AI对TFP的影响

Economic Change and Restructuring (2023) 56:1113–1146
https://doi.org/10.1007/s10644-022-09467-4
The impact of artificial intelligence on total factor

productivity: empirical evidence from China’s
manufacturing enterprises
Ke‑Liang Wang1 · Ting‑Ting Sun1 · Ru‑Yu Xu1
Received: 1 September 2022 / Accepted: 24 November 2022 / Published online: 5 December 2022
© The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature
2022
Abstract
Using the panel data of 938 listed manufacturing companies in China from 2011
to 2020, this paper scientifically examines the impact of artificial intelligence (AI)
on total factor productivity (TFP) of China’s manufacturing enterprises by using
the fixed effect model, mediating effect model and difference-in-differences model.
The results show that AI can significantly improve the TFP of China’s manufactur-
ing enterprises, as confirmed by a series of robustness tests. Technological inno-
vation, human capital optimization and market matching improvement have proved
to be three important channels for AI to affect the TFP of China’s manufacturing
enterprises. The impact of AI on TFP varies greatly among China’s manufacturing
enterprises in different geographical locations, industry characteristics, ownership
and life cycle stages. The findings of this paper can provide theoretical insights and
empirical evidence at the micro enterprise level for policymakers to give full play to
the role of AI in promoting the high-quality development of China’s manufacturing
industry.
Keywords Total factor productivity (TFP) · Artificial intelligence (AI) ·

Manufacturing enterprises · Transmission mechanism · Heterogeneity ·
Exogenous policy shock
1 Introduction
Artificial intelligence (AI), as a critical technique guiding the fourth scientific and
technological revolution (Dwivedi et al., 2021), has continuously permeated various
aspects of people’s lives and enterprises’ operations with its strong permeability and
* Ke‑Liang Wang
[email protected]
1
School of Economics, Ocean University of China, No. 238, Songling Road, Qingdao 266011,
Shandong, China
13
Vol.:(0123456789)
1114 Economic Change and Restructuring (2023) 56:1113–1146
universality (Trajtenberg, 2018), posing an extremely profound impact on economic

growth and social progress worldwide (Aghion et al., 2018; Graetz and Michaels,
2018). According to Capgemini’s Research Report (2020), over half of the top man-
ufacturing companies in Europe are applying AI, with Japan (30%) and the U.S.
(28%) closely behind. Accenture’s Report (2017) predicted that AI applications
will boost the value of manufacturing growth by nearly $4 trillion by 2035, with an
annual growth rate of 4.4%. This indicates that AI will become an important driver
for global manufacturing development in the future. Under this background, major
economies in the world have successively issued strategies to vigorously develop
AI. For instance, the U.S. formulated the “Robot Plan” in 2013, Japan issued the
“New Robot Strategy” in 2015, and European Union promulgated the “Declaration
on Strengthening AI Cooperation” in 2018. The comprehensive integration of AI
with manufacturing has resulted in a radical transformation in the technical para-
digm. However, its application within the enterprise is sophisticated, involving not
just changes to production practices but also modifications to management styles
and mindsets, which pose significant challenges for enterprises (Wen et al., 2022b).
For developing nations and emerging economies, can AI help manufacturing crack
the low-end lock and achieve transformation and upgrading?
As the largest developing country in the world, China’s AI industry has devel-
oped sharply in recent years. According to the China Internet Development Report
(2021), by 2020, the scale of China’s AI industry has already amounted to 303.1 bil-
lion yuan, with 1454 enterprises, ranking second in the world. PwC’s Report (2017)
predicted that AI will raise China’s GDP by 26% in 2030, rendering it the country
that would gain the most from AI. As an emerging economy, China’s manufactur-
ing has increased significantly in both size and growth rate (Brandt et al., 2012).
However, in recent years, China’s manufacturing industry has long been situated
at the bottom of the global value chain. Additionally, as the demographic dividend
gradually disappears, labor costs have risen dramatically. This inevitably leads to the
decline of the comparative advantage of China’s manufacturing industry (Cai, 2012).
Facing the dual pressures from home and abroad, it is urgent for China’s manufac-
turing enterprises to master AI technology and strive to achieve transformation and
upgrading, to regain the competitive advantage in international competition. Thus, it
is critical for developing countries and emerging economies to probe into the asso-
ciation between AI and TFP of manufacturing enterprises based on China’s empiri-
cal experience.
Recently, with the maturity of theories and technologies of AI, it has been exten-
sively applied in numerous areas. It has not only achieved great breakthrough in
high-tech fields such as finance, medical treatment and unmanned driving (Xiao and
Liu, 2019), but also brought dramatic changes to the operation mode in traditional
industries (Ahmad et al., 2022). Specifically, enterprises can adopt AI to digitize
all production links and then realize intelligent and modular management, which
can essentially improve their production efficiency (Poole and Mackworth, 2010).
In addition, AI can help enterprises to better allocate production factors by opti-
mizing production process (Dolgui et al., 2019), thereby significantly optimizing
enterprises’ resource allocation efficiency. Moreover, through the AI platform, more
efficient communication can be achieved between enterprises and consumers, which
13
Economic Change and Restructuring (2023) 56:1113–1146 1115
can greatly reduce their transaction costs, substantially improving the efficiency of
resource utilization (Chen et al., 2020). Based on above analysis, it is demonstrated
that AI has drastically changed the traditional operation model of enterprises, sig-
nificantly improving their production efficiency and resource allocation capability
(Acemoglu and Restrepo, 2018; Liu et al., 2020), which are all the embodiment of
the improvement of enterprise TFP.
In this context, a question will naturally be raised. Can AI promote enterprise
TFP? Particularly, how does AI affect TFP of Chinese manufacturing enterprises?
This needs to be further verified theoretically and empirically. Hence, with panel
data of 938 manufacturing listed companies in China from 2011–2020, this paper
methodically investigates the effect of AI on enterprise TFP, and its transmission
mechanism and heterogeneities are further analyzed. We believe that the findings of
this paper have important theoretical and practical significances for policymakers to
formulate targeted measures to realize the transformation and upgrading, as well as
high-quality development of manufacturing industry in China.
With the rapid expansion of AI in China, this paper investigates the impact of
AI on TFP of manufacturing firms. Due to the lack of authoritative measures of AI,
existing literature universally employs industrial robots, output value of information
transmission industry, and AI patents as proxy variables for AI (Huang et al., 2022;
Nguyen and Vo, 2022). However, such proxy variables are mainly macro-level and
do not accurately reflect the characteristics of firms themselves. Thus, in order to
evaluate AI in enterprises, we use text analysis to extract AI keywords from com-
panies’ annual reports. In addition, the contributions of this paper to the existing
literature are as follows.
Firstly, compared with previous studies mainly focusing on regional and indus-
trial levels as well as micro-firm level in developed countries, this paper is intended
to examine the relationship between AI and TFP at the micro-enterprise level in
China, and the related research still remains scarce.
Secondly, this article offers empirical evidence regarding the relationship between
AI and enterprises’ TFP. We use a two-way fixed effects model to demonstrate that
AI significantly enhances enterprises’ TFP. Moreover, utilizing the mediating effect
model, this article discusses the impact of AI on the TFP of manufacturing enter-
prises from the perspectives of technological innovation, human capital optimiza-
tion and market matching amelioration, which provides theoretical and practical
evidence for further leveraging the role of AI in contributing high-quality develop-
ment of China’s manufacturing industry. Furthermore, we examined the impact of
AI on the enterprise’s TFP from the perspective of policy shocks by using the DID
method, with the Smart Manufacturing Demonstration Project as a quasi-natural
experiment, which verified the feasibility of AI policies.
Finally, the paper presents empirical evidence for policymaking in developing
countries and emerging economies from the perspective of regional, industry, and
firm heterogeneity. Results indicate that AI contributes more to TFP in the eastern
region, technology-intensive industries, state-owned enterprises, and growing enter-
prises. To capitalize on the prominent role of AI in manufacturing firms, authorities
in emerging economies should use AI in accordance with the features of different
enterprises.
13
Fig.1 Research framework
This paper explores the connections between AI and enterprises’ TFP, offering not
only significant empirical evidence for the intelligent transformation of enterprises in
developing countries and emerging economies, but also a novel policy perspective to
cope with the information revolution in developed countries.
The rest of this paper is arranged as follows. Section 2 is literature review; Sect. 3
presents the theoretical mechanism and research hypotheses; Research design, vari-
ables selection and data description are introduced in Sect. 4; Sect. 5 offers empirical
results and discussions; Sect. 6 concludes this study and provides related policy recom-
mendations. The research framework of this paper is presented in Fig. 1.
13
2 Literature Review
As the main participants in the economic activities of a country or a region,

