𝑖 𝑖𝑖
+ 𝜌𝜌 (𝑓𝑓𝑡𝑡+1 𝑖𝑖
𝐸𝐸𝑡𝑡+1 𝑖𝑖
+ 𝑓𝑓𝑡𝑡+1 𝐸𝐸𝑡𝑡+1 ) + 𝑧𝑧𝑡𝑡+1 𝑈𝑈𝑡𝑡+1 ).

The value of being unemployed after having worked in sector i, 𝑈𝑈𝑡𝑡𝑖𝑖 , is a function of the unemployment
benefit 𝑢𝑢𝑢𝑢𝑡𝑡𝑖𝑖 and the probability of finding a job next period in different sectors.
𝑗𝑗 𝑗𝑗
𝑈𝑈𝑡𝑡𝑖𝑖 = 𝑢𝑢𝑢𝑢𝑡𝑡𝑖𝑖 + 𝛽𝛽((1 − 𝑓𝑓𝑡𝑡 )𝑈𝑈𝑡𝑡+1
𝑖𝑖 𝑖𝑖
+ 𝑓𝑓𝑡𝑡+1 𝑖𝑖
𝐸𝐸𝑡𝑡+1 + 𝑓𝑓𝑡𝑡+1 𝐸𝐸𝑡𝑡+1 ).
The unemployment benefit 𝑢𝑢𝑢𝑢𝑡𝑡𝑖𝑖 is modeled as
𝑢𝑢𝑢𝑢𝑡𝑡𝑖𝑖 = 𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑡𝑡 𝑤𝑤𝑡𝑡−1
𝑖𝑖
� ,
+ 𝑟𝑟𝑟𝑟 𝑤𝑤
in which 𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑡𝑡 is a policy parameter capturing the replacement ratio of the previous wage’ i.e., the larger
𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑡𝑡 the higher the unemployment generosity. The government can adjust the replacement ratio over
time, as discussed below. Meanwhile, the second term in the equation is a constant reflecting other
benefits of being unemployed (such as home production).

In each sector, there is a representative “labor firm” that sells labor services to intermediate goods firms at
a price 𝑥𝑥𝑡𝑡𝑖𝑖 . The labor firms hire workers by posting vacancies. Posting a vacancy involves a fixed cost, 𝜅𝜅 >
0, which is paid in every period the vacancy is open. To model frictions in sectoral mobility, the labor firm
is assumed to have a cost of hiring new workers 𝑡𝑡𝑡𝑡𝑡𝑡𝑖𝑖 , for instance due to retraining needs. Active labor
market policy 𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝑡𝑡 is assumed to reduce this hiring cost. The value of an open vacancy is
𝐽𝐽𝑡𝑡𝑖𝑖 = 𝑥𝑥𝑡𝑡𝑖𝑖 − 𝑤𝑤𝑡𝑡𝑖𝑖 − 𝑡𝑡𝑡𝑡𝑡𝑡𝑖𝑖 (1 − 𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝑡𝑡 ) + 𝛽𝛽(1 − 𝜌𝜌) 𝐽𝐽𝑡𝑡+1
𝑖𝑖
,
in which 𝑡𝑡𝑡𝑡𝑡𝑡𝑖𝑖 is assumed to follow 𝑡𝑡𝑡𝑡𝑡𝑡𝑖𝑖 = 𝜌𝜌𝑡𝑡𝑡𝑡 𝑡𝑡𝑡𝑡𝑡𝑡−1
𝑖𝑖
+ (1 − 𝜌𝜌𝑡𝑡𝑡𝑡 )𝛾𝛾 𝑤𝑤𝑡𝑡𝑁𝑁𝑁𝑁 ε𝑖𝑖𝑡𝑡𝑡𝑡𝑡𝑡 , with 𝜌𝜌𝑡𝑡𝑡𝑡 reflecting the persistence of
the cost, 𝛾𝛾 relating the cost to the wage, and ε𝑖𝑖𝑡𝑡𝑡𝑡𝑡𝑡 = {0,1} to capture periods during which the cost is
incurred.
Note that the value of having an open vacancy is based on the “free-entry condition”: the labor firm posts
vacancies if the value of having a vacancy exceeds the cost. Therefore, the net value of opening a
𝜅𝜅
vacancy is driven to zero in equilibrium, or: 𝐽𝐽𝑡𝑡𝑖𝑖 = , in which 𝑓𝑓𝑓𝑓𝑡𝑡𝑖𝑖 is the probability of filling a vacancy in
𝑓𝑓𝑓𝑓𝑡𝑡𝑖𝑖

sector i.

Calibration
Annex Table 1.1 presents the calibration of the key parameters of the model. Specifically, the separation
rate 𝜌𝜌, matching efficiency 𝑚𝑚
�, and matching elasticity 𝜇𝜇 are calibrated to match three targets: (1)
unemployment rate of 7 percent, which ranges between 5 percent in the US and 10 percent in euro area
countries; (2) job-finding probability of workers of 45 percent per quarter; and(3) vacancy-filling probability
of firms of 70 percent per quarter (Christoffel, Kuester, and Linzert 2009; Challe 2020). The replacement
ratio is set at 50 percent. The benefits of unemployment beyond unemployment insurance are calibrated
to match a 20–30 percent consumption loss upon unemployment (Den Haan, Rendahl, and Riegler
2018). The calibration of other parameters is explained in Annex Table 1.1.

28
This reflects the probability of being separated at the beginning of the period (𝜌𝜌) and not finding a job in the same period (1 − 𝑓𝑓𝑡𝑡 ).

