Gondwana Research 131 (2024) 181–195

Contents lists available at ScienceDirect

Gondwana Research
journal homepage: www.elsevier.com/locate/gr

The impact of artificial intelligence application on enterprise

environmental performance: Evidence from microenterprises
_
Yuping Shang a, Silu Zhou a, Delin Zhuang a,⇑, Justyna Zywiołek b
, Hasan Dincer c
a
School of Economics, Hefei University of Technology, Hefei 230009, China
b
Faculty of Management, Czestochowa University of Technology, 42-200 Czestochowa, Poland
c
School of Business, Istanbul Medipol University, Kavacık South Campus, Beykoz 34810, Istanbul, Turkey

a r t i c l e i n f o a b s t r a c t

Article history: The progress of artificial intelligence (AI) technology is an important way to solve the problem of global
Received 7 June 2023 sustainable development. This paper discusses how AI, as a new scientific force, affects the pollution
Revised 17 February 2024 reduction of enterprises from a micro perspective. In this paper, the robot data provided by the
Accepted 18 February 2024
International Federation of Robotics and pollution data of industrial enterprises in China are used for
Available online 19 March 2024
Handling Editor: A. Sinha
empirical analysis to test the effect of AI on pollution performance of enterprises. The results show that
the application of AI can reduce enterprise emissions through technological innovation effect and labor
substitution effect, and this result remains robust after sample substitution and variable substitution.
Keywords:
Further heterogeneity analysis shows that AI has a more significant impact on enterprises in big cities
Artificial intelligence and southern cities, as well as large enterprises, old enterprises and non-technology-intensive enter-
Enterprise environmental performance prises. The research results of this paper provide an effective reference for solving how to achieve green
Robot application and sustainable development of artificial intelligence applications.
Enterprise innovation Ó 2024 International Association for Gondwana Research. Published by Elsevier B.V. All rights reserved.
Labor substitution

1. Introduction and low-carbon advanced environmental protection technology

and equipment. In addition, enterprises take the initiative to carry
Nowadays, environmental problems in the world are attracting out industrial transformation. The enterprise adheres to high-
more and more attention. Reducing pollution emission level and quality development and structural adjustment and optimization
comprehensively improving the operation quality of pollution con- in parallel, and makes a strong contribution to the goal of carbon
trol facilities have become the unified goal of environmental gov- peak and carbon neutrality as scheduled.
ernance in all countries. A healthy and sustainable ecological In November 2022, the World Meteorological Organization
environment is the basic condition for human survival and the (WMO) released its interim report on the State of the Global Cli-
basis for the continuation of human civilization. In order to live mate 2022. The report says the effects of bad weather are getting
in harmony between man and nature, countries around the world worse. Among them, the most obvious features of climate change
are actively launching a series of environmental protection policy are global warming and declining air quality. Since industrializa-
meetings. Governments are actively implementing more effective tion, the concentration of greenhouse gases has been rising, and
measures for global pollution management. According to the cur- the heat has been accumulating, which has led to the extreme
rent data of global climate change and surface water pollution, it melting of glaciers. According to satellite monitoring, the sea level
is necessary to further strengthen the efforts of pollution control. rise in the past two and a half years has reached 10 % of the
It is not only necessary to reduce the emission of pollution from increase in the past three decades. In addition, according to air
the source, but also to carry out efficient treatment of the inevita- quality monitoring results, only 23 % of the world’s people live in
ble discharge of pollution under the current level of technology. As areas with average annual concentrations of NO2 that meet WHO
the largest developing country, China has taken the lead in the field air quality guidelines. In 2022, the annual average concentration
of pollution prevention and control. China has vigorously strength- of nitrogen dioxide and particulate matter such as PM10 and
ened technological innovation and service innovation in its indus- PM2.5 were added in the first update of the WHO Air Quality data-
try. And it strives to actively develop safe, advanced, energy-saving base. These air pollutants do great harm to the human body. The
World Health Organization estimates that more than 13 million
⇑ Corresponding author. people worldwide die each year from preventable environmental
E-mail address: [email protected] (D. Zhuang). causes. Bad environmental conditions can even lead to epidemics

https://doi.org/10.1016/j.gr.2024.02.012
1342-937X/Ó 2024 International Association for Gondwana Research. Published by Elsevier B.V. All rights reserved.
Y. Shang, S. Zhou, D. Zhuang et al. Gondwana Research 131 (2024) 181–195

like the new coronavirus, and humanity has already paid a terrible in 2010 are shown in Fig. 1, and the visual analysis results of word
price. Therefore, governments should attach great importance to cloud in 2020 are shown in Fig. 2. Comparing Fig. 1 and Fig. 2, it can
environmental governance and introduce relevant policies and be seen that the number of intelligent words mentioned in enter-
take actions. prise annual reports shows an increasing trend year by year, and
Faced with the imminent problem of environmental gover- intelligence has become an important development direction and
nance, governments and enterprises of various industries are trend of enterprises, especially in 2020.
thinking deeply and increasing the investment in pollution preven- In such an environment, this paper constructs an AI level index
tion and control. Nowadays, with the deepening of globalization, based on the robot data released by IRF from 2000 to 2014. Based
the development of Internet and computer technology is also very on the industrial intelligence level index and the pollution emis-
fast. AI, as a product of the information revolution in the new era, sion data of Chinese enterprises in the same period, this paper con-
has been attached great importance by all industries. AI is a deriva- ducts correlation analysis and econometric regression. In addition,
tive of computer science, a machine that can react in a way that the paper studies and discusses the influence of AI application on
mimics human intelligence. AI comes from human beings and pollution emission intensity of enterprises and related mecha-
serves human beings. It is a system integrating language recogni- nisms. The results show that the introduction of robots in various
tion, image recognition, natural language processing and chip tech- industries in China has surged since 2005. The level of AI has
nology. Because of the nature of AI, in a sense, its memory storage improved accordingly. Intelligent application has significantly
capacity is better than that of humans. reduced pollution emission intensity of enterprises. This result
In fact, in today’s era of information explosion, the development remains robust after relevant tests such as replacement of sample
of information and digitalization has long become one of the data and explanatory variables. The mechanism test shows that AI
important areas of international competition. To take the lead, promotes the reduction of pollutant emission level by improving
we must accelerate digital development and cultivate our core the level of technological innovation and substituting the labor
technologies in the digital economy. Different from the agricultural force. Further heterogeneity analysis results show that enterprises
economy and industrial economy in the past, the digital economy in big cities, older and larger enterprises and enterprises in south-
has a wider range of radiation, deeper influence and faster develop- ern cities have more obvious benefits in using AI to promote emis-
ment speed. As a representative of digital technology, artificial sion reduction. Moreover, non-technology exchange intensive
intelligence has emerged in all walks of life, showing its unique enterprises have a better effect on improving environmental per-
charm. Based on the characteristics of high resolution and high formance after introducing AI than technology exchange intensive
efficiency of artificial intelligence, it will also have a strong force enterprises. Therefore, promoting the development of AI and
in the future of enterprise emission reduction and pollution pre- encouraging enterprises to use AI to control enterprise pollution
vention. At present, the global climate environment is getting emissions is conducive to sustainable development. This paper
worse and worse, and technical barriers are difficult to break demonstrates the role of AI in improving environmental perfor-
through. Therefore, the development of artificial intelligence and mance through multi-layer data analysis and empirical tests. The
its research in pollution prevention and control are urgent. Only research of this paper has important enlightenment for industrial
by taking the lead in the field of artificial intelligence and master- enterprises to reasonably improve the level of intelligence to pro-
ing its core technology of pollution control can we curb the further mote energy conservation and emission reduction.
expansion of pollution emissions earlier. The innovation and research significance of this paper are as fol-
It can be seen from relevant literature that AI does promote lows: First, the impact of AI on environmental performance is stud-
environmental performance (Zhang et al., 2021). On the other ied from a micro perspective. At present, many international
hand, the introduction of AI technology has improved the produc- researches focus on the macro influence mechanism, but lack of
tion efficiency of enterprise (Graetz & Michaels, 2018). In this pro- in-depth study on the individual of micro economic activities. It
cess pollutants are inevitably increased as concomitants of the final is found that the economic behavior of micro enterprises can be
product. With the continuous development of intelligent technol- used to distinguish more clearly the way of AI’s effect on environ-
ogy and gradually mature theory, how enterprises can reduce pol- mental performance. Based on the analysis of micro-data at the
lution emission through AI has important research significance. At industry and enterprise level, this paper inspires relevant depart-
present, relevant researches on artificial intelligence have long ments around the world to conduct environmental governance
confirmed that the introduction of artificial intelligence can through rational use of AI from the perspective of enterprises, so
improve the innovation performance of enterprises (Klump et al., as to improve global environmental performance. Second, existing
2007). In addition, in the process of enterprises using intelligent theories measure AI indicators from multiple perspectives, but few
machines, ordinary labor will be gradually replaced (Acemoglu & use robot inventory data to describe the trend of AI in detail. There-
Restrepo, 2018; 2020). In addition, studies have shown that the fore, this paper matches the robot inventory data to relevant indus-
extensive use of artificial intelligence can improve the production tries and analyzes its development trend. The analysis shows that
efficiency of enterprises (Graetz & Michaels, 2018) and expand the number of robots increases rapidly after 2005 and 2010. In
the scale of enterprises (Herrendorf et al., 2013). AI can also bring addition, through the mechanism test, this paper confirms that
a certain level of economic growth to businesses. The factors AI promotes the emission reduction of enterprises by improving
affecting the environmental performance of enterprises are closely the level of technological innovation and replacing the labor effect.
related to the level of technological innovation, labor force struc- Although the existing theories have found that the improvement of
ture and production scale of enterprises. Therefore, the effect intelligence level has a positive correlation with environmental
brought by artificial intelligence will affect the environmental per- performance, the researches on the influencing ways are not deep
formance of enterprises. Based on this, this paper deeply studies enough. The research in this paper provides a direction for enter-
the path of artificial intelligence affecting enterprise environmen- prises to use AI to improve their emission reduction ability with
tal performance. higher benefits. It can help enterprises play the role of AI at a
To systematically reveal the intelligent development direction higher level. Finally, in the process of research, this paper also
of enterprises, this paper selects the annual reports of 4327 listed found that the characteristics of enterprises, urban location, urban
enterprises in China during 2010–2020. With the help of machine type, etc., will make a difference in the influence of AI on the emis-
learning algorithm, the key word frequency hotspots related to sion level of enterprises. Third, from the perspective of research
intelligence are mined. The visual analysis results of word cloud data, many existing studies use regularized task intensity
182
Y. Shang, S. Zhou, D. Zhuang et al. Gondwana Research 131 (2024) 181–195