enterprises’ TFP is highly correlated with the comprehensive competitiveness
of an economy (Hsieh and Klenow, 2010). As such, how to enhance the TFP of
enterprises has always been a hot issue concerned by the academic community.
The related research of TFP is mainly conducted from two aspects: measurement
methods and influencing factors.
The methods for measuring enterprise TFP can be divided into three catego-
ries: the parametric methods, such as Solow residual method, Stochastic Fron-
tier Analysis (SFA) and OLS, the nonparametric methods represented by a series
of productivity indexes derived from Data Envelopment Analysis (DEA), and
the semiparametric methods, including Olley-Pakes (OP) and Levinsohn-Petrin
(LP), Ackerberg–Caves–Frazer (ACF) and Wooldridge–Levinsohn–Petrin (WLP).
Among them, the parametric methods tend to overestimate enterprise TFP caused
by endogenous problems and selective error (Van Biesebroeck, 2007). The non-
parametric method takes several production units as the technical effective points
to construct the production frontier without considering the randomness of sam-
ples, so it fails to reflect the heterogeneity of enterprise TFP (Simar and Wil-
son, 2000). By contrast, the semi-parametric method can effectively overcome the
defects of the above two methods. Hence, it has become the mainstream method
for measuring enterprise TFP (Chen et al., 2021; Zheng et al., 2022). At present,
the commonly used semi-parametric methods for enterprise TFP estimation are
the OP and the LP. Among them, the OP proposed by Olley and Pakes (1992)
takes the amount of enterprise investment as the proxy of productivity shock.
However, in this context, enterprises with current investment of zero will be
excluded from the sample, thereby resulting in a large number of enterprise sam-
ples missed. By contrast, the LP proposed by Levinsohn and Petrin (2003) adopts
intermediate inputs as the proxy of productivity shock, which can effectively
address the defects of OP. However, OP and LP methods presume that enter-
prises can instantly adapt inputs costlessly and instantly when facing productivity
shocks. As a result, Wooldridge (2009) presented a GMM-based one-step estima-
tion method (WLP method) by modifying OP and LP methods. The advantage
of this approach is to override the potential identification problem in the estima-
tion of the first step in the revised estimation method presented by Ackerberg
et al. (2006), as well as to obtain robust standard errors under the consideration
of serial correlation and heteroskedasticity, thus better reflecting the changes of
enterprise TFP. Therefore, this study will use the WLP method to measure firm
TFP, and the TFP measured by the OP and LP methods will be used as explana-
tory variables in robustness tests.
Regarding the influencing factors of enterprise TFP, the existing literature
mainly expounds on this topic from two aspects: the internal characteristics and
the external environment. From the perspective of enterprise internal factors,
Guan and Cheng (2020) examined the impact of product complexity on enterprise
TFP. Zhang and Liu (2017) discussed the relationship between corporate leverage
13
and TFP. In addition, some scholars analyzed the impact of R&D investment,
knowledge interaction and financial performance on enterprise TFP (Antonelli
and Scellato, 2013; Hall and Mairesse, 1995; Machek, 2014). With regard to the
macro environmental factors, previous studies mainly focused on the impact of
environmental regulation and technological progress. Such as, Tang et al. (2020)
regarded China’s “Acid Rain Control Zone and Sulfur Dioxide Control Zone (Two
Control Zone)” policy as a quasi-natural experiment, and analyzed the impact of
command-and-control environmental regulation on TFP of Chinese A-shared
listed companies by using DID method. Zhang et al. (2021b) investigated the
impact of tax incentives on TFP of Chinese mining enterprises. Song et al. (2022)
examined the impact of green technological progress and environmental regula-
tion of TFP of resource-based enterprises. Additionally, some scholars have also
found that import and export trade, government subsidies and digital finance all
play a positive role in improving enterprise TFP (Li et al., 2022b; Tang et al.,
2022; Vogel and Wagner, 2010).
Since the productivity paradox was mooted in 1978, there is considerable litera-
ture probing the impact of information and communication technologies (ICT) on
economic performance and productivity at the country, industry, and firm levels. At
the national level, numerous scholars have interpreted the Solow paradox in light of
lagged effects, measurement errors, and redistribution of profits (Brynjolfsson et al.,
2018; Martínez et al., 2008; Oliner and Sichel, 2000). Cross-country data indicate
that, due to disparities in economic growth and network infrastructure, mobile ICT
has a varied effect on productivity in developing and developed nations (Kim et al.,
2021). At the industry level, Heo and Lee (2019) discovered that a significant link-
age between ICT manufacturing and ICT services has taken shape, which is steadily
strengthening. Del Gaudio et al. (2021) investigated the impact of ICT in the bank-
ing sector and revealed that ICT enhances performance and stability in the banking
sector. Based on data from EU and US industries in 1995–2007, Papaioannou and
Dimelis (2017) came to the conclusion that, after accounting for upstream regula-
tion, ICT had a significant role in boosting the efficiency of manufacturing indus-
tries. And at the firm level, existing studies have observed that ICT contributes to
the productivity of firms in developed countries such as New Zealand, Spain and
Eastern Europe (Sanchez et al., 2006; Skorupinska and Torrent-Sellens, 2017; Tor-
rent-Sellens et al., 2022). Similarly, Luo and Bu (2016) found that ICT significantly
improved firm productivity based on data from firms in 27 emerging economies.
One effective approach for addressing the connection between ICT and an enter-
prise’s TFP is technological innovation (Kijek and Kijek, 2019). Reduced infor-
mation search costs are crucial for elevating TFP for businesses (Li et al., 2022a).
Despite variations between macro and micro level findings, most researchers con-
cluded that ICT investments played a constructive role in improving economic per-
formance and productivity.
ICT investments have given rise to the Internet (Wang et al., 2022a; Wu et al.,
2021). According to majority of studies, the emergence of the Internet has reduced
the cost of international information transfer and sharing, and the knowledge and
technological spillovers it has generated have fueled regional economic growth
(Choi and Yi, 2009; Kenny, 2003; Koutroumpis, 2009). During this process, on the
13
one hand, the Internet and the traditional industrial sector are integrated to fuel the
transformation and upgrading of traditional industries (Li and Du, 2021; Pan et al.,
2022). On the other hand, the Internet development has also spawned new indus-
tries, such as Internet finance (Sharma et al., 2020), industrial Internet (Zhang et al.,
2021a), and energy Internet (Yang et al., 2020). It is noteworthy that with the divi-
sion of labor between industries, their technical and economic bonds are more inti-
mate. Internet development not only directly favors employment in this industry, but
also indirectly other industries through "inter-industry transmission effects"(Wang
et al., 2020). Also, in international trade, Internet development helps to decrease
transaction costs. Countries can become more fully integrated into the global com-
mercial system in such an environment, which has a greater favorable impact on
those nations that are remote from the global market (Gnangnon and Iyer, 2018).
AI is underpinned by the rapid development of ICT and the Internet (Gams et al.,
2019), which is recognized as a driving factor of ICT innovation (Lu et al., 2018).
In recent years, with the rapid development of AI worldwide, academia gradu-
ally began to pay attention to its impact on economy and society (Acemoglu and
Restrepo, 2020; Cheng et al., 2019; Furman and Seamans, 2019; Xie et al., 2021).
For example, Autor (2015) argued that the wide application of AI will not only sig-
nificantly improve production efficiency, but also create new jobs, further increas-
ing labor demand. Based on the industrial data of 17 countries from 1993 to 2007,
Graetz and Michaels (2018) examined the economic impact of AI and found that
the high integration of AI and the real economy can not only greatly improve labor
productivity and increase product added value, but also boost the demand of highly
skilled labor force. Using the provincial panel data of China during 2006–2019, Du
and Lin (2022) researched the impact of industrial robot on regional TFP and found
that there was a U-shaped curve relationship between AI and regional TFP, with
most provinces already showing an upward trend across the inflection point. Liu
et al. (2020) disclosed that AI can obtain the required knowledge from massive data
accurately and quickly, thus accelerating knowledge creation and spillover, which
can help enterprises improve their abilities of learning and absorption, so as to pro-
mote technological innovation. Based on corporate data for Taiwan from 2002 to
2018, Yang (2022) states that AI significantly increases the TFP of enterprises and
optimizes the structure of the workforce. Damioli et al. (2021) discovered that AI
patenting has an additional positive impact on the labor productivity of companies,
controlling for other patenting activities. Peng and Tao (2022) discovered that digital
transformation significantly increased business performance. Meanwhile, the main
avenues for resolving the connection between digital transformation and corporate
success are technological innovation, cost reduction, and efficiency enhancements.
Gaglio et al. (2022) also revealed that digital transformation facilitates firm labor
productivity through technological innovation. Besides, some scholars have also
examined the impact of AI on workforce structure (Nguyen and Vo, 2022), tech-
nological innovation (Wen et al., 2022b), and green economy development (Zhao
et al., 2022). The above studies bring empirical evidence for the beneficial effects of
AI on economic development and social progress.
Existing studies demonstrate that AI may enhance operational effectiveness and
reduce expenses to increase business productivity. On the one hand, AI contributes
13
to the operational efficiency of enterprises. With the advent of the digital economy,
the element of data has become an indispensable element for business operations
(Zhao and Zhou, 2022). AI captures vast amounts of data about products and busi-
nesses and further analyzes and predicts these data (Zeng et al., 2022). Administra-
tors can dynamically adjust inventory and new product design with market data to
make the right production decisions (Haefner et al., 2021), thus alleviating issues of
mismanagement and product stagnation (Schuh et al., 2016). On the other hand, AI
facilitates cost reduction for enterprises. Industrial intelligence can not only com-
plete products efficiently and enhance product quality, but also decrease the demand
for low-skilled labor and reduce the labor cost of enterprises (Huang et al., 2021).
High-skilled workforce and AI are significantly compatible with each other, which
can foster the diffusion of both tacit and explicit knowledge, lowering firms’ inno-
vation costs and ultimately contributing to their TFP (Che and Zhang, 2018). For
instance, Sinopec selected four plants as smart refineries in 2013 to upgrade to smart
manufacturing. To date, these plants have shown a significant rise in economic
advantages, with an increase in automated data collection capacity from 10% to 95
and a 10% increase in labor productivity (ARC, 2018).
To sum up, although a great deal of studies has concentrated on the link between
ICT investment, the Internet, digital transformation, AI and economic development,
technological innovation, energy conservation and environmental protection in
recent years, there are still some limitations that need to be improved. Firstly, most
previous studies analyze the impact of AI on economy and society at the national,
industrial and regional levels, and there is still a lack of related research from the
micro enterprises, especially since few studies pay attention to the impact of AI on
enterprise TFP in developing countries, such as China. Secondly, the mechanisms
and channels of AI affecting enterprise TFP have not been explained in detail in
the existing literature, which is not conducive to the formulation of targeted poli-
cies. Moreover, with the implementation of smart manufacturing pilot, it is worth
investigating whether it can have an impact on enterprise’s TFP. Finally, previous
studies have not deeply investigated the heterogenous impact of AI on enterprise
TFP from different regions, industries and enterprises, which also cannot provide a
sufficient reference for the design of tailored measures. Given the above limitations,
based on the balanced panel data of 938 manufacturing listed companies in China
during 2011–2020, this paper systematically examines the impact of AI on China’s
manufacturing enterprises from the aspects of impact degree, mechanisms and het-
erogeneities, which can provide an important theoretical and practical basis for deci-
sion-makers to formulate scientific policies to give full play to the positive role of AI
in promoting the high-quality development of China’s manufacturing industry.
3 Theoretical mechanism and research hypotheses
3.1 AI and enterprise TFP
Incorporated with the Internet, big data, cloud computing and other technologies,
AI is being applied extensively in various aspects of enterprises. It enhances labor
13
productivity by decreasing the marginal production cost, enabling enterprises’ R&D,

production, organization and operation more intelligent, so as to improve enterprise
TFP. The direct impact of AI on enterprise TFP is primarily embodied in the follow-
ing three aspects.
Firstly, AI can accelerate the pace of updating and intelligence of machines
within the organization by using intelligent devices like industrial robots (Aghion
et al., 2018). These machines have the merits of automation, self-determination
and high precision, which enable companies to continuously enhance the efficiency
of their production operations. Meanwhile, companies can modify their existing
resource endowments, such as manpower, raw materials, and machinery and equip-
ment, based on the varied data supplied by AI and in accordance with the scenario.
It enables them to maintain a relatively stable production capacity and enhance the
product yield rate, which regulates the decrease in allocation efficiency caused by
excessive consumption of corporate resources (Graetz and Michaels, 2018).
Secondly, AI can lessen the demand for low-skilled labor in companies, thus
reducing labor input and labor costs. AI devices can work continuously and with
far higher productivity and quality than human workers, which in turn increases the
efficiency of the completion of production processes (Graetz and Michaels, 2018).
For example, in the warehouse logistics stage, AI algorithms can subdivide vari-
ous major business processes to realize automatic screening, handling, and loading
and unloading of products and other operational processes. As AI becomes more
productive, it tends to boost TFP by reducing a company’s investment in labor for
regular tasks.
Finally, enterprises can also establish intelligent factories based on artificial neu-
ral networks, cluster analysis and other AI algorithms, and thus realize real-time
monitoring and active early warning of danger in the whole industrial chain (Hu
et al., 2019). Besides, with the adoption of AI into the enterprise, enterprise can
communicate with each other immediately, saving time cost and making the overall
process more agile. This can not only optimize the enterprise organization and man-
agement mode (Dixon et al., 2021), but also strengthen the collaborative production
capacities of various departments and achieve the delicacy management of produc-
tion process, so as to improve the enterprise TFP (Arakpogun et al., 2021). Based on
foregoing analysis, we formulate Hypothesis 1.
Hypothesis 1 AI can improve the TFP of manufacturing enterprises.
3.2 AI, technological innovation and enterprise TFP
Endogenous growth theory holds that technological progress is the decisive factor
to maintain stable economic growth (Romer, 1990). By directly participating in the
R&D process, AI can enrich enterprises’ innovation knowledge, and reduce enter-
prises’ innovation costs, as well as essentially improve enterprises’ technological
innovation level, realizing the transition from traditional enterprises to intelligent
manufacturing (Liu et al., 2020).
13
To be specific, firstly, AI can increase innovative knowledge. depending on