INTERNATIONAL MONETARY FUND 29

STAFF DISCUSSION NOTES Broadening the Gains from Generative AI: The Role of Fiscal Policies

This baseline simulation calibrates a series of shocks that increase the productivity of automated capital
by 300 percent by 2030, resulting in a decrease of labor in the affected sector by 20 percent. This is
consistent with McKinsey (2023)’s projection that automation could replace the time spent on work
activities by 20 to 30 percent by 2030. Note however that the projection includes both the extensive and
intensive margins, whereas the model captures only the extensive margin.

Annex Table 1.1. Calibration of Model Parameters

Parameters Symbol Value Sources
Sectoral Production Parameters
Elasticity of substitution in labor- 𝜂𝜂𝑖𝑖𝑖𝑖 2.5 Berg, Buffie, and Zanna (2018)
automated capital (LAC) bundle
Elasticity between traditional capital and 𝜂𝜂𝑖𝑖 0.50 Berg, Buffie, and Zanna (2018)
the LAC in sectoral production
Share of labor in LAC bundle 𝑠𝑠𝑖𝑖 0.99 Labor income share of 61 percent

Share of traditional capital in sectoral 𝛼𝛼𝑖𝑖 0.54 Traditional capital income share of
production 35 percent
Share of sectoral input in final goods 𝜍𝜍 0.5 Assumption of symmetric sectors
production
Labor Market Parameters

Separation rate 𝜌𝜌 0.06 See text for discussion

Matching efficiency 𝑚𝑚
� 0.56 See text for discussion
Matching elasticity 𝜇𝜇 0.5 See text for discussion
Vacancy cost 𝜅𝜅 0.04 About 15 percent of average wage
as in standard value
Worker’s wage bargaining power 𝜓𝜓 0.65 Standard value
Replacement rate initial value 𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟 0.5 See text for discussion
Other benefit of unemployed worker 𝑟𝑟𝑟𝑟 0.26 See text for discussion
Response of unemployment income 𝜆𝜆 0.4 See discussion in scenario
support to unemployment rate
Persistence of hiring cost 𝜌𝜌𝑡𝑡𝑡𝑡 0.5 See discussion in scenario
Hiring cost parameter 𝛾𝛾 0.15 See discussion in scenario
Macroeconomic Parameters
Discount factor 𝛽𝛽 0.989 Challe (2020)
Firm owners’ risk aversion 𝜎𝜎� 0.283 Challe (2020)
Depreciation rate 𝛿𝛿 0.025 Standard value
Price elasticity 𝜖𝜖𝑝𝑝 6 Standard value
Price adjustment cost 𝜅𝜅 30 Match the average frequency of
price changes (every three
quarters)
Taylor rule: smoothing 𝜌𝜌 0.5 Standard value

Taylor rule: response to inflation 𝛾𝛾𝜋𝜋 2 Standard value

Taylor rule: response to output gap 𝛾𝛾𝑦𝑦 0.125 Standard value

Source: IMF staff compilation.

Note: Standard values are based on Smets and Wouters (2007); Christiano, Eichenbaum, and Evans (2005); Gomes, Jacquinot,
and Pisani (2012); and Gertler, Sala, and Trigari (2008).

INTERNATIONAL MONETARY FUND 30

STAFF DISCUSSION NOTES Broadening the Gains from Generative AI: The Role of Fiscal Policies

Smaller Impact Scenario

The alternative scenario considers a smaller increase in the productivity of automated capital—about half
the magnitude of that in the baseline presented in Figure 4—similar to Cazzaniga and others (2024).
Under this assumption, output increases by about 17 percent, while the labor income share in the
affected sector falls by 10 percentage points, reducing total labor income share in the economy by 5
percentage points (Annex Figure 1.1). The same qualitative results remain. The transition is still costly
due to sectoral mobility frictions, but less severe than in the baseline (Annex Figure 1.2).

Annex Figure 1.1. Alternative Scenario with Smaller Sectoral Automation Shocks

Source: IMF staff calculations.

Note: In the alternative scenario, sectoral productivity shocks to automated capital are about half the magnitude of those in the
baseline scenario.

Annex Figure 1.2. Transitional Dynamics in

the Alternative Scenario

Source: IMF staff calculations.

Note: In the alternative scenario, sectoral productivity
shocks to automated capital are about half those of the
baseline scenarios.

INTERNATIONAL MONETARY FUND 31

STAFF DISCUSSION NOTES Broadening the Gains from Generative AI: The Role of Fiscal Policies

Annex 2. Effective Tax Rates

This annex discusses details of two tax measures used in the analysis: marginal effective tax rates
obtained from simulations of countries’ prevailing tax rules and average tax rates, obtained from
macroeconomic data.

Marginal Effective Tax Rates

The effective tax burden of the corporate income tax (CIT) depends not only on the statutory CIT rate but
also on detailed depreciation rules for various asset types and other capital allowances. Many countries
aim to match tax depreciation to economic depreciation. Yet accelerated (or bonus) depreciation for tax
purposes has been used extensively to support particular investments that are deemed to have positive
spillovers. For instance, the US and the UK allow investments in some tangible assets to be expensed
immediately. Such provisions can have significant implications for investment because of the time value
of money.

Incentives for incremental investment decisions by firms can be measured by the marginal effective tax
rate (METR). It captures firms’ investment incentives at the point at which the marginal product of capital
equals the cost of capital—that is, at the breakeven point. This margin determines the scale of
investment. METRs can be calculated for investments in alternative assets and different sources of
finance, using parameters in the tax code for the tax rate, tax depreciation, deductions for financing costs,
and other capital allowances or tax credits. If METRs are zero, this means that the tax system is neutral
regarding investment; if METRs are positive but equivalent across asset types, the tax system distorts
overall investment but not the allocation between assets. By exploring the variation in METRs for different
asset types, one can gauge the tax preference for certain asset types over others.