Fig. 1. Results of word cloud map analysis in 2010.

(Gregory et al., 2022) and total factor productivity (Autor & on enterprise pollution emission. The sixth part is the conclusion
Salomons, 2018) to measure intelligence level. The core explana- and policy enlightenment.
tory variable of this paper uses the robot inventory index at the
industry level. It avoids the deviation caused by uncontrollable fac-
tors caused by other indicators and makes the research results 2. Literature review and mechanism analysis
more authoritative. In addition, through the data analysis of Chi-
nese industrial enterprises, combined with the intelligence level 2.1. Literature review
index to obtain a more convincing influence mechanism.
The rest of this paper is arranged as follows: The second part is Due to the basic performance of strong applicability of AI, all
literature review and theoretical mechanism analysis. The third walks of life introduce AI to improve the development of the indus-
part is the index measure and data fact. The fourth part reports try. In the future, as the main body of economic development in the
the empirical estimation results and analyzes them. The fifth part market, enterprises will also increase the proportion of intelligent
further examines the mechanism test and heterogeneity of AI level machines (Bughin et al., 2018). As intelligent technology becomes
183
Y. Shang, S. Zhou, D. Zhuang et al. Gondwana Research 131 (2024) 181–195

Fig. 2. Results of word cloud map analysis in 2020.

more and more mature, a large number of scholars also have a new generation of digital technologies, artificial intelligence plays
strong interest in the impact of intelligence and the impact of dig- an important role in the process of digital transformation of enter-
ital technology. Siebel (2019) indicates that digital transformation prises. It is found that the most direct phenomenon after AI enters
is a process in which enterprises innovate and reform traditional the market is to replace the ordinary labor force (Zhao et al., 2022;
production mode, management mode and core business through Frey & Osborne, 2017; Acemoglu & Restrepo, 2018; 2020). As a
digital technology. Existing literature suggests that digital transfor- result, working hours are automated and unemployment rises
mation can not only drive innovation, but also transfer knowledge (Autor et al., 2006; Fossen & Sorgner, 2022) and the income gap
and technology (Acs et al., 2009). At the enterprise level, digital increases accordingly (Michaels et al., 2014).
transformation can improve the economic efficiency and financial However, some studies show that although AI has a substitution
performance of enterprises (Gal et al., 2019). In addition, digital effect on labor force, it will not affect the employment rate (Arntz
transformation can also affect the productivity of enterprises et al., 2016; Dauth et al., 2017; Graetz & Michaels, 2018). At the
through the upgrading of organizational structure (Galindo- same time, it will not increase the income gap (Kunst, 2020;
Martín et al., 2019) and the reduction of operating costs Acemoglu & Restrepo, 2020). In addition, Furman and Seamans
(Goldfarb and Tucker, 2019). As one of the representatives of the (2019) found that the use of AI can improve enterprise production

184
Y. Shang, S. Zhou, D. Zhuang et al. Gondwana Research 131 (2024) 181–195

(Graetz & Michaels, 2018) and economic growth (Gasteige & The research results of this paper further expand the understand-
Prettner, 2017; Graetz & Michaels, 2018) as well as reducing enter- ing of digital technology, and enrich the path of enterprise pollu-
prise costs and expanding enterprise scale (Herrendorf et al., tion reduction.
2013). In addition, there are also literatures on the impact of AI
on household finance from a micro perspective. The introduction 2.2. Mechanism analysis
of AI has led to a general decline in household income. At the same
time, it reduces household investment in the capital market and Based on the literature review, this paper further studies the
increases the gap between rich and poor (Sachs & Kotlikoff, micro-mechanism of AI on environmental performance. In this
2012; Jansson & Karabulut, 2018). Another kind of literature about part, the effects of AI on reducing pollution emissions of enter-
the study of corporate environmental performance. Environmental prises are studied from two aspects: technological innovation
problem is a hot topic all over the world, and also an accompanying effect and labor substitution effect, as shown in Fig. 3.
problem on the road of production and development of enterprises. The use of intelligent machines can drive technological compe-
Modern China is actively transforming from an agricultural power tition in the industry and comprehensively improve the level of
to an industrial power. It is inevitable that there will be some technological innovation (Klump et al., 2007). If an enterprise
enterprises in order to seize the market and vigorously develop wants to take advantage in the industry to become the leader of
production while ignoring the control of pollution emissions. Due the industry, in addition to increasing production to occupy market
to global climate change and serious greenhouse effect, the deteri- share, the more critical is the competition of technology. The intro-
oration of air and water quality has gradually attracted the atten- duction of AI helps technicians solve basic tasks, such as searching
tion of governments all over the world. Since then, the pollution for information and processing data, more efficiently. In this way,
emission of industrial enterprises in various countries has been technical researchers can be more focused on the research and
strictly controlled. Environmental performance has also become development of innovative products. Mutual learning and compe-
the basis for measuring enterprise ethical standards. It has been tition between different enterprises in the industry will lead to the
observed from the existing literature that the environmental per- continuous improvement of the level of innovative technology.
formance of enterprises is affected by many factors. It includes Technological innovation will improve the original working mode
government intervention and regional environmental regulation of enterprises and make the production level of enterprises more
(Henriques & Sadorsky, 1996; Shapiro & Walker, 2018; Zhang modern and intelligent. From the perspective of production equip-
et al., 2022). On the other hand, from the perspective of the inter- ment of enterprises, technological innovation brought by AI can
nal conditions of enterprises, the types and scale of enterprises and improve and upgrade old equipment of enterprises (Ye et al.,
the governance level of enterprise management (Lewis et al., 2014; 2020). The upgraded new equipment has a higher level of technol-
Endo, 2020; Zhang et al., 2020; Hoang et al., 2021) can also affect ogy. It can treat the factors of production efficiently and reduce the
the pollution emission level of enterprises. output of pollutants. In addition, from the perspective of produc-
At present, many scholars have raised questions about the tion procedures, technological innovation can enable enterprises
impact of AI on environmental performance and conducted rele- to carry out intelligent transformation (Bessen, 2019). In the com-
vant studies. As early as in the late 20th century, some scholars plicated production process, enterprises can reasonably use inno-
proposed the possible relationship between AI and environmental vative robot equipment for collaborative management. And
performance (Cortés et al., 1999). However, AI is not widely used, innovative robots have higher resolution. It can separate and treat
and the data is not comprehensive enough to conduct in-depth pollutants in time while manufacturing products, so as to reduce
research on it. Zhang (2021) studied data at the provincial level pollution emissions. Therefore, this paper proposes:
and found that AI technology could indeed promote green produc- Hypothesis 1 The application of AI can improve the environ-
tion and improve environmental performance. The research on mental performance of enterprises through the effect of technolog-
water pollution control shows that the application of intelligent ical innovation.
technology can help enterprises improve the ability of pollution The most direct advantage of AI is to substitute labor force
control. In addition, it combines with big data to grasp the circula- (Zhao et al., 2022; Frey & Osborne, 2017; Acemoglu & Restrepo,
tion of real-time information and realize the detection and preven- 2020). In particular, assembly line labor that does not require high
tion of pollution (Ye et al., 2020). However, due to the availability technology or labor that is highly fungible (Autor & Dorn, 2012).
of data, detailed micro-level exploration was not done at that time. From the perspective of enterprise production, the working effi-
Therefore, on the basis of the existing papers, this paper deeply ciency of AI is higher than that of ordinary labor force. At the same
studies the micro-mechanism of AI on environmental performance. productivity intensity, the cost of purchasing AI machines and
In addition, this paper provides relevant suggestions for global maintaining them is much lower than the cost of wages, benefits,
environmental governance at the enterprise level. labor search and training. Moreover, the application of AI can not
Through the review of the above literature, it can be seen that AI only improve production efficiency, but also reduce commodity
is of great significance for enterprises to improve environmental prices (Graetz & Michaels, 2018). By using AI to replace ordinary
performance. However, the existing literature on AI mainly dis- labor force, enterprises can invest more money in pollutant dis-
cusses the impact of AI on the labor market at the macro level charge equipment and pollutant discharge technology, so as to
(Acemoglu & Restrepo, 2018; 2020). With the wide application of promote their emission reduction. On the one hand, AI can under-
artificial intelligence, some scholars have begun to study the take work intensity and precision that ordinary labor force cannot
impact of artificial intelligence on the economic efficiency of enter- undertake in the industrial production process. AI can reduce the
prises at the micro level (Graetz & Michaels, 2018). But at present, production of pollutants at their source and improve the environ-
there are few researches on the micro level. In addition, in terms of mental performance of enterprises. On the other hand, the substi-
corporate environmental performance, most of the existing litera- tution effect of AI on labor force is often reflected in lower-skilled
ture is considered from the perspective of internal management positions. Low-tech jobs are replaced by AI, and more labor force
and external environment. At present, there is little literature will move to higher-tech jobs in order to prevent unemployment
focusing on the positive environmental effects of artificial intelli- (Autor & Dorn, 2012; Kunst, 2020). Such competition for jobs will
gence for enterprises. Therefore, from the perspective of micro lead to a constantly improving skill level of the workforce. It will
enterprises, this paper innovatively studies the impact of artificial also promote enterprises to reduce pollution to a certain extent.
intelligence on environmental performance and its mechanism. Therefore, this paper proposes:
185
Y. Shang, S. Zhou, D. Zhuang et al. Gondwana Research 131 (2024) 181–195