AI’s advantage in data acquisition, enterprises can obtain their required industrial
information and knowledge conveniently, and further identify the acquired knowl-
edge through precise algorithms, which greatly improves the ability of enterprises’
knowledge acquisition and absorption, and enriches the innovation knowledge of
enterprises (Olan et al., 2022). Secondly, AI optimizes efficiency and reduces the
innovation risk within the company. For instance, enterprises can use virtual real-
ity, AI and other related technologies to establish a virtual test system, map and
iterate the physical equipment in the virtual space, and quickly stimulate the tradi-
tional test process. This significantly reduces enterprises’ innovation costs, shortens
enterprises’ R&D cycle and significantly improves enterprises’ innovation efficiency
(Patel and Shah, 2021). Finally, AI enables companies to enhance profitability while
using fewer resources (Lee et al., 2022). Against the background of higher return
on AI capital, companies can have more funds for R&D design, dynamically adapt
the direction of new product, and modify products on time, which further facilitates
technological innovation in manufacturing (Liu et al., 2020). In summary, AI can
not only enrich enterprises’ innovation knowledge, but also optimize the innovation
process and enterprises’ innovation ability.
According to Schumpeter’s theory, entrepreneurial innovative behavior is an
intrinsic force of economic growth (Schumpeter, 1942). Technological innovation
is conducive to enterprises to quickly absorbing advanced knowledge and technolo-
gies, thus increasing the production capacity of traditional factors such as labor and
capital, and realizing more professional, standardized and intelligent production.
This significantly reduces enterprise production costs and improves enterprise TFP.
Based on analysis above, Hypothesis 2 can be proposed.
Hypothesis 2 AI can promote manufacturing enterprises to improve their techno-

logical innovation ability, and then improve their TFP.
3.3 AI, human capital optimization and enterprise TFP
ACEMOGLU (1996) stated that the maximum marginal productivity can be

achieved only when there is an effective match between technology and human capi-
tal. Indeed, AI is also inseparable from a highly skilled workforce to cooperate so
as to increase productivity more efficiently (Berg et al., 2018). With the increasing
application and penetration of AI in the field of manufacturing, the original employ-
ment structure has been broken, which is reflected in the increased market demand
for a highly-skilled labor force (Benzell et al., 2015; Michaels et al., 2014). Hence, it
promotes human capital optimization in manufacturing industry.
The impact of AI on the structure of human capital is manifested in the following
three aspects. Firstly, AI can replace low-skilled labor through substitution effects.
AI is mostly applied in production of manufacturing, such as with industrial robots
that can perform labor-intensive operations like welding, spray painting, loading, and
unloading in place of workers. Some scholars have discovered that routine tasks that
are reproducible and take a long time to operate are more susceptible to replacement
13
by AI (Acemoglu and Restrepo, 2020). The replaced low-skilled workers can obtain
the opportunities to improve their professional skills by learning and training pro-
vided by AI, which enable them to be competent for the new labor positions in the
intelligent environment. This can also directly optimize enterprises’ human capital,
thus optimizing the human capital structure of the whole industry (Mitchell et al.,
1990). Secondly, given that AI, as an integrator of automation technology, necessi-
tates a large number of specialists in the fields of programming, operation and main-
tenance for its application, the demand for highly skilled workforce and specialized
technical training will skyrocket. This will significantly promote the optimization
of enterprises’ human capital. Finally, due to its high permeability and universal-
ity, AI has changed traditional enterprises’ operation model and spawned intelligent
manufacturing industry. In this context, more and more highly-skilled workers are
attracted into this industry, and the quality of human capital has been significantly
improved.
Hence, high-quality human capital can not only promote enterprise technologi-
cal progress, but also optimize enterprise factor allocation, which are beneficial to
the improvement of enterprise TFP (David and Dorn, 2013). Numerous studies have
verified that high-quality human capital is a significant impetus for the transforma-
tion and upgrading of manufacturing industry (Che and Zhang, 2018). Building on
this, hypothesis 3 can be provided.
Hypothesis 3 AI can improve the TFP of manufacturing enterprise by optimizing

human capital structure.
3.4 AI, market matching amelioration and enterprise TFP
Currently, consumers’ demand has shown the characteristics of alienation, refine-

ment and customization. As such, it becomes increasingly difficult for enterprises
to comprehend consumers’ real needs, and thus resource waste and inventory accu-
mulation occur frequently. To address this, enterprises can adopt AI technologies to
promote the matching of demand and supply, trying to avoid the market mismatch,
so as to boost enterprise TFP.
To be specific, on the demand side, enterprises can collect and analyze mas-
sive data related to consumers’ needs through AI, so as to identify target cus-
tomer groups and make personalized product recommendations (Tung, 2019).
This helps to improve the circulation efficiency of goods and reduce the search
matching cost (Wang et al., 2022b). On the supply side, there are two major
impacts of AI on market matching. During the R&D and design phase, enter-
prises can adopt AI to establish market prediction models, compressively ana-
lyzing the market’s real-time and historical data. Companies can concentrate on
the general requirements of consumers while developing and designing prod-
ucts, as well as when enhancing their quality and packaging, which results in a
systematic approach for product optimization. Enterprises utilize AI to establish
real-time monitoring of resource usage, product quality, and proactive warning
of hazards during the production operation stage (Hu et al., 2019). In this way,
13
Fig.2 Mechanism analysis diagram
production factors in the workshop can be allocated effectively, and resource

allocation is further optimized (Huang et al., 2017). Therefore, AI can to some
extent improve the matching efficiency of market supply and demand. This is
conducive to reducing enterprises’ search and matching costs, enabling enter-
prises to allocate more resources in R&D and thereby promoting the improve-
ment of enterprise TFP. Based on above analysis, Hypothesis 4 can be proposed.
Hypothesis 4 AI can improve the TFP of manufacturing enterprise by improving

market matching efficiency.
The transmission mechanism of AI affecting manufacturing enterprise TFP is

illustrated in Fig. 2.
13
4 Research design, variables selection and data description
4.1 Research design
4.1.1 Benchmark regression model
To quantitively examine the impact of AI on the TFP of China’s manufacturing

enterprises, this study constructs the benchmark regression model as follows.
n
∑
TFPit = 𝛼0 + 𝛼1 AIit + 𝜑j Xjit + 𝛿t + 𝜇i + 𝜀it (1)
j=1
where i represents manufacturing listed companies; t represents year; TFP is the

explained variable, which is the total factor productivity of each listed company; AI
represents the core explanatory variable, which is artificial intelligence; X represents
a series of control variables, consisting of firm age, assets-liability ratio, return on
assets, growth, total assets turnover; j denotes different control variables,𝛼0 , 𝛼1 , 𝜑j ,
represent the coefficients to be estimated; 𝛿t , 𝜇i , 𝜀it represent time-fixed effect, firm-
fixed effect and stochastic error, respectively.
4.1.2 The mediating effect model
To examine the impact channels between AI and the TFP of China’s manufactur-
ing enterprises, referring to Baron and Kenny (1986), we construct the mediating
effect model as follows.
n
∑
TFPit = 𝛼0 + 𝛼1 AIit + 𝜑j Xjit + 𝛿t + 𝜇i + 𝜀it (2)
j=1
n
∑
Mit = 𝛽0 + 𝛽1 AIit + 𝛾j Xjit + 𝛿t + 𝜇i + 𝜆it (3)
j=1
n
∑
TFPit = 𝜃0 + 𝜃1 AIit + 𝜃2 Mit + 𝜗j Xjit + 𝛿t + 𝜇i + 𝜏it (4)
j=1
where M is the mediating variable, which includes the variables of technological

innovation, human capital and market matching. Equation (2)-Eq. (4) together con-
stitute the mediating effect model, among which, Eq. (2) examines the impact of AI
on enterprise TFP, Eq. (3) examines the impact of AI on the mediating variables,
and Eq. (4) examines the impact of AI and the mediating effect variables on enter-
prise TFP, respectively. If 𝛼1 in Eq. (2) is significant, it indicates that AI has a signif-
icant effect on enterprise TFP; also, if 𝛽1 in Eq. (3) and 𝜃2 in Eq. (4) are significant,
then it indicates that there is a mediating effect.
13
4.2 Variables selection
4.2.1 The explained variable: enterprise TFP
Enterprise TFP is the explained TFP in the benchmark regression. We use the WLP
method to measure the TFP of China’s 938 listed manufacturing enterprises during
2011–2020, with the production function shown in Eq. (5).
ln Yit = 𝛼0 + 𝛼1 ln Kit + 𝛼2 ln Lit + 𝛼3 ln Mit + 𝜀it (5)
where Yit stands for the operating revenue of firm i in year t; Kit and Lit stand for
capital input and labor input, gauged by net fixed assets and the number of employ-
ees, respectively; and Mit stands for intermediate input, denoted by (operating
cost + period cost—depreciation and amortization—employee compensation). In
this paper, the logarithm of the residuals is calculated to measure the TFP of the
enterprises. Furthermore, the TFP measured by the OP method and the LP method
is also adopted as a robustness check. While using the ACF method, LP method and
OP method for estimation, we referred to the prodest procedure used by Mollisi and
Rovigatti (2017).
4.2.2 The core explanatory variable: AI
Since AI indicators at the industry or regional level are less likely to capture the charac-
teristics of individual firms, and the share of intangible assets associated with AI may
be affected by ostentatious investment (Triplett, 1999). Thus, we draw on previous lit-
erature (Meng and Zhang, 2021; Zhou et al., 2022), based on the China Artificial Intel-
ligence Development Report 2018 published by Tsinghua University, in which a series
of AI-related terms were used as keywords. Then, python software was used to retrieve
and match the text in the annual report of each company separately. Finally, the fre-
quencies of AI-related keywords were computed as enterprise’s artificial intelligence
(AI).
4.2.3 Mediating variables
Based on above analysis, this study chooses technological innovation, human capital
and market matching as mediating variables to test the mediating effect of AI affect-
ing the TFP of manufacturing enterprises. Among them, technological innovation (TI)
is characterized by the number of patent applications of the enterprise (Lin and Zhu,
2019); human capital (HC) is measured by the proportion of enterprise employees with
bachelor degree or above in the total number of employees; market matching (MM) is
reflected by enterprise inventory turnover days (Chao et al., 2021).
4.2.4 Control variables
To obtain the net effect of AI on enterprise TFP, we control another series of influ-
encing factors at enterprise level, including Firm age (Age), Assets-liability ratio
13
Table 1 Definition of main variables