Two internationally comparable datasets on METRs published by the Organisation for Economic Co-
operation and Development (OECD) and the ZEW Leibniz Centre for European Economic Research are
used in this note. Both contain METRs for various countries, asset groups, and time periods and are
based on the same methodological approach. The ZEW data cover three main assets (buildings,
machinery, acquired patents) for 34 countries over the period 1998–2020. The OECD data cover eight
different assets (including acquired software and computer hardware; see Annex Figure 2.1) for 74
countries over the period 2017–22. Annex Figure 2.1 shows that current METRs (based on equity finance
and averaged across 74 countries) are far from neutral across asset types. The averages are well above
zero, with the bulk of observations in the range of 12.5 to 35 percent.

INTERNATIONAL MONETARY FUND 32

STAFF DISCUSSION NOTES Broadening the Gains from Generative AI: The Role of Fiscal Policies

Annex Figure 2.1. METRs by Asset Group, 2017–22, 74 Countries

(Percent)

Sources: Organisation for Economic Co-operation and Development; and IMF staff calculations.
Note: METR = marginal effective tax rate.

Annex Figure 2.2. Corporate Tax Bias for Labor-Saving Assets by Economy: Bottom 10
(Percent)

Sources: Organisation for Economic Co-operation and Development; ZEW Leibniz Centre for European Economic Research;
and IMF staff calculations.
Note: The figure shows the 10 economies with the least corporate tax bias favoring labor-saving assets. The bias is measured as
the METR for each asset type relative to the METR for buildings and is based on a sample of 85 economies for 2022 for acquired
software and computer hardware; for intellectual property the sample covers only the EU27 for 2020. A negative (positive) value
denotes a lower (higher) METR on the asset relative to buildings. METR = marginal effective tax rate.

INTERNATIONAL MONETARY FUND 33

STAFF DISCUSSION NOTES Broadening the Gains from Generative AI: The Role of Fiscal Policies

Average Tax Rates

Average tax rates (ATRs) are constructed by relating historical data on taxes paid to a measure of the tax
base. The ATRs used in this note are based on macroeconomic data compiled by Bachas and others
(2022), who develop a long global time series of ATRs on capital and labor. Based on this approach,
each tax category, as captured in the macroeconomic data, is attributed to one of the two production
factors.

The tax revenue data are structured according to the methodology in OECD Revenue Statistics (OECD
2023e), with category 1200 (CIT) fully attributed to capital, category 2000 (social security contributions
[SSC]) fully attributed to labor, and category 1100 (personal income tax [PIT]) attributed partially to capital
and partially to labor. The PIT share attributed to capital and labor varies by country and year. Although
the definition of the ATR on labor income follows Bachas and others (2022), we apply a different definition
of the ATR on capital income by excluding property and wealth taxes to better reflect taxes affecting firms’
automation decisions. Consumption taxes are excluded from the analysis. Having attributed each tax to
one of the two production factors, totals are divided by the capital and labor income shares computed
from the national accounts data.

Further disaggregating the ATRs into their components shows that the ATR on corporate income and on
capital income (at the personal level) have both decreased since 1980, with the latter experiencing
significant reduction in the early 1980s and in the early 2000s. The CIT component is more affected by
business cycles and has been on a downward trajectory since the late 1970s. Since 2016, automatic
exchange of bank information may have played a role in a slight recovery of the personal tax
components. Annex Figure 2.3 shows trends in the ATRs for labor and capital in individual countries.

Annex Figure 2.3. ATR on Labor and Capital for Individual Countries

Sources: Organisation for Economic Co-operation and Development; and IMF staff calculations.

INTERNATIONAL MONETARY FUND 34

STAFF DISCUSSION NOTES Broadening the Gains from Generative AI: The Role of Fiscal Policies

Annex 3. Corporate Taxes and Labor Income

Share
This annex presents details on the empirical analysis, examining the effect of corporate taxes on labor
income share across countries.

Data
Labor income share. A long time series of labor income shares for 155 countries since 1965 is available
from Bachas and others (2022). It uses a panel of national accounts data from the System of National
Accounts (SNA), produced by the United Nations. Estimation of labor income shares requires information
on all the components of national income, including compensation of employees plus a share of mixed
income (operating surplus of private unincorporated enterprises). While it is relatively straightforward to
measure workers’ compensation, the labor share of mixed income (unincorporated enterprises) is hard to
measure. The benchmark series of the labor share of mixed income is assumed at 75 percent; that is, 25
percent of mixed income is considered capital income. Given that the labor share of mixed income is
time- and country-invariant, the identifying variation in the labor income share comes from changes in the
workers’ compensation across countries and years.

Country-level tax rates and macroeconomic variables. Data on statutory corporate and personal
income tax rates come from IMF Fiscal Affairs Department Tax Rate Database. Data on nontax
macroeconomic determinants of labor income, including average hourly wage, average price of capital
relative to consumption, trade openness, population, financial development index, and inflation are
obtained from the IMF World Economic Outlook database.