Fig. 3. Workflow and approach of this paper.

Hypothesis 2 The application of AI can improve enterprise envi- formance. Finally, from the perspective of enterprise scale, differ-
ronmental performance through artificial substitution effect. ent enterprises have different scales due to the differences in
From the point of view of the city, the geographical location and business. Larger enterprises generally have stronger productivity
natural environment of different regions are different. Therefore, and higher profitability. After introducing AI, enterprises should
different cities have different environmental policies and levels cultivate professional talents who make good use of AI. It is also
of development. From the perspective of Chinese cities, the devel- necessary to maintain and upgrade relevant instruments regularly.
opment of southern regions is generally higher than that of north- The process of AI adaptation is long and requires certain invest-
ern regions in terms of economic level and development level. ment costs. Therefore, the introduction of AI in larger enterprises
Because of its natural geographical advantages, southern cities can better play its role. Therefore, this paper proposes:
have more frequent exchanges with other countries and maritime Hypothesis 3 The application of AI has a greater effect on the
trade, and have more extensive ways to get information (Xu et al., environmental performance improvement of enterprises in big
2023). Therefore, such cities were exposed to the idea of AI earlier cities, enterprises in southern cities, large enterprises and non-
and studied the application of AI in production earlier. Their AI technology exchange intensive enterprises.
technology is more mature. In addition, there are more industrial
enterprises in the southern cities, and the effect of AI on labor sub-
3. Index measure and typical fact
stitution is most obvious in labor-intensive industries. Therefore, it
is more conducive to promoting enterprises in areas with higher
3.1. Model specification
economic levels. In addition, from the perspective of the industry,
generally speaking, the pollution emission level of non-technology
The research focus of this paper is to investigate the influence of
exchange intensive industries is higher than that of technology
AI level on enterprise pollution emission performance. Therefore,
exchange intensive industries. In the process of production and
the following measurement model is set in this paper:
manufacturing in non-technical communication intensive indus-
tries, pollutants will be produced. In many cases, in order to lnenijt ¼ a0 þ a1 lnintelljt þ dY ijt þ FEindu þ FEcity þ FEyear þ eijt ð1Þ
improve the output of enterprises, the pollution level of enterprises
will also increase. However, technology-intensive industries are Where, i, j and t represent enterprise, city and year respectively. The
more profitable by means of information exchange and resource explained variable en represents enterprise pollution emission
services. The products of this kind of industry are mostly virtual intensity. The core explanatory variable intell represents the AI level
products, which will not bring actual pollution products. Therefore, of the industry. Y represents the enterprise control variable. FEindu,
the application of AI in non-technical communication intensive FEcity and FEyear represent the fixed effect of two-digit industry, city
industries can reduce the generation of pollutants from the root and time, respectively. a0 is a constant term and eijt is a random
and promote the improvement of enterprises’ environmental per- error term.
186
Y. Shang, S. Zhou, D. Zhuang et al. Gondwana Research 131 (2024) 181–195