Variable Symbol Definition
Founding age Age Ln (firm age + 1)

Assets-liability ratio LEV Total liabilities / total assets
Return on equity ROE Net profit / [ (owner’s equity at the
beginning of the period + owner’s
equity at period end) / 2]
Growth Growth Enterprise operating income growth rate
Total assets turnover ATO Operating income/ [(total assets at the
beginning of the period + total assets
at the end of the period)/2]
(LEV), Return on assets (ROA), Growth (Growth), Total assets turnover (ATO).
Among them, Firm age (Age), is expressed as the natural logarithm of the firm’s age
of incorporation plus one. There are marked variations in the development patterns
and external environments of firms at different ages, which consequently impact on
TFP. Since older firms usually lack sensitivity to technological change, they are less
likely to adapt to new technologies (Guillén, 2002). Assets-liability ratio (LEV),
expressed as the ratio of capital to debt. This relates to the firm’s access to credit,
which affects the firm’s TFP (Li et al., 2018). Growth (Growth), expressed as the
growth rate of operating income, is commonly used to measure the growth status
of a corporation (Wen et al., 2022a). The higher Growth, the more beneficial it is
for the growth of the firm and thus improve TFP. Return on assets (ROA) is repre-
sented by the ratio of net income to total assets. Higher ROA means stronger profit-
ability, greater potential for growth, and easier access to TFP. Total asset turnover
ratio (ATO) is indicated by the ratio of operating income to total assets. ATO, which
reflects the operational efficiency of a company’s assets. Better ATO implies that a
company has strong sales capacity and risk management ability (Forrest and Night-
ingale, 2017). The specific definitions of the above control variables are shown in
Table 1.
4.3 Data description
The purpose of this study is to examine the impact of AI on the TFP of manufactur-
ing enterprises, and 938 listed companies in China’s manufacturing industry from
2011 to 2020 are chosen as a research sample. Taking into account that the 2008
global financial crisis and its effects negatively impacted many international com-
panies, as well as the limitations imposed by the fact that the majority of companies
began disclosing information on employee education levels in 2011 and that there
was no consistency in industry classification prior to that year, the sample start date
is set to 2011 in this paper.
We select listed companies in manufacturing industry as the research object
for the following reasons: manufacturing industry is more widely distributed in
the country, and the sample of companies in the same segment is larger and more
13
Table 2 Descriptive statistics of Variable Obs Mean Std. Dev Min Max
the main variables
TFP 9380 9.374 1.059 5.585 14.111
AI 9380 0.171 0.518 0.000 5.961
Age 9380 2.809 0.367 1.099 3.829
LEV 9380 0.396 0.188 0.007 1.957
Growth 9380 0.151 0.412 -0.882 15.610
ROA 9380 0.045 0.063 -0.894 0.478
ATO 9380 0.704 0.467 0.013 7.871
TI 9380 2.993 1.609 0.000 9.239
HC 9380 0.214 0.251 -1.270 18.357
MM 9379 4.684 0.797 0.799 9.926
comparable. Moreover, manufacturing industry accounts for a higher proportion of

listed companies and is more representative. In addition, as a pillar industry in the
national economy, China launched an intelligent manufacturing demonstration pro-
ject in 2015 to advance the intelligent transformation of the manufacturing industry
and realize manufacturing powerhouse.
The sample data are obtained from CSMAR Database, WIND database, the Inter-
national Federation of Robotics (IFR), China Statistical Yearbook and China Indus-
trial Economic Statistical Yearbook, etc. Consistent with the later discussion, first,
we remove the financial firms and ST and *ST samples. Second, firms with missing
data are removed. And then, we exclude the sample of firms whose listing year is
after 2011 to construct balanced panel data. Eventually, we obtained data for 938
firms with a total of 9380 samples. The descriptive statistics and correlation coef-
ficient of variables are shown in Table 2 and Fig. 3. As can be shown through Fig. 3,
the correlation coefficients between the variables are relatively small, which implies
that there is no severe problem of multicollinearity among the variables.
5 Empirical results and discussions
5.1 Benchmark regression results
In this paper, the pooled-OLS and the fixed-effect (FE) OLS model are used to esti-
mate the benchmark regression model, respectively, to quantitatively investigate the
direct impact of AI on the TFP of China’s manufacturing enterprises. The regression
results are presented in Table 3.
Column (1) in Table 3 reports the estimated results of pooled-OLS model, and
columns (2)-(4) provide the estimation results of three FE models. It can be seen
that AI coefficients in the four OLS regression models are all significantly positive
at the significance level of 1%, which indicates that AI can significantly improve the
TFP of China’s manufacturing enterprises, and hence Hypothesis 1 is verified. This
is in accordance with the findings of Yang (2022) and Alguacil et al. (2022).
13
Fig.3 Correlation coefficient of the main variables
5.2 Endogenous treatment: IV method
When discerning the impact of AI on TFP, the benchmark regression results above
may be subject to endogeneity, leading to biased regression results. To be specific,
based on the benchmark regression results above, AI has a beneficial influence on
enhancing TFP, and inversely, TFP may also affect AI through technological spillo-
vers and other means, thus a reverse causal relationship may be present between AI
and TFP. In addition, there are numerous factors influencing TFP, and the control
variables involved in this paper are limited, which may difficult to avoid omitted
variables. To overcome the endogenous problem arising from missing variables and
two-way causality, this study employs the IV method to ensure the robustness of the
benchmark regression results.
The key to using IV method to address endogenous problems lies in selecting
appropriate IV variable. Referring to Chao et al. (2020), this study adopts the
number of microelectronic computers produced per 10,000 people in 1994 as
the IV of the development level of AI in China’s manufacturing enterprises. The
main reason lies in the following two aspects. On the one hand, the growth of
AI is inextricably linked to microcomputers, satisfying the relevance condition.
On the other hand, the past production of microcomputers has no direct bearing
on enterprises’ TFP nowadays, fulfilling the exogeneity condition. Moreover, to
13
Table 3 Benchmark regression Variables Pooled-OLS FE FE FE

results
(1) (2) (3) (4)
AI 0.1315*** 0.0841*** 0.0620*** 0.0736***
(9.6754) (10.4760) (4.6414) (9.3241)
Age 0.4190*** 1.1889*** 0.1336*** 0.2520***
(19.9017) (65.6196) (5.7481) (5.1902)
LEV 2.9446*** 0.7795*** 2.9566*** 0.8673***
(51.7515) (18.3695) (52.2681) (20.9493)
Growth -0.0225 0.1284*** -0.0292 0.1183***
(-0.9725) (11.4102) (-1.3057) (10.6794)
ROA 4.3058*** 1.0491*** 4.5452*** 1.0722***
(23.6155) (10.1318) (25.1746) (10.7013)
ATO 0.8379*** 0.6455*** 0.8832*** 0.6810***
(21.5807) (17.3093) (21.4940) (17.3605)
Constant 6.2280*** 5.1910*** 6.9948*** 7.7649***
(106.1566) (88.2228) (109.8687) (58.8330)
Year FE No No Yes Yes
Firm FE No Yes No Yes
Observations 9380 9380 9380 9380
R-squared 0.544 0.951 0.573 0.955
*
, **, and *** mean significant at the level of 10, 5, and 1%, respec-
tively. The values in brackets are t-value
enable the instrumental variable with a time trend, we further multiply it with
the fixed asset investment in information transmission, software and information
technology industry. Considering that AI is mainly carried by above-mentioned
industry, in general, AI is also relatively higher in regions with higher invest-
ment in fixed assets in such industry. Based on above analysis, it can be found
that multiplier term for the number of microcomputers produced per 10,000 peo-
ple in 1994 and the fixed asset investment in information transmission, software
and information technology industry satisfies the prerequisite that an IV must
have, and therefore we believe that it is appropriate to take it as the IV of AI
in China’s manufacturing enterprises. The estimation results of IV method are
shown in Table 4.
The estimation results of the first stage regression show that the coefficient of
IV is significantly positive at the significance level of 1%, and passes the weak
IV test. Hence, the selected IV has been proved to be reasonable. From the esti-
mated results of the second stage regression, it can be found that after the under-
lying endogeneity problems are addressed, the promoted impact of AI on TFP of
China’s manufacturing enterprises is still significant, which confirms the robust-
ness of the conclusion of benchmark regression.
13
Table 4 Results of the IV Variables First-stage regression Second-

estimation stage
regression
AI TFP
IV 0.0025***
(4.7302)
AI 0.3677***
(2.5909)
Controls Yes Yes
Year FE Yes Yes
Firm FE Yes Yes
F value 22.38 –
N 9380 9380
R-squared 0.562 0.277
*
Table 5 Results of robustness tests

Variables The explained variable: TFP
(1) (2) (3) (4) (5) (6) (7) (8)
AI 0.0745*** 0.0431*** 2.8098*** 0.0203*** 0.0694*** 0.0699*** 0.0711*** 0.0678***

(9.3830) (5.9049) (5.8859) (3.7805) (8.6275) (8.3320) (8.5831) (8.4157)
Constant 7.8257*** 6.3193*** 7.7333*** 7.6937*** 7.9239*** 7.7088*** 7.7133*** 7.5025***
(58.8903) (50.4945) (57.7732) (56.1830) (60.6558) (44.4746) (53.8713) (43.6976)
Controls Yes Yes Yes Yes Yes Yes Yes Yes
Year FE Yes Yes Yes Yes Yes Yes Yes Yes
Firm FE Yes Yes Yes Yes Yes Yes Yes Yes
N 9380 9380 9380 8442 7440 8442 8117 7504
R-squared 0.956 0.938 0.955 0.959 0.960 0.951 0.952 0.964
*
, **, and *** mean significant at the level of 10, 5, and 1%, respectively. The values in brackets are
t-value
5.3 Robustness tests
To verify the robustness of benchmark regression results, we use four schemes for
robustness test, and the estimated results are listed in Table 5. Specifically, the four
robustness tests include:
(1) Replacing explanatory variable. We use the TFP of China’s manufacturing enter-
prises calculated by the LP method and OP method to replace the explanatory
variable in Eq. (1) and then re-estimate, and the results are reported in columns
(1)– (2) of Table 5.
13
(2) Substitution of explanatory variables. We employ the frequency of the sum of AI

keywords in the total number of words in the full text as an explanatory variable
to re-run the regression. The results are displayed in column (3). Furthermore,
we apply industrial robots as explanatory variables. Given that AI in production
processes is primarily realized through industrial robots (Bessen et al., 2019),
and industrial robot applications can, to some extent, mirror the actual develop-
ment of AI (Zhao et al., 2022). Referring to Acemoglu and Restrepo (2020),
industrial robot density is employed in this study to measure AI at the level of
Chinese manufacturing enterprises, as is shown in Eq. (6). Limited by the fact
that the data for industrial robots are updated to 2019, we therefore re-regressed
using the panel data from 2011–2019, and the results are shown in column (4).
PWPjit=2011 MRCH
it
robotit = × CH (6)
ManuPWPt=2011 Li,t=2010
where MRCH it
denotes the industrial robot stock in industry i in year t; Li,t=2010
CH
denotes the number of employees in industry i in 2010. And ManuPWPt=2011