Specification
The baseline specification evaluates the impact of corporate and personal income taxes on labor income
share, controlling for the standard determinants of labor income share for 42 advanced and emerging
market economies during 1990–2008:
𝐿𝐿𝐿𝐿𝑖𝑖𝑖𝑖 = 𝛽𝛽1 𝐶𝐶𝐶𝐶𝐶𝐶𝑖𝑖𝑖𝑖 + 𝛽𝛽2 𝑃𝑃𝑃𝑃𝑃𝑃𝑖𝑖𝑖𝑖 + 𝛬𝛬′ 𝑋𝑋𝑖𝑖𝑖𝑖 + 𝛼𝛼𝑖𝑖 + 𝜂𝜂𝑡𝑡 + 𝜀𝜀𝑖𝑖𝑖𝑖 ,

in which 𝐿𝐿𝐿𝐿𝑖𝑖𝑖𝑖 corresponds to the labor income share in country i year t; 𝑋𝑋𝑖𝑖𝑖𝑖 is a vector of structural and
institutional characteristics, including labor costs, defined as the log of average labor income per hour
worked (US dollars, constant prices); capital costs, defined as the average price of capital relative to
consumption; trade openness, defined as the sum of a recipient country’s exports and imports as a share
of GDP; and level of financial development. 𝐶𝐶𝐶𝐶𝐶𝐶𝑖𝑖𝑖𝑖 denotes the statutory corporate income tax rate, and
𝑃𝑃𝑃𝑃𝑃𝑃𝑖𝑖𝑖𝑖 denotes the top statutory personal income tax. The regression also includes country and year fixed
effects, 𝛼𝛼𝑖𝑖 and 𝜂𝜂𝑡𝑡 , respectively. 𝜀𝜀𝑖𝑖𝑖𝑖 denotes the error term.

Results
The results are summarized in Annex Table 3.1, with results for only advanced economies in specification
(3), discussed in Box 1. Estimated coefficients for the CIT rate are positive and insignificant if all countries
are included in specifications (1) and (2). This masks a positive and significant (at the 90 percent level)
CIT coefficient for advanced economies in specification (3) and a negative and insignificant CIT
coefficient for emerging market and developing economies in (4). In contrast, the estimated coefficients
for the PIT rates are negative and significant in specifications (2)–(4). Based on the results from
INTERNATIONAL MONETARY FUND 35
STAFF DISCUSSION NOTES Broadening the Gains from Generative AI: The Role of Fiscal Policies

specification (3), a 1 percentage point reduction in the statutory CIT rate, all else equal, reduces the labor
income share in advanced economies by 0.1 percentage point. A 1 percentage point reduction in the top
statutory PIT rate, in comparison, increases the labor income share by 0.11 percentage point. The
average statutory CIT rate decreased from 27.7 to 23.9 percent in the sample of 32 advanced countries
during 2005–18, or by 3.8 percentage points. The average top statutory PIT rate increased from 41.9 to
43.7 percent during the same period, or by 1.8 percentage points. Together, these would imply a
reduction in labor share of 0.58 percent, all else constant (Box Figure 1.1).

Annex Table 3.1. Estimation Results

(1) (2) (3) (4)

Statutory CIT rate 0.01 0.02 0.10* -0.01

(0.02) (0.02) (0.06) (0.03)

-0.04** -0.11*** -0.06**

Statutory PIT rate
(0.02) (0.04) (0.03)

Controls included Y Y Y Y
Year FE Y Y Y Y
Country FE Y Y Y Y
Observations 1086 1086 399 687
R squared 0.89 0.89 0.93 0.91

Sources: Bachas and others 2022; IMF, World Economic Outlook database; and IMF staff calculations,
Note: IMF staff calculations are based on an ordinary least squares panel regression with country and year fixed effects,
controlling for macroeconomic factors and the CIT and PIT rate. The sample covers 42 advanced and emerging market
economies during 1990–2018. Values in parentheses indicate standard errors. Column 1 includes only the statutory CIT rate for
all countries in the sample. Column 2 adds the top PIT rate to the full sample. Columns 3 and 4 estimate the effects of tax rates
for advanced and emerging market and developing economies, respectively. CIT = corporate income tax; FE = fixed effects; PIT
= personal income tax; Y = yes.
*p < .10; **p < .05; ***p < .01.

INTERNATIONAL MONETARY FUND 36

STAFF DISCUSSION NOTES Broadening the Gains from Generative AI: The Role of Fiscal Policies

References
Acemoglu, Daron. 2021. “Harms of AI.” NBER Working Paper 29247, National Bureau of Economic
Research, Cambridge, MA

Acemoglu, Daron. 2024. “The Simple Macroeconomics of AI.” NBER Working Paper 32487, National
Bureau of Economic Research, Cambridge, MA.

Acemoglu, Daron, David Autor and Simon Johnson. 2023. “Policy Insight 123: Can We Have Pro-Worker
AI? Choosing a Path of Machines in Service of Minds”, CEPR Policy Insight No 123, CEPR Press,
Paris & London.

Acemoglu, Daron, David H. Autor, and Simon Johnson. 2023. “Can We Have Pro-worker AI?: Choosing a
Path of Machines in Service of Minds.” CEPR Policy Insight 123, Centre for Economic Policy
Research, London.

Acemoglu, Daron, and Simon Johnson. 2023. Power and Progress: Our 1000-Year Struggle over
Technology and Prosperity. New York: Public Affairs.

Acemoglu, Daron, Andrea Manera, and Pascual Restrepo. 2020. “Does the U.S. Tax Code Favor
Automation?” Brookings Papers on Economic Activity (Spring): 231–300.

Acemoglu, Daron, and Pascual Restrepo. 2018. “The Race between Man and Machine: Implications of
Technology for Growth, Factor Shares, and Employment.” American Economic Review 108 (6):
1488–542.

Acemoglu, Daron, and Pascual Restrepo. 2020. “Robots and Jobs: Evidence from US Labor
Markets.” Journal of Political Economy 128 (6): 2188–244.

Acemoglu, Daron, Claire Lelarge, and Pascual Restrepo. 2020. "Competing with Robots: Firm-Level
Evidence from France." AEA Papers and Proceedings, 110: 383-88.