3.2. Index measure robots, resulting in the number of robots in many industries before
this is zero. The core explanatory variable of this paper is measured
(1) Enterprise pollution emission intensity (en). Because the by adding one millionth on the basis of the ratio of robot inventory
pollution emission of industrial enterprises is related to many fac- to the total number of people in the industry. The millionth here is
tors, and the availability of data is restricted. Therefore, this paper also to eliminate the deviation caused by data missing or zero. In this
constructs the index of pollution emission of industrial enterprises paper, artificial intelligence industry data is used to match pollution
by referring to relevant literatures (Shang et al., 2022; Song et al., data at the enterprise level. The data of artificial intelligence level is
2022). Taking the above situation into consideration, this paper limited to the industry level, and there is no measurement of the
finally selected the four most concerned pollutant indexes of intelligence level at the enterprise level. Therefore, this paper makes
industrial wastewater emission intensity (iw), chemical oxygen reference to relevant literature (Geng et al., 2022), subdivides the AI
demand emission intensity (cdo), sulfur dioxide emission intensity data at the industry level, matches and measures the pollution data
(sde) and nitrogen oxide emission intensity (ne) to build a compre- at the enterprise level.
hensive pollutant index. Firstly, the data of these four indicators (3) Control variable. In order to avoid the possibility of missing
are standardized (Dagestani et al., 2023). Take sulfur dioxide emis- variable bias, this paper controls the characteristic variables at the
sion intensity as an example: enterprise level. In this paper, the average annual number of all
employees (cyrs) is used to represent the enterprise size. The age
sdeijt  minsdeit of the enterprise (age) is represented by the difference between
rsdeit ¼ ð2Þ
maxsdeit  minsdeit the establishment time of the enterprise and the date of data inter-
After processing, this paper adjusts the coefficient of each index ception. The ratio of capital to labor (assetp) is expressed by the
relative to the mean value. This paper still takes sulfur dioxide ratio of total assets to the number of employees. The Income-
emission intensity as an example: labor ratio (incomp) is expressed as primary business sales revenue
per unit of employment. Finally, it should be noted that the
rsdeit enterprise-level data in this paper are from the database of Chinese
wsdeit ¼  ð3Þ
rsdeit industrial enterprises from 2001 to 2014. Moreover, the method of
Brandt et al. (2012) is used for intertemporal matching of indus-
Finally, the comprehensive index of pollution emission required trial enterprise database. To avoid the possibility of collinearity
in this paper is constructed according to the equal weight: and heteroscedasticity, this paper also takes the natural logarithm
of related data variables for regression. In addition, samples with
1
enijt ¼ ðriwit  wiwit þ rcdoit  wcdoit þ rsdeit  wsdeit total assets, number of employees, industry category, year of oper-
4
ation, and sales less than or equal to zero or missing data were
þ rneit  wneit Þ ð4Þ removed. In addition, the sample of enterprises with fewer than
8 employees, sales value less than 5 million and enterprise age over
Where, max and min represent the maximum and minimum values
 100 is deleted. In general, if the age of an enterprise is greater than
respectively, and rsdeit represent the mean value of sulfur dioxide 100, it was born before the 20th century and is far away from now.
emission intensity. In Formula (4), the larger the value of en, the The nature and system of enterprises are quite different from those
higher the pollution intensity. of new enterprises, so the comparative study is of little signifi-
(2) Artificial intelligence level (intell). The data of AI level in this cance. At the same time, we find that enterprises with enterprise
paper mainly come from China’s robot inventory data from 2001 to age greater than 100 have serious missing variables, which is not
20141 published by the International Federation of Robotics (IRF). At conducive to the research.
present, IRF’s robot data has been widely used in relevant researches Chinese Industrial Enterprise Database is one of the most used
on AI (Cheng et al., 2019; Acemoglu & Restrepo, 2020). The industry databases in domestic academic research in recent years. Accord-
codes in the data classification details published by IRF are not con- ing to the incomplete statistics of Google Academic Database, by
sistent with the industry classification in China. In order to unify the the end of 2016, more than 200 overseas academic papers had
data, this paper compares the industries corresponding to the robot been used in the database of Chinese industrial enterprises. His
inventory data released by IRF with the two industries in the Indus- publications include the Quarterly Journal of Economics, American
try Classification of National Economy (2012).2 Based on the Economic Review, Review of Economic Studies, and Economic Journal,
research method of Graetz & Michaels (2018), this paper adds one Journal of Development Economics, Journal of Urban Economics, etc.
part per million to the total number of working robots in the indus- The above-mentioned international mainstream economics jour-
try. Finally, lnintell is taken to represent the level of AI. Before 2005, nals have successively published empirical papers based on Chi-
AI was not fully popularized in China. Many industries do not use nese Industrial Enterprise Database, in addition to the excellent
papers themselves. Also because of the high quality of the data-
1
The AI in this article can measure data from recent years. However, because the base, it can make a good micrometrology paper, which is not avail-
pollution data of Chinese industrial enterprises is not fully published, the data that able in other data. Chen’s (2018) research also confirmed that the
can be measured in this paper can only be updated until 2014 (Shang et al., 2022).
data base of Chinese Industrial Enterprise Database from 1996 to
Therefore, in order to match the pollution data of enterprises, the data from 2001 to
2014 is selected as a whole for measurement.
2013 was consistent with the sample coverage in the industrial
2
The one-to-one correspondence between the IRF classification and the industry section of China Statistical Yearbook and China Industrial Statisti-
classification of China’s national economy is as follows: Food and beverage industry cal Yearbook in previous years. To ensure the regularity and accu-
IRF (10–12) corresponds to (13–15). textile industry IRF (13–15) corresponds to (17– racy of regression results. The final data required by this research is
19). wood and furniture industry IRF (16) corresponds to (20–21). paper industry IRF
obtained through integration, and the statistical descriptive data is
(17–18) corresponds to (22–23). pharmaceutical and cosmetics industry IRF (19)
corresponds to (2 6 ) 27). other chemical products industry IRF (20–21) corresponds shown in Table 1.
to (26). rubber and plastics industry IRF (22) corresponds to (29–30). glass, ceramics,
stone and mineral products industry IRF (23) corresponds to (31). basic metal 3.3. Typical fact
industry IRF (24) corresponds to (33). metal products industry IRF (25) corresponds to
(34)). automobile industry IRF (29) corresponds to (3 7 2). other transportation
industry IRF (30) corresponds to (37). other manufacturing industry IRF (91)
(1) Evolution characteristics of AI level in China. According to
corresponds to (42). electronic components and equipment industry IRF (2 6 0) the robot inventory in different years, this paper draws the annual
corresponds to (39). semiconductor industry IRF (2 6 1) corresponds to (4 0 5). growth trend chart, as shown in Fig. 4. It is not difficult to find that
187
Y. Shang, S. Zhou, D. Zhuang et al. Gondwana Research 131 (2024) 181–195

Table 1
Results of descriptive statistics of main variables.

Variable Obs Mean Std. Dev. Min Max

lnen 3,162,929 16.0188 1.7508 16.5601 12.3878
lnintell 3,162,929 12.3721 3.7774 16.1181 0.1726
lncyrs 3,162,929 4.9402 1.1002 2.0794 12.3717
lnage 3,162,929 1.9977 0.7695 0.0000 4.7449
lnassetp 3,162,929 5.1630 1.1294 0.0017 15.8705
lnincomep 3,162,929 5.6612 1.0604 0.2841 16.1837

before 2005, the stock of robots in China was close to 0. Since 2005,
it has grown. After four or five years of mature development, the
application level of AI has been improved. Intelligent development
has increased, and the number of robots has increased further
since 2009. According to the actual situation of intelligent develop-
ment in China, the level of intelligence was almost zero before
2005. The reason is that before 2005, China paid more attention
to infrastructure construction and AI research and development
did not receive much attention. But there was a significant spike
after 2005, as the Chinese government invested more in AI. In
the Chinese market, after several years of running-in, the combina-
tion of AI and traditional industry achieved initial maturity and
stable growth after 2009. This is a trend in line with China’s
national conditions.
In addition, as shown in Fig. 5, this paper considers the different
demands for intelligence in different industries. This paper
extracted the stock data of different industries from China’s robot
Fig. 4. Data of China’s robot stock from 2001 to 2014. inventory data from 2001 to 2014 for comparative analysis. It

Fig. 5. Data of China’s robot stock by industry from 2001 to 2014.

188
Y. Shang, S. Zhou, D. Zhuang et al. Gondwana Research 131 (2024) 181–195

4. Empirical analysis

4.1. Benchmark regression

The baseline regression results of this paper are shown in

Table 2. According to column (1), the estimated coefficient of AI
level on enterprise pollution emission intensity is significantly
negative at the 1 % level. This means that the improvement of AI
level can promote enterprises to reduce pollution. And this paper
preliminarily verifies the judgment of linear fitting graph in corre-
lation analysis. Control variables are added starting from column
(2), and columns (3) - (5) control the two-digit industry fixed
effect, city fixed effect and time fixed effect respectively. column
(6) further controls the clustering standard error, and the signifi-
cance of the regression coefficient does not change. The regression
coefficients of 6 columns in Table 2 are plotted as Fig. 7. It can be
seen that the regression coefficient of artificial intelligence applica-
Fig. 6. Correlation between AI and enterprise pollution emission intensity.
tion on enterprise pollution emission is always significantly nega-
tive. Therefore, the improvement and wide application of AI can
can be clearly observed from the trend chart that the stock of help enterprises reduce pollution emissions and improve environ-
robots in the automobile industry (29) grows faster than that in mental performance.
other industries. Non-technology-intensive industries such as tex- As for the direction of control variables, employment scale, age
tiles (13–15) are the least intelligent. In addition, it can be seen and capital-labor ratio of enterprises all have positive effects on
from Fig. 5 that the intelligence level of various industries in China pollutant discharge of enterprises. This indicates that the larger
has increased to varying degrees. The growth rate of non- the employment scale, the age of the enterprise and the ratio of
technology exchange intensive industries (13–15, 16, 17–18) is capital to labor, the higher the emission level of the enterprise.
lower than that of other industries. Among them, the higher the income labor ratio of enterprises will
(2) Correlation analysis. Based on the need of preliminary anal- promote the reduction of pollution emission. This is because com-
ysis of research objectives, this paper matches the AI data with the panies with a larger income to labor ratio belong to industries with
pollution emission data of enterprises from 2001 to 2014. On this intensive technology exchange. Technology industries typically
basis, this paper draws binscatter - linear fitting maps, as shown emit less pollution than non-technology industries. And
in Fig. 6. Through the linear graph, it is not difficult to find that technology-intensive industries cut emissions better than less
the application of AI has an obvious negative correlation with technologically demanding industries. After adding control vari-
enterprise pollution emission. Therefore, it is preliminarily judged ables, the corresponding core explanatory variables are still signif-
that the extensive application of AI is helpful to reduce the pollu- icantly negative, as shown in column (2) of Table 2. This means
tion emission of enterprises. Thus, it is beneficial to promote the that the introduction of AI to these enterprises have played a role
improvement of enterprise environmental performance. Of course, in curbing pollution. In general, the significance level of the core
further accurate measurement tests are needed to determine explanatory variables concerned in this paper did not change after
whether this conclusion is strictly statistically significant. the addition of control variables and fixed effects. Therefore, the

Table 2
Benchmark regression results.