denotes the median percentage of employees in production departments of all
enterprises in manufacturing industry in 2011.
(3) Excluding strategic corporate behavior. Zhao et al. (2020) discovered that enter-
prises have strategic behavior in disclosing information about "Internet + ".
Hence, enterprises may exaggerate their AI applications in their annual reports.
To rule out bias caused by firms’ strategic disclosure behavior, we use discipli-
nary notifications from the China Securities Regulatory Commission, Shanghai
Stock Exchange and Shenzhen Stock Exchange to remove those firms that have
been disciplined before, and then regress them. The results are displayed in
column (5).
(4) Lagging the explanatory variable for one period. We lag all explanatory variables
in Eq. (1) for one period and then re-estimate, the results are shown in column
(6) of Table 5.
(5) Adjusting the sample. Given that there are great differences between municipali-
ties and other prefecture-level cities in economic development, enterprise scale
and the development of AI, we exclude the samples of enterprises registered in
municipalities and then re-regress Eq. (1), and the results are provided in column
(7) of Table 5.
(6) Changing the sample time. We shorten the original sample time from 2011–2020
to 2012–2019, and then re-regress Eq. (1), and the results are presented in col-
umn (8) of Table 5.
It can be seen in columns (1)-(8) of Table 5, AI coefficients in the four men-

tioned regression models above are significantly positive at the significance level
of 1%, which is consistent with benchmark regression estimated results. This
indicates that the conclusion that AI can enhance the TFP of China’s manufactur-
ing enterprises is reliable.
13
5.4 Heterogeneity analysis
5.4.1 Regional heterogeneity
Due to the uneven regional development in China, the impact of AI on TFP of Chi-
na’s manufacturing enterprises may vary in different regions. Therefore, based on
the geographical location, we divide the enterprise samples into three groups: the
East, The Central and the West for regression, and the results are shown in columns
(1)-(3) in Table 6.
It can be seen that the coefficients of AI in the East, Central and West are sig-
nificantly positive, and the coefficients are larger in the East than in the Central and
West, which indicates that the promotion of AI to TFP of eastern enterprises is sig-
nificantly stronger than the central and western enterprises. The main reason is that
due to a higher level of economic development and regional innovation, the eastern
AI development is significantly faster than that in the central and western, laying a
solid foundation for the integration of AI into the manufacturing industry. By com-
parison, due to the backward development in high-tech industry, insufficient R&D in
AI technologies and smaller AI industry scale, the integration of AI in manufactur-
ing enterprises in central and western China is substantially hindered.
5.4.2 Industrial heterogeneity
Given that there exist obvious differences between different manufacturing indus-
tries in industrial chain characteristics, technical level and market environment, the
impact of AI on TFP may vary greatly across different manufacturing industries. To
verify this, we categorize the sample enterprises into two groups with reference to
Zhu and Li (2019): the Technology-intensive and the Non-technology-intensive,1 to
analyze the industrial heterogeneity of AI affecting TFP of China’s manufacturing
enterprises. The results are reported in columns (4)-(5) in Table 6.
It is evident that AI has a greater contribution to the TFP of technology-intensive
enterprises relatively to non-technology-intensive enterprises. It is known that tech-
nology-intensive industry with high degree of automation, advanced hardware and
software conditions, and high-quality employees, can give full play to the effect of
AI in improving enterprise TFP. However, non-technology-intensive industries, with
low automation, low dependence on technology, and lack of professionals, cannot
provide the basis for the integration of AI and enterprise operation, which limits the
impact of AI in promoting enterprise TFP.
1
The Technology-intensive consist of 7 sub-industries, which includes: (1) chemical raw materials and
chemical products manufacturing, (2) pharmaceutical manufacturing, (3) chemical fiber manufacturing,
(4) automobile manufacturing, (5) railroad, ship, aerospace and other transportation equipment manufac-
turing, (6) electrical machinery and equipment manufacturing (7) computer, communications and other
electronic equipment manufacturing. The rest of the manufacturing industry segments are classified as
the Low-tech.
13
1134
13
Table 6 Regression results of heterogeneity analysis
Variables Regional heterogeneity Industrial heterogeneity Business life cycle hetero- Enterprise ownership heterogene-
geneity ity
East Central West Non-technologi- Technology-intensive Growing Mature Non-State-owned State -owned
cal-intensive
(1) (2) (3) (4) (5) (6) (7) (8) (9)
AI 0.0763*** 0.0567*** 0.0479* 0.0582*** 0.0711*** 0.0839*** 0.0188 0.0725*** 0.0752***

(8.5224) (3.0034) (1.8744) (5.2244) (7.1315) (7.6314) (1.5406) (6.9450) (5.9246)
Constant 7.7684*** 8.1399*** 6.7054*** 7.5963*** 8.0415*** 7.6401*** 10.3129*** 7.9346*** 7.3709***
(51.8530) (24.7999) (13.3614) (43.0687) (44.1675) (38.9942) (9.0400) (51.8665) (25.6211)
Controls Yes Yes Yes Yes Yes Yes Yes Yes Yes
Year FE Yes Yes Yes Yes Yes Yes Yes Yes Yes
Firm FE Yes Yes Yes Yes Yes Yes Yes Yes Yes
N 6428 1792 1140 4350 5030 4931 4359 6149 3227
R-squared 0.955 0.953 0.960 0.963 0.952 0.964 0.968 0.939 0.968
*
, **, and *** mean significant at the level of 10, 5, and 1%, respectively. The values in brackets are t-value
Economic Change and Restructuring (2023) 56:1113–1146
5.4.3 Business life cycle heterogeneity
The impact of AI on TFP of manufacturing enterprises in different business life

cycle stages may be different. To verify this, we divide the sample into two groups
based on Han et al. (2021): the Mature groups (enterprises older than the industry
median) and the Growth groups (enterprises younger than the industry median).
Then we re-estimate correspondingly, and the estimated results are shown in col-
umns (6)-(7) of Table 6.
It can be found that AI can significantly promote the TFP in growing enter-
prises, which is not significant for mature enterprises. As can be explained in this
way, the growing enterprises, which are smaller and more energetic, are more
reactive to new technologies. New technologies have gradually or widely partici-
pated in their operation processes, and the technological and talent foundation
for AI to give full play to its effect has been consolidated. It enables AI to sig-
nificantly promote enterprises’ TFP. In contrast, mature enterprises are in a stable
development stage and account for a certain market share, but the huge produc-
tion scale also makes enterprises reach the bottleneck.
5.4.4 Enterprise ownership heterogeneity
Given that numerous studies have proved that different ownership enterprises
have great differences in resource allocation and technological utilization
(Gadenne et al., 2009), the impact of AI on manufacturing enterprises with dif-
ferent ownership is likely to be different. Therefore, we divide the sample enter-
prises into two groups: state-owned enterprises and non-state-owned enterprises
for regression. The results are presented in columns (8)-(9) in Table 6.
It is revealed that AI can significantly promote TFP both in state-owned and
non-stated-owned enterprises, and its effect is greater in state-owned enterprises.
The main reason lies in that, compared with non-state-owned enterprises, state-
owned enterprises are more driven to respond to the call for the development of
AI technologies, they are better positioned to complete the full integration of AI
into corporate operations. They are also in charge of developing new models for
integrating AI with manufacturing, depending on AI to reorganize operational
processes and enhance total factor productivity. Hence, AI can better promote
TFP in state-owned manufacturing enterprises.
5.5 Transmission mechanism analysis
The above research has confirmed that AI can significantly promote the TFP of
China’s manufacturing enterprises. However, what is the transmission mechanism
needs to be further studied. As such, in this section, based on the previous theo-
retical analysis, we adopt the mediating effect model to empirically verify how
AI affects the TFP of China’s manufacturing enterprises from three channels:
13
Table 7 Results of transmission mechanism analysis

Variables TFP TI TFP HC TFP MM TFP
(1) (2) (3) (4) (5) (6) (7)
AI 0.0736*** 0.0585** 0.0717*** 0.0063** 0.0733*** − 0.0200** 0.0704***

(9.3241) (2.2231) (9.1394) (2.3659) (9.2966) (− 2.1213) (9.0386)
TI 0.0316***
(8.8493)
HC 0.0446*
(1.8122)
MM − 0.1680***
(− 12.8888)
Constant 7.7649*** 2.9939*** 7.6704*** 0.2080 7.7556*** 5.1616*** 8.6353***
(58.8330) (7.6022) (58.1835) (1.4010) (58.6560) (31.1591) (59.4741)
Controls Yes Yes Yes Yes Yes Yes Yes
Year FE Yes Yes Yes Yes Yes Yes Yes
Firm FE Yes Yes Yes Yes Yes Yes Yes
N 9380 9380 9380 9380 9380 9379 9379
R-squared 0.955 0.774 0.956 0.430 0.955 0.863 0.958
*
, **, and *** mean significant at the level of 10, 5, and 1%, respectively. The values in brackets are
t-value
technological innovation (TI), human capital (HP) and market matching (MM),
and the estimated results are reported in Table 7.
As shown in column (1) of Table 7, the coefficient of AI is significantly positive,
indicating that AI can promote the TFP of China’s manufacturing enterprises, which
again confirms Hypothesis 1.
The estimated results with TI as the mediating variable are shown in columns
(2)-(3) in Table 7. It can be found in column (2), the coefficient of AI is 0.0585 and
significant, indicating that AI can dramatically drive enterprise technological inno-
vation. Moreover, as shown in column (3), the coefficients of AI and TI are signifi-
cantly positive, and the coefficient of AI is 0.0717, less than 0.0736, which suggests
that AI can significantly improve TFP of China’s manufacturing enterprises by pro-
moting technological innovation. Hence, Hypothesis 2 is verified.
The estimation results with HC as the mediating variable are provided in col-
umns (4)-(5) in Table 7. As shown in column (4), the coefficient of AI is 0.0063 and
significant, showing that AI significantly improves the level of human capital. In
addition, it can be seen in column (5) that the coefficient of AI is 0.0733 and signifi-
cant, less than 0.0736, indicating that AI can significantly promote TFP of China’s
manufacturing enterprises by boosting human capital level. Therefore, Hypothesis 3
is verified.
Columns (6)-(7) in Table 7 present the estimated results with MM as the mediat-
ing variable. As shown in column (6), the coefficient of AI on enterprise inventory
turnover days is -0.0200, and is significant, indicating that with the wide applica-
tion of AI, the market matching degree of China’s manufacturing industry has been
13
significantly improved. Furthermore, it can be found that the coefficient of AI in col-