Adão, Rodrigo, Martin Beraja, and Nitya Pandalai-Nayar. 2024. “Fast and Slow Technological
Transitions.” https://economics.mit.edu/sites/default/files/inline-
files/transitions_JPE_Macro_revised_final.pdf.

Aghion, Philippe, Ufuk Akcigit, Antonin Bergeaud, Richard Blundell, and David Hemous. 2019.
“Innovation and Top Income Inequality.” Review of Economic Studies 86 (1): 1–45.

Aghion, Philippe, Celine Antonin, Simon Bunel, and Xavier Jaravel. 2022. “The Effects of Automation on
Labor Demand: A Survey of the Recent Literature.” CEPR Discussion Paper DP16868, Centre for
Economic Policy Research, London. https://ssrn.com/abstract=4026751.

Amaglobeli, David, Ruud A. de Mooij, Andualem Mengistu, Mariano Moszoro, Manabu Nose, Soheib
Nunhuck, Sailendra Pattanayak, Lorena Rivero del Paso, Frankosiligi Solomon, Rebecca Sparkman,
Hervé Tourpe, Gerardo Uña. 2023. Transforming Public Finance Through GovTech, IMF Staff
Discussion Note 2023/4, International Monetary Fund, Washington, DC.

INTERNATIONAL MONETARY FUND 37

STAFF DISCUSSION NOTES Broadening the Gains from Generative AI: The Role of Fiscal Policies

Atkinson, Anthony, and Joshep E. Stiglitz. 1972. “The Structure of Indirect Taxation and Economic
Efficiency.” Journal of Public Economics 1 (1): 97–119.

Atkinson, Anthony, and Joshep E. Stiglitz. 1976. “The Design of Tax Structure: Direct versus Indirect Taxation,”
Journal of Public Economics, 6 (1-2), 55-75.

Aum, Sangmin, and Yongseok Shin. 2022. “Is Software Eating the World?” Federal Reserve Bank of San
Francisco Working Paper 2022/10/21. https://www.frbsf.org/wp-content/uploads/sites/4/10-21-2022-
Shin-paper.pdf.

Autor, D. 2022. “The Labor Market Impacts of Technological Change: From Unbridled Enthusiasm to
Qualified Optimism to Vast Uncertainty.” NBER Working Paper 30074, National Bureau of Economic
Research, Cambridge, MA

Autor, David. 2024. “Applying AI to Rebuild Middle Class Jobs.” NBER Working Paper 32140, National
Bureau of Economic Research, Cambridge, MA.

Autor, David H., Lawrence F. Katz, and Alan B. Krueger. 1998. Computing Inequality: Have Computers
Changed the Labor Market? Quarterly Journal of Economics 113 (4): 1169–213.

Autor, David H., William R. Kerr, and Adriana D. Kugler. 2007. “Does Employment Protection Reduce
Productivity? Evidence from US States.” Economic Journal 117 (521): 189–217.

Bachas, Pierre, Matthew H. Fisher-Post, Anders Jensen, and Gabriel Zucman. 2022. “Capital Taxation,
Development, and Globalization: Evidence from a Macro-Historical Database.” NBER Working Paper
29819, National Bureau of Economic Research, Cambridge, MA.

Baily, Martin Neil, Erik Brynjolfsson, and Anton Korinek. 2023. “Machines of Mind: The Case for an AI-
Powered Productivity Boom.” Brookings Institution, Washington, DC.

Baldwin, Richard. The Globotics Upheaval: Globalization, Robotics, and the Future of Work. 2019.
Oxford, UK: Oxford University Press.

Banks, James, and Peter Diamond. 2011. “The Base for Direct Taxation.” In Dimensions of Tax Design,
edited by Stuart Adam, Tim Besley, Richard Blundell, Stephen Bond, Robert Chote, Malcolm
Gammie, Paul Johnson, Gareth Myles, and James Poterba, 548–674. Oxford, UK: Oxford University
Press.

Beraja, Martin, and Nathan Zorzi. 2024 “Inefficient Automation.” The Review of Economic Studies.

Berg, Andrew, Lahcen Bounader, Nikolay Gueorguiev, Hiroaki Miyamoto, Kenji Moriyama, Ryota
Nakatani, and Luis-Felipe Zanna. 2021. “For the Benefit of All: Fiscal Policies and Equity-Efficiency
Trade-offs in the Age of Automation.” IMF Working Paper 2021/187, International Monetary Fund,
Washington, DC.

Berg, Andrew, Edward F. Buffie, and Luis-Felipe Zanna. 2018. Should We Fear the Robot Revolution?
(The correct answer is yes). Journal of Monetary Economics 97: 117–48.

Berman, Eli, John Bound, and Zvi Griliches. 1994. “Changes in the Demand for Skilled Labor within U.S.
Manufacturing: Evidence from the Annual Survey of Manufacturers.” Quarterly Journal of Economics 109
(2): 367–97.

INTERNATIONAL MONETARY FUND 38

STAFF DISCUSSION NOTES Broadening the Gains from Generative AI: The Role of Fiscal Policies

Berndt, Ernst R., and Catherine J. Morrison. 1995. “High-Tech Capital Formation and Economic
Performance in U.S. Manufacturing Industries; An Exploratory Analysis,” Journal of Econometrics 65
(1): 9-43.

Blanchard, Olivier J., and Jean Tirole. 2008. “The Joint Design of Unemployment Insurance and
Employment Protection: A First Pass.” Journal of the European Economic Association 6 (1): 45–77.

Bond, Stephen, and Jing Xing. 2015. “Corporate Taxation and Capital Accumulation: Evidence from
Sectoral Panel Data for 14 OECD Countries.” Journal of Public Economics 130: 15–31.