Univariate Add control Add Industry fixed Add year fixed Add city fixed Control clustering standard
regression variables effects effects effects error
(1) (2) (3) (4) (5) (6)
lnen lnen lnen lnen lnen lnen
lnintell 0.0240*** 0.0314*** 0.0353*** 0.0057*** 0.0047*** 0.0047**
(0.0003) (0.0003) (0.0003) (0.0006) (0.0006) (0.0020)
lncyrs 0.2064*** 0.2132*** 0.2410*** 0.2297*** 0.2297***
(0.0009) (0.0009) (0.0010) (0.0010) (0.0068)
lnage 0.1613*** 0.1614*** 0.1637*** 0.1934*** 0.1934***
(0.0013) (0.0013) (0.0013) (0.0013) (0.0070)
lnassetp 0.2202*** 0.2233*** 0.2188*** 0.2237*** 0.2237***
(0.0011) (0.0011) (0.0011) (0.0011) (0.0073)
lnincomep 0.2117*** 0.2336*** 0.2160*** 0.2027*** 0.2027***
(0.0012) (0.0012) (0.0012) (0.0012) (0.0065)
Constant 16.3155*** 17.6880*** 17.6616*** 17.5139*** 17.6055*** 17.6055***
(0.0034) (0.0100) (0.0103) (0.0112) (0.0111) (0.0576)
Observations 3,162,929 3,162,929 3,162,929 3,162,929 3,162,929 3,162,929
Adjusted R2 0.0027 0.0477 0.0991 0.1024 0.1378 0.1378
Industry FE NO NO YES YES YES YES
Year FE NO NO NO YES YES YES
City FE NO NO NO NO YES YES

Note: Robust standard errors in parentheses, *** p < 0.01, ** p < 0.05, * p < 0.1. Same below.

189
Y. Shang, S. Zhou, D. Zhuang et al. Gondwana Research 131 (2024) 181–195

explanatory variable were removed. Only the ratio of the original

data robot stock to the total number of people in the industry is
returned, and the result is shown in column (3). Finally, again
without adding one part per million, logarithm of the robot stock
data is directly taken for regression. The results obtained are
shown in column (4). In general, after replacing the AI measure,
all regression coefficients are still significantly negative at the level
of 1 %. It is further confirmed that AI can promote the pollution
reduction of enterprises.
(2) Change the explained variable. In order to prevent the
empirical results from being biased due to the undiversified
explained variables, this paper also introduced industrial wastew-
ater emission intensity (iw), chemical oxygen demand emission
intensity (cdo), sulfur dioxide emission intensity (sde) and nitrogen
oxide emission intensity (ne) as explained variables. According to
Fig. 7. The coefficient of the baseline regression.
the unified standard, the logarithm of the introduced substitution
variable is taken for regression. The regression results are shown
theoretical hypothesis of this paper is further verified. The conclu- in Table 4. Taking column (1) as an example, the application of
sion of this paper is consistent with the reality and provides empir- AI has a significantly negative effect on the discharge of industrial
ical evidence for the reality development. wastewater. This indicates that if the AI level increases by 1 %, the
industrial wastewater discharge will decrease by about 0.005 %. It
4.2. Robustness test can be seen that AI is conducive to reducing pollution emission
intensity of Chinese enterprises. Except for column (2) in Table 4,
(1) Change explanatory variables. For the validity of the empir- all regression results are as significantly negative as those in col-
ical results, the robot data is processed and logarithm is taken. umn (1). Chemical oxygen demand emission is the sum of chemical
Firstly, regardless of the influence of the number of industry oxygen demand emission in industrial wastewater and domestic
employees, the index of industrial robot stock is added by one unit sewage. And the regression of industrial wastewater discharge
and logarithm is taken as the substitute index of AI. The regression intensity indicated by column (1) has been significantly negative.
was conducted again, and the results were shown in column (1) of However, since the domestic sewage data is too broad and cannot
Table 3. Secondly, drawing on the method of Acemoglu & Restrepo obtain complete data, which is an uncontrollable factor, the regres-
(2020), the number of robots added can also be used to replace the sion result does not show negative effect.
robot stock. Similarly, after adding one unit, the logarithm is taken (3) Change the sample. In order to eliminate the interference of
as the substitute index of AI index, and the regression is carried out other factors in the sample to the results, this paper eliminated the
again. The results are shown in column (2) in Table 3. In addition, extreme value in the sample. Since the quality of enterprise sample
the parts per million that were originally added to the core data in 2014 needs to be investigated, data of 2014 is excluded in

Table 3
Robustness test 1: changing explanatory variables.

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

lnen lnen lnen lnen
lnstock 0.0048***
(0.0006)
lninstallations 0.0036***
(0.0006)
lnintell2 0.0163***
(0.0013)
lnstock2 0.0091***
(0.0015)
Observations 3,162,929 3,162,929 1,924,725 1,924,725
Adjusted R2 0.1378 0.1378 0.1138 0.1137
control variable YES YES YES YES
Industry FE YES YES YES YES
Year FE YES YES YES YES
City FE YES YES YES YES

Table 4
Robustness test 2: changing the explained variable.

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

lniw lncdo lnsde lnne
lnintell 0.0049*** 0.0069*** 0.0208*** 0.0340***
(0.0018) (0.0021) (0.0020) (0.0034)
Observations 413,085 390,621 362,090 236,295
Adjusted R2 0.3621 0.3824 0.4120 0.3979
control variable YES YES YES YES
Industry FE YES YES YES YES
Year FE YES YES YES YES
City FE YES YES YES YES

190
Y. Shang, S. Zhou, D. Zhuang et al. Gondwana Research 131 (2024) 181–195

Table 5
Robustness test 3: changing samples.

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

lnen lnen lnen lnen
lnintell 0.0043*** 0.0141*** 0.0047*** 0.0076**
(0.0006) (0.0008) (0.0006) (0.0032)
lnintell  lnintell 0.0001
(0.0001)
Observations 2,988,762 2,263,576 2,854,840 3,162,929
Adjusted R2 0.1410 0.1204 0.1388 0.1378
control variable YES YES YES YES
Industry FE YES YES YES YES
Year FE YES YES YES YES
City FE YES YES YES YES

Table 6
Regression results of endogeneity test.

Further controls the fixed effect Instrumental variable regression

Add industry-city Add city-year Add industry FE Add year FE Add city FE
(1) (2) (3) (4) (5)
lnen lnen lnen lnen lnen
lnintell 0.0057*** 0.0051*** 0.0447*** 0.0071*** 0.0028***
(0.0006) (0.0006) (0.0004) (0.0014) (0.0014)
First stage regression
IV – – 0.7690*** 0.3941*** 0.3939***
(0.0004) (0.0005) (0.0005)
F – – 3.9e + 06*** 6.4e + 05*** 6.4e + 05***
Anderson canon. corr. LM statistic – – 1.8e + 06*** 5.3e + 05*** 5.3e + 05***
Cragg-Donald Wald F – – 3.9e + 06*** 6.4e + 05*** 6.4e + 05***
industry FE YES YES YES YES YES
year FE YES YES NO YES YES
city FE YES YES NO NO YES
industry-city YES YES NO NO NO
city-year NO YES NO NO NO
N 3,162,788 3,162,788 3,162,929 3,162,929 3,162,929

this paper. The results are shown in column (1) of Table 5. For the of artificial intelligence can help improve corporate environmental
problem mentioned above that the robot inventory and flow sam- performance” has not changed. In addition, in order to eliminate
ples before 2005 are 0, the previous data will also be eliminated. the endogeneity problems such as variable simultaneity and two-
The results in column (2) of Table 5 are obtained. As for the data way causality as much as possible, this paper adopts the instru-
at the enterprise level, the economic level of China’s four munici- mental variable regression method. Based on the correlation and
palities is relatively high. These cities have a larger urban scale exclusivity conditions of the instrumental variables, this paper
and a larger number of enterprises. The higher development level introduces the stock of industrial robots in the same industry in
of the four municipalities also leads to a higher probability of AI countries with the most similar per capita GDP levels as the instru-
exposure for enterprises in their fields, which will interfere with mental variable (Acemoglu & Restrepo, 2020). Since China is the
the regression results of the overall data. Taking these factors into second largest economy in the world after the United States, this
account, samples from the four municipalities directly under the paper selects the robot inventory data of the United States as the
Central government are removed and then returned. The results instrumental variable. As a developed country, the United States’
obtained are shown in column (3) of Table 5. In addition, the quad- robot stock can reflect the world’s frontier level of artificial intelli-
ratic term is also investigated in this paper, as shown in column (4) gence in the industry and meet the conditions for relevance. In
of Table 5. The regression coefficient of the quadratic term is not addition, the use of robots in the United States has nothing to do
significant, so it can be seen that the linear correlation between with the degree of robot application of Chinese enterprises, so it
the promotion effect of AI application on the improvement of meets the conditions of exclusivity. The regression results of
enterprise environmental performance is correct. To sum up, the instrumental variables are shown in columns (3) - (5) of Table 6.
results obtained by regression of replacement samples are always The unidentifiable test statistics all significantly reject the uniden-
significantly negative, which confirms the research direction of this tifiable hypothesis. The Cragg Donald Wald F statistics are all
paper again. greater than the 10 % critical value of the Stock-Yogo test, rejecting
the null hypothesis of weak instrumental variables. This shows
4.3. Endogeneity test that the instrumental variables are reasonable and reliable.