umn (7) is 0.0704 and significant, indicating that AI can significantly drive TFP of
China’s manufacturing enterprises by ameliorating the degree of market matching.
Hence, Hypothesis 4 is verified.
5.6 Further analysis: exogenous policy shock effect
5.6.1 Continuous DID model
The Ministry of industry and information technology of China selected a batch of

enterprises as the pilot units of “Intelligent Manufacturing Pilot” project every year
from 2015 to 2018, aiming to steadily promote the “Made in China 2025” strategy
by giving full play to the role of AI in improving enterprises informatization level
and production efficiency. Therefore, in order to further confirm the previous con-
clusions, we take the intelligent manufacturing pilot policy as a quasi-natural experi-
ment, and adopt the DID method to qualify the impact of the exogenous policy
shock on TFP of China’s manufacturing enterprises based on the panel data of 938
manufacturing listed companies during 2011–2020. The continuous DID model is
presented as follows.
n
∑
TFPit = 𝛼 + 𝛽didit + 𝛾j Xjit + 𝛿i +𝜇i +𝜀it (7)
j=1
where i represents enterprise; t represents year; did represents the dummy variable
of “Intelligent Manufacturing Pilot” policy. If enterprise i is selected into the pilot
project in year t, then did is assigned as 1 in year t and the later years; otherwise, did
is assigned as 0. 𝛿i and 𝜇i represent the fixed effect of enterprise and year, respec-
tively; 𝜀it represents the random error term.
5.6.2 Parallel trend test
One prerequisite of policy evaluation using DID method is to satisfy the assump-
tion of parallel trend. With reference to Beck et al. (2010), we adopt the event study
approach to test parallel trend assumption. To reduce multicollinearity, the dummy
variable (pre7) in the 7th year before policy implementation is excluded before esti-
mation. The regression model is shown as follows.
6 4 n
∑ ∑ ∑
TFPit = 𝛼 + 𝛽n preNit + 𝜆currentit + 𝜃n postNit + 𝛾j Xjit + 𝛿i +𝜇i +𝜀it
n=1 n=1 j=1
(8)
where current denotes the current year the policy was implemented; pre1-pre6,
respectively, denote the 6th year to the 1st year before policy implementation; post1-
post4 denote the 1st and 4th year after policy implementation.
Figure 4 illustrates the statistical results of the regression coefficients 𝛽 and 𝜃 . It
can be found that the 𝛽 coefficient is not significant from pre_6 to pre_1, indicating
13
Fig.4 Results of the parallel trend test
that there is no significant difference in TFP between the treatment group and the
control group before policy implementation, and thus the parallel trend assump-
tion is satisfied. Additionally, the impact of intelligent manufacturing pilot policy
on enterprise TFP is gradually and significantly positive in the 2nd year after the
implementation of the policy, which indicates that the smart manufacturing pilot
policy has a certain lag. The main explanation may be that the smart manufactur-
ing pilot policy mainly improves production efficiency through artificial intelligence
R&D and application, which has a large investment scale and a long completion
time. Hence, there is a lag in the effect of the smart manufacturing pilot policy.
5.6.3 Policy effect estimation
Table 8 offers the estimated results of the continuous DID method. It is disclosed
that the intelligent manufacturing pilot policy has a significant impact on TFP
improvement of China’s manufacturing enterprises. Compared with non-pilot enter-
prises, the policy can increase the TFP of pilot enterprises by 5.54%. This finding
once again proves that AI can significantly improve the TFP of China’s manufactur-
ing enterprises, strongly supporting the benchmark regression result.
5.6.4 Placebo test
To diminish the effect of random factors on DID regression results, referring to

Cai et al. (2016), we randomly select intelligent manufacturing pilot enterprises for
13
Table 8 Estimation results of Variables Did

continuous DID method
Did 0.0554***
(3.4965)
Constant 7.6859***
(57.0767)
Controls Yes
Year FE Yes
Firm FE Yes
N 9380
R-squared 0.955
*
placebo test. Figure 5 displays the kernel density of did coefficient and its corre-
sponding p-value distribution after 500 random samples. It is found that the mean
value of did coefficient is close to zero, far less than 0.0554, and the coefficient dis-
tribution approaches the normal distribution. This indicates that the estimation result
of the DID model is robust.
On the whole, the analysis results of the exogenous policy shock effect further
confirm that AI can significantly boost the improvement of TFP in China’s manufac-
turing enterprises.
Fig.5 Results of the placebo test
13
6 Conclusions and policy implications
AI has been proved to be an important diver to promote TFP in different economies,

but related research at micro-enterprises level still remains limited. Based on the bal-
anced panel data of 938 China’s manufacturing listed companies from 2011 to 2020,
this paper systematically examines the impact of AI on TFP of China’s manufactur-
ing enterprises by using the fixed-effect model, the mediating effect model and the
DID approach, on basis of which its transmission mechanism, heterogeneities and
the policy shock effect are further analyzed. The main conclusions are as follows:
(1) AI can significantly promote the improvement of TFP of manufacturing enter-

prises in China, and after the endogenous treatment and an array of robustness
tests, the conclusion remains valid.
(2) The impact of AI on TFP varies dramatically in manufacturing enterprises with
different geographical location, industrial characteristics, ownership and life
cycle stage, and it is more significant in the eastern enterprises, the technology-
intensive industries, the state-owned enterprises, and the growing enterprises.
(3) Technological innovation, human capital optimization and market matching ame-
lioration are confirmed as three important channels for AI to facilitate the TFP
of China’s manufacturing enterprises.
(4) The exogenous policy shock effect studied by DID approach also proves that AI
can significantly improve the TFP of China’s manufacturing enterprises.
Drawing on the advanced outcomes of developed countries in AI and smart man-

ufacturing (Hossain et al., 2022; Kehayov et al., 2022; Rathore et al., 2022), emerg-
ing economies such as China should further develop new AI technologies and utilize
the data mining of AI, together with the technology spillover features, to interoperate
with manufacturing industries. The above results can provide scientific references
for formulating smart manufacturing strategies in emerging economies, as well as
empirical evidence for industrial shifts and AI changes in developed countries.
Based on the findings above, to ensure that the development of AI better improves
China’s manufacturing enterprises’ TFP, this paper proposes the following policy
implications:
(1) Given that this paper has confirmed that AI can significantly promote the TFP
of China’s manufacturing enterprises, the Chinese government should continue
to increase the construction of AI infrastructure and strengthen the protection of
AI patents and intellectual property rights, creating a healthy and benign external
environment for the development of AI. In the meantime, the Chinese govern-
ment should optimize the strategic layout of AI development in manufacturing
industry and promote the in-depth integration of AI with enterprise operation
management, data analysis, product service, and product sales, and so on.
(2) Given that the results of transmission mechanism analysis has proved that tech-
nological innovation, human capital optimization and market matching ame-
lioration are the three important channels for AI to affect the TFP of China’s
13
manufacturing enterprises, the following three points should be achieved: first,