Branch, William A., Nicolas Petrosky-Nadeau, and Guillaume Rocheteau. 2016. “Financial Frictions, the
Housing Market, and Unemployment.” Journal of Economic Theory 164: 101–35.

Braxton, J. Carter, and Bledi Taska. 2023. “Technological Change and the Consequences of Job Loss.”
American Economic Review 113 (2): 279–316.

Brollo, Fernanda. 2024. “Strengthening Social Protection to Pave the Way for Technological Innovation:
Evidence from the U.S.” IMF Working Paper 2024/095, International Monetary Fund, Washington,
DC.

Brollo Fernanda, David Coady, Samir Jahan, and Riki Matsumoto. 2024a. “Challenges Facing SSNs in
Emerging and Developing Economies.” IMF Working Paper 2024/096, International Monetary Fund,
Washington, DC.

Brollo Fernanda, Gabriel Lara Ibarra, and Ricardo Campante Vale. 2024b. “Strengthening Income
Stabilization through Social Protection in Emerging and Developing Economies: The Brazilian
Experience.” IMF Working Paper 2024/052, International Monetary Fund, Washington, DC.

Brussevich, Mariya, Era Dabla-Norris, and Salma Khalid. Forthcoming. “Is Technology Widening the
Gender Gap? Automation and the Future of Female Employment.” International Monetary Fund,
Washington, DC.

Brynjolfsson, Erik, Danielle Li, and Lindsey R. Raymond. 2023. “Generative AI at Work.” NBER Working
Paper 31161, National Bureau of Economic Research, Cambridge, MA.

Bürgisser, Reto. 2023. “Policy Responses to Technological Change in the Workplace.” JRC Working
Papers Series on Social Classes in the Digital Age, Joint Research Centre, Brussels.

Cazzaniga, Mauro, Florence Jaumotte, Longji Li, Giovanni Melina, Augustus J. Panton, Carlo Pizzinelli,
Emma J. Rockall, and Marina Mendes Tavares. 2024. “Gen-AI: Artificial Intelligence and the Future of
Work.” IMF Staff Discussion Note SDN2024/001, International Monetary Fund, Washington, DC.

Challe, Edouard. 2020. “Uninsured Unemployment Risk and Optimal Monetary Policy in a Zero-Liquidity
Economy.” American Economic Journal: Macroeconomics 12 (2): 241–83.

Challe, Edouard, Julien Matheron, Xavier Ragot, and Juan F. Rubio-Ramirez. 2017. “Precautionary
Saving and Aggregate Demand.” Quantitative Economics 8 (2): 435–78.

Chamley, Christophe. 1986. “Optimal Taxation of Capital Income in General Equilibrium with Infinite
Lives.” Econometrica 54 (3): 607–22.

Chetty, Raj. 2008. “Moral Hazard versus Liquidity and Optimal Unemployment Insurance.” Journal of
Political Economy 116 (2).

INTERNATIONAL MONETARY FUND 39

STAFF DISCUSSION NOTES Broadening the Gains from Generative AI: The Role of Fiscal Policies

Chetty, Raj, and Adam Looney. 2006. “Consumption Smoothing and the Welfare Consequences of Social
Insurance in Developing Economies.” Journal of Public Economics 90 (12): 2351–56.

Chirinko, Robert S., and Daniel J. Wilson. 2023. “Job Creation Tax Credits, Fiscal Foresight, and Job
Growth: Evidence from US States.” National Tax Journal 76 (3): 481–523.

Christiano, Lawrence J., Martin Eichenbaum, and Charles L. Evans. 2005. “Nominal Rigidities and the
Dynamic Effects of a Shock to Monetary Policy.” Journal of Political Economy 113 (1): 1–45.

Christoffel, Kai, Keith Kuester, and Tobias Linzert. 2009. “The Role of Labor Markets for Euro Area
Monetary Policy.” European Economic Review 53 (8): 908–36.

Comunale, Mariarosaria, and Andrea Manera. 2024. "The Economic Impacts and the Regulation of AI: A
Review of the Academic Literature and Policy Actions." IMF Working Paper 2024/065, International
Monetary Fund, Washington, DC.

Costinot, Arnaud, and Ivan Werning. 2023. “Robots, Trade, and Luddism: A Sufficient Statistic Approach
to Optimal Technology Regulation.” Review of Economic Studies 9 (5): 2261–91.

Dauth Wolfgang, Sebastian Findeisen, Jens Suedekum, and Nicole Woessner. 2021. “The Adjustment of
Labor Markets to Robots.” Journal of the European Economic Association 19 (6): 3104–153.

de Mooij, Raud A., and Gaëtan Nicodème. 2008. “Corporate Tax Policy and Incorporation in the
EU.” International Tax and Public Finance 15: 478–98.

de Vries, Alex. 2023. “The Growing Energy Footprint of Artificial Intelligence.” Joule 7 (10): 2191–94.

Den Haan, Wouter, Pontus Rendahl, and Markus Riegler. 2018. “Unemployment (Fears) and Deflationary
Spirals.” Journal of the European Economic Association 16 (5):1281–349.

Devereux, Michael P., Li Liu, and Simon Loretz. 2014. “The Elasticity of Corporate Taxable Income: New
Evidence from UK Tax Records.” American Economic Journal: Economic Policy 6 (2): 19–-53.

Di Pace, Federico, and Matthias S. Hertweck. 2019. “Labor Market Frictions, Monetary Policy and
Durable Goods.” Review of Economic Dynamics 32: 274–304.

Diamond, Peter A., and James A. Mirrlees. 1971. “Optimal Taxation and Public Production I: Production
Efficiency.” American Economic Review 61 (1): 8–27.