In order to eliminate endogeneity problems such as missing 5. Further analysis

variables, this paper further controls the fixed effect on the basis
of the benchmark regression model. The specific estimated results 5.1. Mechanism test
are shown in columns (1) - (2) of Table 6. Among them, column (1)
of Table 6 controls the industry-city effect, and column (2) controls In this part, new intermediate variables are selected to further
the city-year effect on the basis of column (1). The regression study the action mechanism proposed by the research hypothesis
results are still significant, and the core conclusion ‘‘the application of this paper. The research hypothesis mentions that the applica-
191
Y. Shang, S. Zhou, D. Zhuang et al. Gondwana Research 131 (2024) 181–195

Table 7
Mechanism of action test results.

Technological innovation effect Labor substitution effect

(1) (2) (3) (4) (5) (6)
lnauthor lnapply lnpc_author1 lnpc_author2 lnsos2 lnemp2
lnintell 0.0056*** 0.0065*** 0.0005*** 0.0006*** 0.0035*** 0.0112***
(0.0001) (0.0002) (0.0000) (0.0000) (0.0004) (0.0001)
Observations 3,162,929 3,162,929 3,124,665 3,162,929 313,633 3,162,929
Adjusted R2 0.1098 0.1246 0.0105 0.0110 0.2386 0.9159
Control variable YES YES YES YES YES YES
Industry FE YES YES YES YES YES YES
Year FE YES YES YES YES YES YES
City FE YES YES YES YES YES YES

tion of AI mainly promotes the emission reduction of industrial will reduce the cost of enterprises, so it is necessary for enterprises
enterprises by improving the level of innovation and the artificial to invest in AI.
substitution effect. Therefore, this paper mainly discusses the pro-
moting role of AI application from the perspectives of technological
5.2. Analysis of heterogeneity
innovation and substitution effect. Therefore, the above mecha-
nism is tested empirically3.
(1) Regional heterogeneity. The impact of AI on enterprise pol-
(1) Technological innovation effect. This paper replaces the
lution reduction will vary due to different regions. Therefore, this
explained variable in Equation (1) with the intermediate variable
paper refers to the population size of municipal district (lnsize_y1),
representing technological innovation. And keep the meaning of
the population size of the whole city (lnsize_x1) and the north–
other variables consistent for regression. The results are shown
south city indicators to interact with AI (Zhao et al., 2023a). As
in columns (1) - (5) of Table 7. The intermediate variables used
shown in columns (1) - (2) of Table 8, the interaction terms
here to represent technological innovation include the total num-
between the population size of municipal district and the popula-
ber of patents granted (author), the total number of patent applica-
tion size of the whole city and AI are significantly negative. It
tions (apply), the number of patents cited (author1), the number of
shows that the larger the urban population size is, the more obvi-
patents cited (author2) and the index of waste gas treatment
ous the effect of introducing AI to reduce the level of pollution is.
capacity (sos2). The data of innovation level in this paper come
column (3) of Table 8 shows the regression test of the interaction
from the patent data of industrial enterprises from 2001 to 2014.
term between the classification dummy variable of North-South
This paper matches the industrial enterprise database with the
cities (1 for northern cities, 0 for southern cities) and AI. The results
patent data of the State Intellectual Property Office according to
show that the interaction term of northern cities is significantly
the enterprise name, year and organization code, and obtains rele-
positive. That is, the introduction of AI in northern cities does not
vant patent information. The index of exhaust gas treatment capac-
have obvious benefits on emission reduction, or even reduces such
ity in this paper is obtained by matching China’s industrial
benefits. In contrast, the introduction of AI in southern cities will
enterprise pollution database from 2001 to 2014 with the indus-
enhance emission reduction benefits.
trial enterprise pollution emission data of the National Bureau of
(2) Heterogeneity by industry. The indicators of technical
Statistics. Theoretically, the improvement of AI is conducive to
exchange intensity (contact), non-technical exchange intensity
the improvement of enterprise production efficiency and cost sav-
(discuss) and maturity (mature), which characterize different
ing, so as to increase the investment in technological innovation.
industries, are introduced into the benchmark regression (Zhao
On the other hand, after the improvement of AI, it is more conve-
et al., 2023b). And its interaction term with AI is used as a regres-
nient and fast for enterprises to collect information and process
data. This also contributes to the improvement of enterprise inno-
vation capability. It is also easy to see from the regression results Table 8
that the regression coefficient of AI level on technological innova- Heterogeneity test results: regional heterogeneity.
tion is always significant. Therefore, Hypothesis 1 is verified. Heterogeneity of city size North-south urban
(2) Artificial substitution effect. In this paper, the variable is heterogeneity
replaced by the indicator of the number of employees in the indus- (1) (2) (3)
try (emp2), as shown in column (6) of Table 7. The regression
lnen lnen lnen
results show that the improvement of AI level will lead to a decline lnintell 0.0046*** 0.0047*** 0.0047***
in the number of employees in the industry. It can be seen that AI (0.0006) (0.0006) (0.0006)
has a substitution effect on labor, and Hypothesis 2 is verified. It is lnintell  lnsize_y1 0.0009***
not difficult to understand that the automation technology devel- (0.0003)
lnsize_y1 0.0026
oped by AI has obvious advantages over ordinary labor in terms
(0.0038)
of production efficiency. Therefore, the bargaining power of ordi- lnintell  lnsize_x1 0.0013***
nary labor in the job market will be reduced, and AI will replace (0.0004)
ordinary labor in basic jobs. The use of AI instead of ordinary labor lnsize_x1 0.0235***
(0.0050)
lnintell  north 0.0138***
(0.0005)
3
Although the intermediary effect model has been widely used, it has some defects Observations 3,162,929 3,162,929 3,162,929
and has been widely criticized by economists recently. While verifying the influence Adjusted R2 0.1378 0.1378 0.1380
of the core variable on the intermediate factor, there is an endogenous relationship Control variable YES YES YES
between the intermediate factor and the final explained variable. Therefore, the Industry FE YES YES YES
reliability of the second regression result of the intermediate variable is doubtful. Year FE YES YES YES
Therefore, this paper focuses on examining the influence of core variables on City FE YES YES YES
intermediate factors.

192
Y. Shang, S. Zhou, D. Zhuang et al. Gondwana Research 131 (2024) 181–195

Table 9
Heterogeneity test results: industry heterogeneity, scale heterogeneity.

Heterogeneity of industries Heterogeneity of size

(1) (2) (3) (4)
lnen lnen lnen lnen
lnintell 0.0049*** 0.0073*** 0.0082*** 0.0047***
(0.0006) (0.0006) (0.0006) (0.0006)
lnintell  contact 0.0015***
(0.0006)
contact 0.0382***
(0.0037)
lnintell  discuss 0.0280***
(0.0006)
discuss 0.3380***
(0.0049)
lnintell  lncyrs 0.0140***
(0.0002)
lnintell  lnage 0.0093***
(0.0003)
Observations 3,162,929 3,162,929 3,162,929 3,162,929
Adjusted R2 0.1378 0.1395 0.1388 0.1380
Control variable YES YES YES YES
Industry FE YES YES YES YES
Year FE YES YES YES YES
City FE YES YES YES YES