to establish a perfect technological innovation incentive mechanism and increase
the investment in AI technologies R&D, substantially improving enterprises’
technological innovation capabilities, so as to reduce their production and
operation costs; second, continuously to improve the human capital level of
manufacturing enterprises through education, training, and production-learning-
research and other means, to lay a solid talent foundation for accelerating the
integration of AI into enterprises’ operation processes; finally, the government
should encourage enterprises and consumers to join the AI sharing platform for
realizing the cooperation between supply and demand sides, and achieve market
matching, which can effectively reduce enterprise search and matching costs.
(3) Given that the results of heterogeneity analysis demonstrate that the impact of
AI on TFP varies greatly in different enterprises, the central government should
do a good job in top-level design in promoting the application of AI in manu-
facturing industry, giving full play to the important role of AI in realizing the
high-quality development of China’s manufacturing industry. The government
should strengthen policy support for AI in underdeveloped industries and regions
by implementing some preferential policies such as tax incentives and fiscal
subsidies, so as to improve the penetration of AI technologies in manufacturing
industry, fully integrating it into enterprise’s process of R&D, production, sales,
and operation management.
Funding National Natural Science Foundation in China,71973131,Ke-Liang Wang
Declarations
Conflict of Interest All authors declare no conflict of interest.
References
Acemoglu D (1996) A microfoundation for social increasing returns in human capital accumulation. Q J
Econ 111:779–804
Acemoglu D, Restrepo P (2020) Robots and jobs: evidence from US labor markets. J Polit Econ
128:2188–2244
Acemoglu D, Restrepo P 2018. Artificial intelligence, automation, and work, The economics of artificial
intelligence: An agenda. University of Chicago Press, pp. 197-236.
Ackerberg D, Caves K, Frazer G (2006) Structural identification of production functions.
Aghion P, Jones BF, Jones CI 2018. Artificial intelligence and economic growth, The economics of artifi-
cial intelligence: An agenda. University of Chicago Press, pp. 237-282.
Ahmad T, Zhu H, Zhang D, Tariq R, Bassam A, Ullah F, AlGhamdi AS, Alshamrani SS (2022) Energet-
ics systems and artificial intelligence: applications of industry 4.0. Energ Rep 8:334–361
Alguacil M, Lo Turco A, Martínez-Zarzoso I (2022) Robot adoption and export performance: firm-level
evidence from Spain. Econ Model 114:105912
Antonelli C, Scellato G (2013) Complexity and technological change: knowledge interactions and firm
level total factor productivity. J Evol Econ 23:77–96
13
Arakpogun EO, Elsahn Z, Olan F, Elsahn F (2021) Artificial intelligence in Africa: Challenges and
opportunities, The fourth industrial revolution: implementation of artificial intelligence for grow-
ing business success. Springer, pp. 375-388.
ARC.2018. What Is China’s smart manufacturing strategy?https://www.arcweb.com/industry-best-pract
ices/what-chinas-smart-manufacturing-strategy. Last accessed on 4th Nov., 2022.
Autor DH (2015) Why are there still so many Jobs? the history and future of workplace automation. J
Econ Perspect 29:3–30
Baron RM, Kenny DA (1986) The moderator–mediator variable distinction in social psychological
research: conceptual, strategic, and statistical considerations. J Pers Soc Psychol 51:1173
Beck T, Levine R, Levkov A (2010) Big bad banks? The winners and losers from bank deregulation in
the United States. J Financ 65:1637–1667
Benzell SG, Kotlikoff LJ, LaGarda G, Sachs JD (2015) Robots are us: some economics of human replace-
ment. National Bureau of Economic Research.
Berg A, Buffie EF, Zanna L-F (2018) Should we fear the robot revolution? (The correct answer is yes). J
Monet Econ 97:117–148
Bessen J, Goos M, Salomons A, Van den Berge W (2019) Automatic reaction-what happens to workers at
firms that Automate? SSRN J. https://doi.org/10.2139/ssrn.3328877
Brandt L, Van Biesebroeck J, Zhang Y (2012) Creative accounting or creative destruction? Firm-level
productivity growth in Chinese manufacturing. J Dev Econ 97:339–351
Brynjolfsson E, Rock D, Syverson C (2018) Artificial intelligence and the modern productivity paradox:
a clash of expectations and statistics, The economics of artificial intelligence: An agenda. Univer-
sity of Chicago Press, pp. 23-57.
Cai F (2012) The coming demographic impact on china’s growth: the age factor in the middle-income
trap*. Asian Econ Pap 11:95–111
Chao X, Lian Y, Luo L (2021) Impact of New Digital Infrastructure on High-quality Development of
Manufacturing. Financ Trade Res 32:1–13 ([In Chinese])
Chao, X., Xue, Z., Sun, Y., 2020. How the new digital infrastructure affects the upgrading of foreign
trade: evidence from chinese cities. Economic Science, 46–59 [In Chinese].
Che Y, Zhang L (2018) Human capital, technology adoption and firm performance: impacts of China’s
higher education expansion in the late 1990s. Econ J 128:2282–2320
Chen M, Liu Q, Huang S, Dang C (2020) Environmental cost control system of manufacturing enter-
prises using artificial intelligence based on value chain of circular economy. Enterp Inf Syst.
https://doi.org/10.1080/17517575.2020.1856422
Chen H, Guo W, Feng X, Wei W, Liu H, Feng Y, Gong W (2021) The impact of low-carbon city pilot
policy on the total factor productivity of listed enterprises in China. Resour Conserv Recycl
169:105457
Cheng H, Jia R, Li D, Li H (2019) The rise of robots in China. J Econ Perspect 33:71–88
Choi C, Yi MH (2009) The effect of the Internet on economic growth: evidence from cross-country panel
data. Econ Lett 105:39–41
Damioli G, Van Roy V, Vertesy D (2021) The impact of artificial intelligence on labor productivity. Eura-
sia Bus Rev 11:1–25
David H, Dorn D (2013) The growth of low-skill service jobs and the polarization of the US labor mar-
ket. Am Econ Rev 103:1553–1597
Del Gaudio BL, Porzio C, Sampagnaro G, Verdoliva V (2021) How do mobile, internet and ICT diffusion
affect the banking industry? An empirical analysis. Eur Manag J 39:327–332
Dixon J, Hong B, Wu L (2021) The robot revolution: managerial and employment consequences for
firms. Manage Sci 67:5586–5605
Dolgui A, Ivanov D, Sethi SP, Sokolov B (2019) Scheduling in production, supply chain and Industry
4.0 systems by optimal control: fundamentals, state-of-the-art and applications. Int J Prod Res
57:411–432
Du L, Lin W (2022) Does the application of industrial robots overcome the Solow paradox? Evid Ch
Technol Soc 68:101932
Dwivedi YK, Hughes L, Ismagilova E, Aarts G, Coombs C, Crick T, Duan Y, Dwivedi R, Edwards J,
Eirug A (2021) Artificial Intelligence (AI): multidisciplinary perspectives on emerging challenges,
opportunities, and agenda for research, practice and policy. Int J Inf Manage 57:101994
Forrest JY-L, Nightingale J (2017) Successfully transition into the era of transient competitive advan-
tages. Northeast Assoc Bus, Econ Technol 73.
Furman J, Seamans R (2019) AI and the economy. Innov Policy Econ 19:161–191
13
Gadenne DL, Kennedy J, McKeiver C (2009) An empirical study of environmental awareness and prac-
tices in SMEs. J Bus Ethics 84:45–63
Gaglio C, Kraemer-Mbula E, Lorenz E (2022) The effects of digital transformation on innovation and
productivity: firm-level evidence of South African manufacturing micro and small enterprises.
Technol Forecast Soc Chang 182:121785
Gams M, Gu IY-H, Härmä A, Muñoz A, Tam V (2019) Artificial intelligence and ambient intelligence. J
Ambient Intell Smart Environ 11:71–86
Gnangnon SK, Iyer H (2018) Does bridging the internet access divide contribute to enhancing countries’
integration into the global trade in services markets? Telecommun. Policy 42:61–77
Graetz G, Michaels G (2018) Robots at work. Rev Econ Stat 100:753–768
Guan S, Cheng L (2020) Does product complexity matter for firms’ TFP? Technol Forecast Soc Chang
160:120233
Guillén MF (2002) Structural inertia, imitation, and foreign expansion: South Korean firms and business
groups in China, 1987–1995. Acad Manage J 45:509–525
Haefner N, Wincent J, Parida V, Gassmann O (2021) Artificial intelligence and innovation management:
a review, framework, and research agenda✰. Technol Forecast Soc Chang 162:120392
Hall BH, Mairesse J (1995) Exploring the relationship between R&D and productivity in French manu-
facturing firms. J Econom 65:263–293
Han X, Wang J, Cheng H (2021) The effect of corporate tax avoidance on salary distribution——empiri-
cal evidence from publicly listed companies in China. Int Rev Financ Anal 78:101917
Heo PS, Lee DH (2019) Evolution of the linkage structure of ICT industry and its role in the economic
system: the case of Korea*. Inform Technol Dev 25:424–454
Hossain MA, Agnihotri R, Rushan MRI, Rahman MS, Sumi SF (2022) Marketing analytics capability,
artificial intelligence adoption, and firms’ competitive advantage: evidence from the manufacturing
industry. Ind Mark Manage 106:240–255
Hsieh C-T, Klenow PJ (2010) Development accounting. Am Econ J-Macroecon 2:207–223
Hu L, Miao Y, Wu G, Hassan MM, Humar I (2019) iRobot-factory: an intelligent robot factory based on
cognitive manufacturing and edge computing. Futur Gener Comput Syst 90:569–577
Huang Y-S, Hung J-S, Ho J-W (2017) A study on information sharing for supply chains with multiple
suppliers. Comput Ind Eng 104:114–123
Huang Z, Shen Y, Li J, Fey M, Brecher C (2021) A Survey on ai-driven digital twins in industry 4.0:
Smart manufacturing and advanced robotics. Sensors 21:6340
Huang G, He L-Y, Lin X (2022) Robot adoption and energy performance: evidence from Chinese indus-
trial firms. Energy Econ 107:105837
Kehayov M, Holder L, Koch V (2022) Application of artificial intelligence technology in the manufactur-
ing process and purchasing and supply management. Proced Comput Sci 200:1209–1217
Kenny C (2003) The Internet and economic growth in less-developed countries: a case of managing
expectations? Oxf Dev Stud 31:99–113
Kijek T, Kijek A (2019) Is innovation the key to solving the productivity paradox? J Innov Knowl
4:219–225
Kim J, Park JC, Komarek T (2021) The impact of mobile ICT on national productivity in developed and
developing countries. Inf Manage 58:103442
Koutroumpis P (2009) The economic impact of broadband on growth: a simultaneous approach. Tel-
ecommun Policy 33:471–485
Lee C-C, Qin S, Li Y (2022) Does industrial robot application promote green technology innovation in
the manufacturing industry? Technol Forecast Soc Chang 183:121893
Levinsohn J, Petrin A (2003) Estimating production functions using inputs to control for unobservables.
Rev Econ Stud 70:317–341
Li M, Du W (2021) Can Internet development improve the energy efficiency of firms: empirical evidence
from China. Energy 237:121590
Li YA, Liao W, Zhao CC (2018) Credit constraints and firm productivity: microeconomic evidence from
China. Res Int Bus Financ 45:134–149
Li D, Chen Y, Miao J (2022a) Does ICT create a new driving force for manufacturing?—Evidence from
Chinese manufacturing firms. Telecommun Policy 46:102229
Li M, Jin M, Kumbhakar SC (2022b) Do subsidies increase firm productivity? Evidence from Chinese
manufacturing enterprises. Eur J Oper Res 303:388
Lin B, Zhu J (2019) The role of renewable energy technological innovation on climate change: empirical
evidence from China. Sci Total Environ 659:1505–1512
13
Liu J, Chang H, Forrest JY-L, Yang B (2020) Influence of artificial intelligence on technological innova-
tion: evidence from the panel data of china’s manufacturing sectors. Technol Forecast Soc Chang
158:120142
Lu H, Li Y, Chen M, Kim H, Serikawa S (2018) Brain Intelligence: go beyond artificial intelligence.
Mobile Netw Appl 23:368–375
Luo Y, Bu J (2016) How valuable is information and communication technology? A study of emerging
economy enterprises. J World Bus 51:200–211
Machek O (2014) The relationship between financial performance and total factor productivity: evidence
from the Czech agricultural sector. Int J Soc Sci Humanit Stud 6:57–67
Martínez D, Rodríguez J, Torres JL (2008) The productivity paradox and the new economy: the Spanish
case J. Macroecon 30:1569–1586
Meng H, Zhang M (2021) How artificial intelligence affects labor employment demand: empirical evi-
dence from chinese enterprises. J Xi’an Jiaotong Univ (Soc Sci) 41, 65–73+93 [in Chinese].
Michaels G, Natraj A, Van Reenen J (2014) Has ICT polarized skill demand? Evidence from eleven
countries over twenty-five years. Rev Econ Stat 96:60–77
Mitchell TM, Mabadevan S, Steinberg LI (1990) LEAP: A learning apprentice for VLSI design, Machine
learning. Elsevier, pp. 271-289.
Mollisi V, Rovigatti G (2017) Theory and practice of TFP estimation: the control function approach using
Stata. SSRN Electron J. https://doi.org/10.2139/ssrn.2916753
Nguyen QP, Vo DH (2022) Artificial intelligence and unemployment: an international evidence. Struct
Chang Econ Dyn 63:40–55
Olan F, Ogiemwonyi Arakpogun E, Suklan J, Nakpodia F, Damij N, Jayawickrama U (2022) Artificial
intelligence and knowledge sharing: contributing factors to organizational performance. J Bus Res
145:605–615
Oliner SD, Sichel DE (2000) The resurgence of growth in the late 1990s: is information technology the
story? J Econ Perspect 14:3–22
Olley S, Pakes A (1992) The dynamics of productivity in the telecommunications equipment industry.
National Bureau of Economic Research Cambridge, Mass., USA.
Pan M, Bai M, Ren X (2022) Does internet convergence improve manufacturing enterprises’ competitive
advantage? Empirical research based on the mediation effect model. Technol Soc 69:101944
Papaioannou SK, Dimelis SP (2017) Does upstream regulation matter when measuring the efficiency
impact of information technology? Evidence across EU and US industries. Inf Econ Policy
41:67–80
Patel V, Shah M. (2021) A comprehensive study on artificial intelligence and machine learning in drug
discovery and drug development. Intell Med
Peng Y, Tao C (2022) Can digital transformation promote enterprise performance? —From the perspec-
tive of public policy and innovation. J Innov Knowl 7:100198
Poole DL, Mackworth AK (2010) Artificial Intelligence: foundations of computational agents. Cam-
bridge University Press
Rathore AS, Nikita S, Thakur G, Mishra S (2022) Artificial intelligence and machine learning appli-
cations in biopharmaceutical manufacturing. Trend Biotechnol. https://doi.org/10.1016/j.tibtech.
2022.08.007
Romer PM (1990) Endogenous technological change. J Polit Econ 98:S71–S102
Sanchez JIL, Rata BM, Duarte AR, Sandulli FD (2006) Is the Internet productive? A firm-level analysis.
Technovation 26:821–826
Schuh G, Blum M, Reschke J, Birkmeier M (2016) Der digitale schatten in der auftragsabwicklung.
Zeitschrift Für Wirtschaftlich Fabrikbetrieb 111:48–51
Schumpeter, J.A., 1942. Socialism, capitalism and democracy. Harper and Brothers
Sharma R, Singh G, Sharma S (2020) Modelling internet banking adoption in Fiji: a developing country
perspective. Int J Inf Manage 53:102116
Simar L, Wilson PW (2000) A general methodology for bootstrapping in non-parametric frontier models.
J Appl Stat 27:779–802
Skorupinska A, Torrent-Sellens J (2017) ICT, innovation and productivity: evidence based on eastern
european manufacturing companies. J Knowl Econ 8:768–788
Song M, Peng L, Shang Y, Zhao X (2022) Green technology progress and total factor productivity of
resource-based enterprises: a perspective of technical compensation of environmental regulation.
Technol Forecast Soc Chang 174:121276
13
Tang H-L, Liu J-M, Wu J-G (2020) The impact of command-and-control environmental regulation on
enterprise total factor productivity: a quasi-natural experiment based on China’s “Two Control
Zone” policy. J Clean Prod 254:120011
Tang X, Ding S, Gao X, Zhao T (2022) Can digital finance help increase the value of strategic emerg-
ing enterprises? Sust Cities Soc 81:103829
Torrent-Sellens J, Díaz-Chao Á, Miró-Pérez A-P, Sainz J (2022) Towards the tyrell corporation? Dig-
itisation, firm-size and productivity divergence in Spain. J Innov Knowl 7:100185
Trajtenberg, M., 2018. AI as the next GPT: a political-economy perspective. national bureau of eco-
nomic research.
Triplett JE (1999) The Solow productivity paradox: what do computers do to productivity? Can J
Econ-Rev Can Econ 32:309–334
Tung K (2019) AI, the internet of legal things, and lawyers. J Manag Anal 6:390–403
Van Biesebroeck J (2007) Robustness of productivity estimates. J Indust Econ 55:529–569
Vogel A, Wagner J (2010) Exports and profitability–first evidence for German business services
enterprises. Appl Econ Q 56:7–30
Wang H, Ding L, Guan R, Xia Y (2020) Effects of advancing internet technology on Chinese employ-
ment: a spatial study of inter-industry spillovers. Technol Forecast Soc Chang 161:120259
Wang K-L, Sun T-T, Xu R-Y, Miao Z, Cheng Y-H (2022a) How does internet development pro-
mote urban green innovation efficiency? Evidence from China. Technol Forecast Soc Chang
184:122017
Wang Z, Li M, Lu J, Cheng X (2022b) Business Innovation based on artificial intelligence and block-
chain technology. Inf Process Manage 59:102759
Wen H, Wen C, Lee C-C (2022a) Impact of digitalization and environmental regulation on total factor
productivity. Inf Econ Policy 61:101007
Wen H, Zhong Q, Lee C-C (2022b) Digitalization, competition strategy and corporate innovation: evi-
dence from Chinese manufacturing listed companies. Int Rev Financ Anal 82:102166
Wooldridge JM (2009) On estimating firm-level production functions using proxy variables to control
for unobservables. Econ Lett 104:112–114
Wu H, Hao Y, Ren S, Yang X, Xie G (2021) Does internet development improve green total factor
energy efficiency? Evid China Energy Policy 153:112247
Xiao Y, Liu S (2019) Collaborations of industry, academia, research and application improve the
healthy development of medical imaging artificial intelligence industry in China. Chin Med Sci
J 34:84–88
Xie M, Ding L, Xia Y, Guo J, Pan J, Wang H (2021) Does artificial intelligence affect the pattern of skill
demand? Evidence from Chinese manufacturing firms. Econ Model 96:295–309
Yang C-H (2022) How artificial intelligence technology affects productivity and employment: firm-level
evidence from Taiwan. Res Policy 51:104536
Yang S-X, Zhu C-X, Qiao L, Chi Y-Y (2020) Dynamic assessment of Energy Internet’s emission reduc-
tion effect – a case study of Yanqing. Beijing J Clean Prod 272:122663
Zeng H, Ran H, Zhou Q, Jin Y, Cheng X (2022) The financial effect of firm digitalization: evidence from
China. Technol Forecast Soc Chang 183:121951
Zhang D, Liu D (2017) Determinants of the capital structure of Chinese non-listed enterprises: is TFP
efficient? Econ Syst 41:179–202
Zhang W, You J, Lin W (2021a) Internet plus and China industrial system’s low-carbon development.
Renew Sust Energ Rev 151:111499
Zhang Y, Li X, Song Y, Jiang F (2021b) Can green industrial policy improve total factor productivity?
Firm-level evidence from China. Struct Change Econ Dyn 59:51–62
Zhao Y, Zhou Y (2022) Measurement method and application of a deep learning digital economy scale
based on a big data cloud platform. J Organ End User Comput 34:1–17
Zhao C, Chen S, Cao W (2020) Internet plus information disclosure: substantive statement or strategic
manipulation——evidence based on the risk of stock price crash. China Ind Econ 3:174–192 ([In
Chinese])
Zhao P, Gao Y, Sun X (2022) How does artificial intelligence affect green economic growth?—Evidence
from China. Sci Total Environ 834:155306
Zheng X, Wu C, Nepal R (2022) Did the administrative approval reform in China affect the productivity
of energy firms? – A quasi-natural experimental approach. Energy Econ 105:105564
Zhou S, Zhou P, Ji H (2022) Can digital transformation alleviate corporate tax stickiness: the mediation
effect of tax avoidance. Technol Forecast Soc Chang 184:122028
13
Zhu, J., Li, D., 2019. Is technological innovation the effective way to achieve “double dividend” of envi-
ronmental protection and employment Growth?——An empirical research based on the mediating
effect model of 34 subdivided industries in China. China soft sci 1–13 [In Chinese].
Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps
and institutional affiliations.
Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under
a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted
manuscript version of this article is solely governed by the terms of such publishing agreement and
applicable law.
13