Eloundou, Tyna, Sam Manning, Pamela Mishkin, and Daniel Rock. 2023. “GPTs are GPTs: An Early Look
at the Labor Market Impact Potential of Large Language Models.”
https://doi.org/10.48550/arXiv.2303.10130.

EUTAX Observatory (EUTAX). 2024. “Global Tax Evasion Report 2024.” Paris.

Fatica, Serena. 2017. “Measurement and Allocation of Capital Inputs with Taxes: A Sensitivity Analysis
for OECD Countries.” Review of Income and Wealth 63 (1): 1–29.

Felten, Edward W., Manav Raj, and Robert Seamans. 2023. “Occupational Heterogeneity in Exposure to
Generative AI.” https://dx.doi.org/10.2139/ssrn.4414065.

Frey, Carl Benedikt (2019). The Technology Trap: Capital, Labor, and Power in the Age of Automation.
Princeton: Princeton University Press.

INTERNATIONAL MONETARY FUND 40

STAFF DISCUSSION NOTES Broadening the Gains from Generative AI: The Role of Fiscal Policies

Fuest, Clemens, Andreas Peichl, and Sebastian Siegloch. 2018. "Do Higher Corporate Taxes Reduce
Wages? Micro Evidence from Germany." American Economic Review 108 (2): 393–418.

Garrett, Daniel G., Eric Ohrn, and Juan Carlos Suárez Serrato. 2020. “Tax Policy and Local Labor Market
Behavior.” American Economic Review: Insights 2 (1): 83–100.

Gertler, Mark, Luca Sala, and Antonella Trigari. 2008. “An Estimated Monetary DSGE Model with
Unemployment and Staggered Nominal Wage Bargaining.” Journal of Money, Credit and Banking 40
(8):1713–64.

Gomes, Sandra, Pascal Jacquinot, and Massimiliano Pisani. 2012. “The EAGLE: A Model for Policy
Analysis of Macroeconomic Interdependence in the Euro Area.” Economic Modelling 29 (5): 1686–
714.

Graetz, Georg, and Guy Michaels. 2018 “Robots at Work.” Review of Economics and Statistics 100 (5):
753–68.

Guerreiro, Joao, Sergio Rebelo, and Pedro Teles. 2022. “Should Robots Be Taxed?” Review of Economic
Studies 89 (1): 279–311.

Gurmu, Shiferaw, David L. Sjoquist, and Laura Wheeler. 2021. “The Effectiveness of Job Creation Tax
Credits.” Regional Science and Urban Economics 90 (C).

Hebous, Shafik, and Michael Keen. 2023. “Pareto-Improving Minimum Corporate Taxation.” Journal of
Public Economics 225: 104952.

Hebous, Shafik, Alexander D. Klemm, Geerten Michielse, and Carolina Osorio Buitron. 2024. “How to Tax
Wealth.” IMF How-to Note 2024/001, International Monetary Fund, Washington, DC.

Hebous, Shafik, Dinar Prihardini, and Nate Vernon. 2022. "Excess Profit Taxes: Historical Perspective
and Contemporary Relevance." IMF Working Paper 2022/187, International Monetary Fund,
Washington, DC.

House, Christopher L., and Matthew D. Shapiro. 2008. “Temporary Investment Tax Incentives: Theory
with Evidence from Bonus Depreciation.” American Economic Review 98 (3): 737–68.

Humlum, Anders, Jakob Munch, and Mette Rasmussen. 2023. “What Works for the Unemployed?
Evidence from Quasi-Random Caseworker Assignments.” University of Chicago Becker Friedman
Institute for Economics Working Paper 2023-43, Chicago, IL.

Hyman, Benjamin G., Brian K. Kovak, Adam Leive, and Theodore Naff. 2021. “Wage Insurance and
Labor Market Trajectories.” AEA Papers and Proceedings 111: 491–95.

International Labour Organization (ILO). 2022. “Youth Not in Employment, Education or Training in Asia
and the Pacific: Trends and Policy Considerations.” Geneva.

International Monetary Fund (IMF). 2017a. Fiscal Monitor: Achieving More with Less. Washington, DC,
April.

International Monetary Fund (IMF). 2017b. Fiscal Monitor: Tackling Inequality. Washington, DC, October.

International Monetary Fund (IMF). 2022. “IMF Engagement on Social Safety Net Issues for Surveillance
and Program Work.” IMF Technical Notes and Manuals 2022/000, Washington, DC.

INTERNATIONAL MONETARY FUND 41

STAFF DISCUSSION NOTES Broadening the Gains from Generative AI: The Role of Fiscal Policies

International Monetary Fund (IMF). 2024. “Expanding Frontiers: Fiscal Policies for Innovation and
Technology Diffusion.” Fiscal Monitor, Chapter 2, Washington, DC, April.

Jerbashian, Vahagn. 2022. “On the Elasticity of Substitution between Labor and ICT and IP Capital and
Traditional Capital.” CESifo Working Paper 9989, CESifo GmbH, Munich.

Judd, Kenneth L. 1985. “Redistributive Taxation in a Simple Perfect Foresight Model. Journal of Public
Economics 28 (1): 59–83.

Katz, Lawrence F., Jonathan Roth, Richard Hendra, and Kelsey Schaberg. 2022. “Why Do Sectoral
Employment Programs Work? Lessons from WorkAdvance.” Journal of Labor Economics 40 (S1):
S249–S291.

Kaymak, Barış, and Immo Schott. 2023. “Corporate Tax Cuts and the Decline of the Manufacturing Labor
Share.” International Finance Discussion Paper 1379, Board of Governors of the Federal Reserve
System, Washington, DC.