sion indicator. As shown in column (1) of Table 9, the regression introduction in non-technical exchange intensive enterprises is
result of the interaction term between technical exchange inten- more obvious than that in technical exchange intensive
sity and AI is significantly positive. This shows that non- enterprises.
technology exchange intensive enterprises have a better emission In the era of rapid development of the Internet and science and
reduction effect than technology exchange intensive enterprises technology, the popularization of AI is also more extensive. Intelli-
after introducing AI. In this paper, the technical exchange intensive gent machines have gradually penetrated into almost all walks of
and non-technical exchange intensive enterprises are distin- life. In addition to the efficiency benefits of AI, there are also many
guished according to the ranking of the exchange intensity of the negative effects. So how to correctly introduce AI to improve the
manufacturing industry. Those ranked after the 15th are classified pollution level of enterprises and bring better benefits has become
as non-technical exchange intensive enterprises. There are many a topic we need to study now. Based on this, this paper puts for-
kinds of technical communication. This paper chooses ‘‘contact ward the following policy implications for the rational use of enter-
with others” and ‘‘face-to-face discussion” in the degree of work prise AI in dealing with enterprise pollution emission reduction:
to measure the intensity of technical communication. First, in the process of upgrading and accelerating production
(3) Heterogeneity of firm size. This paper introduces the vari- development, the introduction of AI is an important way to achieve
able total number of employees in the industry and the enterprise the long-term goal of green and low-carbon development. The fun-
age indicator that can represent the size of the enterprise. The damental purpose of this paper is to explore the benefits of AI for
interaction term with AI is regressed, and the results are shown enterprise pollution reduction. And this paper guides enterprises to
in Table 9. From the regression results in columns (4) - (5) of pay attention to the rational use of AI. Enterprises in various indus-
Table 9, it can be seen that the number of employees in the indus- tries should transform the existing unreasonable configuration and
try and enterprise age have a positive effect on enterprise pollution use of AI. In this way, the efficient development of enterprises can
reduction. This means that the efficient use of AI has a better effect be taken into account, and the environmental performance of
on emission reduction for enterprises with a larger labor force or a enterprises can be guaranteed. In addition, government depart-
larger age. ments should also give certain rewards to enterprises that intro-
duce AI. Through this reward mechanism to promote the
intelligent development of enterprises to achieve energy conserva-
6. Conclusions and policy recommendations tion, emission reduction and sustainable development of enter-
prises. The social sector should also vigorously publicize the
Green, low-carbon, energy conservation and emission reduction impact of AI on environmental performance and improve the social
are hot topics of global concern. Protecting the healthy and sus- awareness that AI can save energy and reduce emissions.
tainable development of the ecological environment is the goal Second, the impact of the extensive application of AI on the
that all countries should adhere to. This paper examines the pollu- environmental performance of enterprises depends on the means
tion reduction effect brought by AI from a micro perspective. In of technological innovation ability and labor substitution effect. If
addition, this paper draws the degree of the use of AI in enterprises AI plays a better role in technological innovation ability and labor
in various industries in recent years. Moreover, the Chinese indus- substitution effect, the pollution reduction effect of enterprises will
trial enterprise database and pollution data are regressed. The be more considerable. Specifically, enterprises introduce AI to
study found that after 2005, the number of robots in Chinese enter- upgrade the original technology innovatively. This will allow the
prises in various industries increased significantly, especially in the original machinery to have a higher precision and enable the iden-
automobile industry. This paper also finds that AI promotes enter- tification and filtration of harmful pollutants in the production pro-
prise pollution reduction through technological innovation effect cess. AI can also be used to innovate equipment for handling
and labor substitution effect. For enterprises in big cities, enter- pollutants. More effective treatment of pollution emissions is also
prises in southern cities and enterprises with larger scale, AI has the key to energy conservation and emission reduction. On the
a better effect on enterprise emission reduction. The effect of AI other hand, improving the level of intelligence of enterprises can
193
Y. Shang, S. Zhou, D. Zhuang et al. Gondwana Research 131 (2024) 181–195

replace the basic work of labor. Through the high intensity and Acknowledgements
high precision work of the machine in the production of products
while minimizing the manufacture of pollutants. Reducing the This work was supported by the National Natural Science Foun-
generation of pollutants at their source is also an essential part dation of China (Grant No. 72304084), the Fundamental Research
of emission reduction. Funds for the Central Universities of China (Grant Nos.
Third, different strategies should be developed for different JZ2023HGQA0083; JZ2023HGTA0210), Lanzhou University 2023
enterprises according to the differentiation factors. From the Graduate Student ‘‘Innovation Star” (Grant No. 2023CXZX-019)
heterogeneity analysis in this paper, it is not difficult to see that and The Major Project of Anhui University Philosophy and Social
AI’s promotion of enterprise pollution reduction will have different Science Research Project (2023AH040229).
effects due to the city size, enterprise size, enterprise age and the
location of the city where the enterprise is located. There is also
the special case of firms intensive in technical exchange and those
Appendix A. Supplementary data
intensive in non-technical exchange. The detection data of micro
enterprises in this paper mainly come from industrial enterprises.
Supplementary data to this article can be found online at
It can be seen from the inspection process that technology-based
https://doi.org/10.1016/j.gr.2024.02.012.
enterprises with small pollution emissions sometimes fail to
reduce their pollution emissions by using AI. It could even backfire.
Therefore, enterprises should make decisions on how to introduce
References
AI to reduce their emission levels according to their own
characteristics. Acemoglu, D., Restrepo, P., 2018. The race between man and machine: Implications
Using Chinese industrial enterprises as a sample, this paper of technology for growth, factor shares, and employment. Am. Econ. Rev. 108
innovatively examines the role of artificial intelligence in promot- (6), 1488–1542. https://doi.org/10.1257/aer.20160696.
Acemoglu, D., Restrepo, P., 2020. Robots and jobs: Evidence from us labor markets. J.
ing corporate environmental performance. As the world’s second Polit. Econ. 128. https://doi.org/10.1086/705716.
largest economy, China is one of the important trading entities in Acs, Z., Braunerhjelm, P., Audretsch, D., Carlson, B., 2009. The knowledge spillover
the world economic market. With the deepening of the globaliza- theory of entrepreneurship. Small Bus. Econ. 32 (1), 15–30. https://doi.org/
10.1007/s11187-008-9157-3.
tion of China’s financial market, Chinese enterprises continue to Arntz, M., Gregory, T., Zierahn, U. (2016). The risk of automation for jobs in OECD
trade with enterprises in other countries. In this process, we learn countries: a comparative analysis. Oecd Social Employment & Migration
from each other and narrow the gap between enterprises in the Working Papers. 189, 34. 10.1787/5jlz9h56dvq7-en.
Autor, D., Dorn, D., 2012. The growth of low skill service jobs and the polarization of
same industry. Therefore, this paper studies from the micro enter-
the us labor market. Am. Econ. Rev. 103 (5), 1553–1597. https://doi.org/
prise level, and its research results are universal in the world. The 10.1257/aer.103.5.1553.
research of this paper provides the basis for how to use artificial Autor, D.H., Salomons, A., 2018. Is Automation Labor-Displacing? Productivity
intelligence to reduce pollution and emission in various countries. Growth, Employment, and the Labor Share. Social Science Electronic Publishing
https://www.jstor.org/stable/26506212.
Although this paper has carried out some exploration and anal- Autor, D.H., Katz, L.F., Kearney, M.S., 2006. The polarization of the U.S. Labor Market.
ysis, other AI measurement indicators have not been analyzed in National Bureau of Economic Research. Am. Econ. Rev. 96 (02), 189–194.
detail due to data availability and other problems. In the future, https://doi.org/10.1257/000282806777212620.
Bessen, J., 2019. Automation and jobs: When technology boosts employment. Econ.
with the disclosure of more and more detailed relevant data, this Policy 34 (100), 589–626. https://doi.org/10.1093/epolic/eiaa001.
paper will further measure AI indicators. In the future, we will fur- Brandt, L., Biesebroeck, J.V., Zhang, Y., 2012. Creative accounting or creative
ther study its impact on corporate environmental performance destruction? Firm-level productivity growth in Chinese manufacturing. J. Dev.
Econ. 97, 339–351. https://doi.org/10.1016/j.jdeveco.2011.02.002.
based on more accurate artificial intelligence indicators. At the Bughin, J., Seong, J., Manyika, J., Chui, M., Joshi, R. (2018). Notes from the AI Frontier:
same time, it can also study the path of the impact of artificial Modeling the Impact of AI on the World Economy. McKinsey Global Institute
intelligence on the enterprise environment and conduct a more Discussion Paper.
Cheng, H., Jia, R., Li, D., Li, H., 2019. The rise of robots in China. J. Econ. Perspect. 33
comprehensive analysis. In addition, this paper mainly describes (2), 71–88. https://doi.org/10.1257/jep.33.2.71.
how artificial intelligence improves corporate environmental per- Cortés, U., Sànchez-Marrè, M., I. R-Roda, et al. (1999). Binding environmental
formance through innovation effect and artificial substitution sciences and artificial intelligence on ecai’98. Ai. Commun. 12(4), 261-265.
doi:10.1142/9789812561749_0001.
effect. But in fact, there may still be unexplained paths that need
Dagestani, A.A., Shang, Y., Schneider, N., Cifuentes-Faura, J., Zhao, X., 2023. Porter in
to be explored further. China: A quasi-experimental view of market-based environmental regulation
effects on firm performance. Energ. Econ. 126,. https://doi.org/10.1016/j.
eneco.2023.106966 106966.
Dauth, W., Findeisen, S., J Südekum, Wner, N. (2017). German robots - the impact of
Credit author statement industrial robots on workers. IAB-Discussion Paper 201730. https://cepr.org/
publications/DP12306.
Yuping Shang: Conceptualization, Investigation, Data curation, Endo, K., 2020. Corporate governance beyond the shareholder–stakeholder
dichotomy: Lessons from Japanese corporations’ environmental performance.
Writing original draft; Writing-review & editing. Silu Zhou: Con- Bus. Strateg. Environ. 29 (4), 1625–1633. https://doi.org/10.1002/bse.2457.
ceptualization, Investigation, Data curation, Writing original draft, Fossen, F.M., Sorgner, A., 2022. New digital technologies and heterogeneous wage
Writing-review & editing. DelinZhuanga: Conceptualization, and employment dynamics in the United States: Evidence from individual-level
data. Technol. Forecast. Soc. 175,. https://doi.org/10.1016/j.
Investigation, Data curation, Writing original draft, Writing- techfore.2021.121381 121381.
_
review & editing. Justyna Zywiołek: Conceptualization, Investiga- Frey, C.B., Osborne, M.A., 2017. The future of employment: how susceptible are jobs
to computerization? Technol. Forecast. Soc. Change. 114, 254–280. https://doi.
tion, Data curation; Writing original draft, Writing-review & edit-
org/10.1016/j.techfore.2016.08.019.
ing. Hasan Dincerc: Data curation, Writing original draft, Furman, J., Seamans, R., 2019. Ai and the economy. Innov. Policy. Econ. 19 (1), 161–
Writing-review & editing. 191. https://doi.org/10.1086/699936.
Gal, P., Nicoletti, G., Theodore, R., Stéphane, S., Christina, T. (2019). Digitalization
and Productivity: In Search of the Holy Grail - Firm-level Empirical Evidence
from European Countries. Int Prod Monit. 63. 10.1787/5080f4b6-en.
Declaration of Competing Interest Galindo-Martín, M., Castaño-Martínez, M., Méndez-Picazo, M., 2019. Digital
transformation, digital dividends and entrepreneurship: A quantitative
The authors declare that they have no known competing finan- analysis. J. Bus. Res. 101, 522–527. https://doi.org/10.1016/j.
jbusres.2018.12.014.
cial interests or personal relationships that could have appeared Gasteige, E., Prettner, K. (2017). A note on automation, stagnation, and the
to influence the work reported in this paper. implications of a robot tax. Discussion Papers. 10.17169/REFUBIUM-25260.