AI对TFP的影响

Uploaded by

Copyright:

Available Formats

AI对TFP的影响

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

AI对TFP的影响

Uploaded by

Copyright:

Available Formats

Economic Change and Restructuring (2023) 56:1113–1146

The impact of artificial intelligence on total factor

Ke‑Liang Wang1 · Ting‑Ting Sun1 · Ru‑Yu Xu1

Keywords Total factor productivity (TFP) · Artificial intelligence (AI) ·

universality (Trajtenberg, 2018), posing an extremely profound impact on economic

As the main participants in the economic activities of a country or a region,

3 Theoretical mechanism and research hypotheses

3.1 AI and enterprise TFP

productivity by decreasing the marginal production cost, enabling enterprises’ R&D,

Hypothesis 1 AI can improve the TFP of manufacturing enterprises.

3.2 AI, technological innovation and enterprise TFP

To be specific, firstly, AI can increase innovative knowledge. depending on

Hypothesis 2 AI can promote manufacturing enterprises to improve their techno-

3.3 AI, human capital optimization and enterprise TFP

ACEMOGLU (1996) stated that the maximum marginal productivity can be

Hypothesis 3 AI can improve the TFP of manufacturing enterprise by optimizing

3.4 AI, market matching amelioration and enterprise TFP

Currently, consumers’ demand has shown the characteristics of alienation, refine-

Fig.2 Mechanism analysis diagram

production factors in the workshop can be allocated effectively, and resource

Hypothesis 4 AI can improve the TFP of manufacturing enterprise by improving

The transmission mechanism of AI affecting manufacturing enterprise TFP is

4 Research design, variables selection and data description

4.1.1 Benchmark regression model

To quantitively examine the impact of AI on the TFP of China’s manufacturing

where i represents manufacturing listed companies; t represents year; TFP is the

4.1.2 The mediating effect model

where M is the mediating variable, which includes the variables of technological

4.2.1 The explained variable: enterprise TFP

4.2.2 The core explanatory variable: AI

Table 1 Definition of main variables

Founding age Age Ln (firm age + 1)

comparable. Moreover, manufacturing industry accounts for a higher proportion of

5 Empirical results and discussions

5.1 Benchmark regression results

Fig.3 Correlation coefficient of the main variables

5.2 Endogenous treatment: IV method

Table 3 Benchmark regression Variables Pooled-OLS FE FE FE

Table 4 Results of the IV Variables First-stage regression Second-

Table 5 Results of robustness tests

(1) (2) (3) (4) (5) (6) (7) (8)

AI 0.0745*** 0.0431*** 2.8098*** 0.0203*** 0.0694*** 0.0699*** 0.0711*** 0.0678***

(2) Substitution of explanatory variables. We employ the frequency of the sum of AI

denotes the number of employees in industry i in 2010. And ManuPWPt=2011

It can be seen in columns (1)-(8) of Table 5, AI coefficients in the four men-

AI 0.0763*** 0.0567*** 0.0479* 0.0582*** 0.0711*** 0.0839*** 0.0188 0.0725*** 0.0752***

5.4.3 Business life cycle heterogeneity

The impact of AI on TFP of manufacturing enterprises in different business life

5.4.4 Enterprise ownership heterogeneity

5.5 Transmission mechanism analysis

Table 7 Results of transmission mechanism analysis

AI 0.0736*** 0.0585** 0.0717*** 0.0063** 0.0733*** − 0.0200** 0.0704***

significantly improved. Furthermore, it can be found that the coefficient of AI in col-

5.6 Further analysis: exogenous policy shock effect

5.6.1 Continuous DID model

The Ministry of industry and information technology of China selected a batch of

5.6.2 Parallel trend test

Fig.4 Results of the parallel trend test

5.6.3 Policy effect estimation

To diminish the effect of random factors on DID regression results, referring to

3 Theoretical mechanism and research hypotheses

3.1 AI and enterprise TFP

3.2 AI, technological innovation and enterprise TFP

3.3 AI, human capital optimization and enterprise TFP

3.4 AI, market matching amelioration and enterprise TFP

4 Research design, variables selection and data description

4.1.1 Benchmark regression model

4.1.2 The mediating effect model

4.2.1 The explained variable: enterprise TFP

4.2.2 The core explanatory variable: AI

5 Empirical results and discussions

5.1 Benchmark regression results

5.2 Endogenous treatment: IV method

AI 0.0745* 0.0431* 2.8098* 0.0203* 0.0694* 0.0699* 0.0711* 0.0678*

AI 0.0763* 0.0567* 0.0479* 0.0582* 0.0711* 0.0839* 0.0188 0.0725* 0.0752***

5.4.3 Business life cycle heterogeneity

5.4.4 Enterprise ownership heterogeneity

5.5 Transmission mechanism analysis

AI 0.0736* 0.0585 0.0717* 0.0063 0.0733* − 0.0200 0.0704***

5.6 Further analysis: exogenous policy shock effect

5.6.1 Continuous DID model

5.6.2 Parallel trend test

5.6.3 Policy effect estimation

6 Conclusions and policy implications