Kehrig, Matthias, and Nicolas Vincent. 2021. “The Micro-Level Anatomy of the Labor Share Decline.”
Quarterly Journal of Economics 136 (2): 1031–87.

Korinek, Anton. 2023a. “Generative AI for Economic Research: Use Cases and Implications for
Economists.” Journal of Economic Literature 61 (4):1281–317.

Korinek, Anton. 2023b. “Scenario Planning for an A(G)I Future.” Finance and Development 60 (4): 30–33.

Korinek, Anton, Martin Schindler, and Joseph E. Stiglitz. 2022. “Technological Progress and Artificial
Intelligence.” In How to Achieve Inclusive Growth, edited by Valerie Cerra and others, 163–211.
Washington, DC: International Monetary Fund.

Korinek, Anton, and Donghyun Suh. 2024. “Scenarios for the Transition to AGI.” NBER Working Paper
32255, National Bureau of Economic Research, Cambridge, MA.

Liu, Li. 2011. “Do Taxes Distort Corporations’ Investment Choices? Evidence from Industry-Level Data.”
University of Oxford Centre for Business Taxation Working Paper 2010/104, Oxford, UK.

Maffini, Giorgia, Jing Xing, and Michael P. Devereux. 2019. “The Impact of Investment Incentives:
Evidence from UK Corporation Tax Returns.” American Economic Journal: Economic Policy 11 (3):
361–89.

Marimon, Ramon, and Fabrizio Zilibotti. 1999. “Unemployment vs. Mismatch of Talents: Reconsidering
Unemployment Benefits.” Economic Journal 109 (455): 266–91.

McKinsey 2023. “Generative AI and the future of work in America.” McKinsey Global Institute’s Report

Nichols, Austin, and Jesse Rothstein. 2015. “The Earned Income Tax Credit (EITC).” NBER Working
Paper 21211, National Bureau of Economic Research, Cambridge, MA.

Noy, Shakked, and Whitney Zhang. 2023. “Experimental Evidence on the Productivity Effects of
Generative Artificial Intelligence.” Science 381 (6654): 187–92.

Organisation for Economic Co-operation and Development (OECD). 2023a. “OECD Employment Outlook
2023: Artificial Intelligence and the Labour Market.” Paris.

Organisation for Economic Co-operation and Development (OECD). 2023b. “Income Support for
Jobseekers: Trade-offs and Current Reforms.” Paris.

INTERNATIONAL MONETARY FUND 42

STAFF DISCUSSION NOTES Broadening the Gains from Generative AI: The Role of Fiscal Policies

Organisation for Economic Co-operation and Development (OECD). 2023c. “OECD Skills Outlook 2023:
Skills for a Resilient Green and Digital Transition.” Paris. https://doi.org/10.1787/27452f29-en.

Organisation for Economic Co-operation and Development (OECD). 2023d. “Corporate Tax Statistics
2023.” Paris. https://doi.org/10.1787/f1f07219-en.

Organisation for Economic Co-operation and Development (OECD). 2023e. “Revenue Statistics 2023:
Tax Revenue Buoyancy in OECD Countries.” Paris. https://doi.org/10.1787/9d0453d5-en.

Ohrn, Eric. 2019. “The Effect of Tax Incentives on U.S. Manufacturing: Evidence from State Accelerated
Depreciation Policies.” Journal of Public Economics 180: 104084.

Pizzinelli, Carlo, Augustus J. Panton, Marina Mendes Tavares, Mauro Cazzaniga, and Longji Li. 2023.
“Labor Market Exposure to AI: Cross-Country Differences and Distributional Implications.” IMF
Working Paper 2023/216, International Monetary Fund, Washington, DC.

Prettner, Klaus, and Holger Strulik. 2020. “Innovation, Automation, and Inequality: Policy Challenges in
the Race Against the Machine.” Journal of Monetary Economics 116: 249–65.

Ravn, Morten O., and Vincent Sterk. 2021. “Macroeconomic Fluctuations with HANK & SAM: An
Analytical Approach.” Journal of the European Economic Association 19 (2): 1162–202.

Restrepo, Pascual. 2023. “Automation: Theory, Evidence, and Outlook.” NBER Working Paper 31910,
National Bureau of Economic Research, Cambridge, MA.

Schaller, Huntley. 2006. “Estimating the Long-Run User Cost Elasticity.” Journal of Monetary Economics
53 (4): 725–36.

Smets, Frank, and Rafael Wouters. 2007. “Shocks and Frictions in US Business Cycles: A Bayesian
DSGE Approach.” American Economic Review 97 (3): 586–606.

Straub, Ludwig, and Iván Werning. 2020. “Positive Long-Run Capital Taxation: Chamley-Judd Revisited.”
American Economic Review 110 (1): 86–119.

Thuemmel, Uwe. 2023. “Optimal Taxation of Robots.” Journal of the European Economic Association 21
(3): 1154–90.

United Nations. 2023. “Interim Report: Governing AI for Humanity.” New York.

Walsh, Carl E. 2012. Monetary policy and resource mobility. Technical report.

Webb, Michael. 2020. “The Impact of Artificial Intelligence on the Labor Market.” Unpublished.

ZEW. 2021. “Effective Tax Levels Using the Devereux/Griffith Methodology—Update 2020, Final Report.”
Project for the EU Commission TAXUD/2020/DE/308, Leibniz Centre for European Economic
Research, Mannheim.

Zwick, Eric, and James Mahon. 2017. “Tax Policy and Heterogeneous Investment Behavior.” American
Economic Review 107 (1): 217–48.

INTERNATIONAL MONETARY FUND 43

Broadening the Gains from Generative AI: The Role of Fiscal Policies
Staff Discussion Note No. SDN/2024/002