194
Y. Shang, S. Zhou, D. Zhuang et al. Gondwana Research 131 (2024) 181–195

Geng, H., He, L.-Y., Lin, X., 2022. Robot adoption and energy performance: Evidence Shapiro, J.S., Walker, R., 2018. Why is pollution from us manufacturing declining?
from Chinese industrial firms. Energy. Econ. 107, (6). https://doi.org/10.1016/j. The roles of environmental regulation, productivity, and trade. Am. Econ. Rev.
eneco.2022.105837 105837. 108 (12), 3814–3854. https://doi.org/10.1257/aer.20151272.
Goldfarb, A., Tucker, C., 2019. Digital economics. J Econ Lit. 57 (1), 3–43. https://doi. Siebel, T. (2019). Digital Transformation: Survive and Thrive in an Era of Mass
org/10.1257/jel.20171452. Extinction. Rosettabook.
Graetz, G., Michaels, G., 2018. Robots at work. Rev. Econ. Stat. 100 (5), 753–768. Song, M., Peng, L., Shang, Y., Zhao, X., 2022. Green technology progress and total
https://doi.org/10.1162/rest_a_00754. factor productivity of resource-based enterprises: A perspective of technical
Gregory, T., Salomons, A., Zierahn, U., 2022. Racing with or against the machine? compensation of environmental regulation. Technol. Forecast. Soc. Change.
Evidence from Europe. J. Eur. Econ. Assoc. 20 (2), 869–906. https://doi.org/ 174,. https://doi.org/10.1016/j.techfore.2021.121276 121276.
10.1093/jeea/jvab040. Xu, Y., Yuan, L., Lee, H., Baire, S., Nakonieczny, J., et al., 2023. Fintech development
Henriques, I., Sadorsky, P., 1996. The determinants of an environmentally and firm technological innovation efficiency: Empirical findings in China. IEEE
responsive firm: An empirical approach. J. Environ. Econ. Manag. 30 (3), 381– T. Eng. Manage. 1–11. https://doi.org/10.1109/TEM.2023.3239499.
395. https://doi.org/10.1006/jeem.1996.0026. Ye, Z., Yang, J., Zhong, N., Tu, X., Jia, J., Wang, J., 2020. Tackling environmental
Herrendorf, B., Rogerson, R., Valentinyi, Á., 2013. Two perspectives on preferences challenges in pollution controls using artificial intelligence: a review. Sci. Total.
and structural transformation. Am. Econ. Rev. 103 (7), 2752–2789. https://doi. Environ. 699, (1). https://doi.org/10.1016/j.scitotenv.2019.134279 134279.
org/10.1257/aer.103.7.2752. Zhang, C., Liu, Q., Ge, G., Hao, Y., Hao, H., 2021. The impact of government
Hoang, T., Przychodzen, W., Przychodzen, J., Segbotangni, E.A., 2021. Environmental intervention on corporate environmental performance: Evidence from China’s
transparency and performance: Does the corporate governance matter? Ecol. national civilized city award. Financ. Res. Lett. 39, (3). https://doi.org/10.1016/j.
Indic. 10, (2–3). https://doi.org/10.1016/j.indic.2021.100123 100123. frl.2020.101624 101624.
Jansson, T., Karabulut, Y., 2018. Do robots increase wealth dispersion? SSRN Zhang, W., Luo, Q., Liu, S., 2022. Is government regulation a push for corporate
Electron. J. https://doi.org/10.2139/ssrn.3229980. environmental performance? Evidence from China. Econ. Anal. Policy 74 (1),
Klump, Ｒ., Mcadam, P., Willman, A., 2007. Factor substitution and factor– 105–121. https://doi.org/10.1016/j.eap.2022.01.018.
augmenting technical progress in the United States: A normalized supply-side Zhang, L., Ren, S., Chen, X., 2020. CEO hubris and firm pollution: State and market
system approach. Rev. Econ. Stat. 89 (1), 183–192. https://doi.org/10.1162/ contingencies in a transitional economy. J. Bus. Ethics. 161, 459–478. https://
rest.89.1.183. doi.org/10.1007/s10551-018-3987-y.
Kunst, D. (2020). Deskilling among manufacturing production workers. Tinbergen Zhao, P., Gao, Y., Sun, X., 2022. How does artificial intelligence affect green economic
Institute Discussion Papers. https://papers.tinbergen.nl/19050.pdf. growth? —Evidence from China. Sci. Total Environ. 834,. https://doi.org/
Lewis, B.W., Walls, J.L., Dowell, G.W.S., 2014. Difference in degrees: CEO 10.1016/j.scitotenv.2022.155306 155306.
characteristics and firm environmental disclosure. Strategic. Manage. J. 35 (5), Zhao, X., Shang, Y., Magazzino, C., Madaleno, M., Mallek, S., 2023a. Multi-step
712–722. https://doi.org/10.1002/smj.2127. impacts of environmental regulations on green economic growth: Evidence in
Michaels, G., Natraj, A., Reenen, J.V., 2014. Has ICT polarized skill demand? Evidence the lens of natural resource dependence. Resour. Policy. 85,. https://doi.org/
from eleven countries over twenty-five years. Rev. Econ. Stat. 96 (1), 60–77. 10.1016/j.resourpol.2023.103919 103919.
https://doi.org/10.1162/REST_a_00366. Zhao, X., Xu, Y., Vasa, L., Shahzad, U., 2023b. Entrepreneurial ecosystem and urban
Sachs, J.D., Kotlikoff, L.J. (2012). Smart machines and long-term misery. NBER innovation: Contextual findings in the lens of sustainable development from
Working Papers. 18629. https://www.nber.org/papers/w18629. China. Technol. Forecast. Soc. 191,. https://doi.org/10.1016/j.
Shang, Y., Xu, J., Zhao, X., 2022. Urban intensive land use and enterprise emission techfore.2023.122526 122526.
reduction: New micro-evidence from China towards COP26 targets. Resour.
Policy. 79,. https://doi.org/10.1016/j.resourpol.2022.103158 103158.

195