Sage Journals: Discover world-class research

Abstract

Export technical complexity is an effective indicator to measure the upgrading of a country’s trade structure and high-quality development of trade. This study examines the impact of artificial intelligence on the export technical complexity upgrading of Chinese manufacturing enterprises at the enterprise and industry levels. It analyses matched data from the China Customs and Chinese Industrial Enterprises databases from 2000 to 2015. The results show that artificial intelligence promotes the upgrading of export technical complexity in Chinese manufacturing enterprises. The impact of artificial intelligence on upgrading export technical complexity is dynamic, showing an “inverted U-shape.” Heterogeneity test results indicate that artificial intelligence has a greater effect on export technical complexity upgrading if it features high-tech complexity, or the enterprises are local or in the eastern region of China. The mechanism results indicate that artificial intelligence significantly promotes labor productivity and innovation to upgrade China’s manufacturing enterprises’ export technical complexity. Finally, industry-level analysis shows that the application intensity of artificial intelligence substantially affects the average export technical complexity upgrades of Chinese manufacturing enterprises through spillover effects and resource reallocation effects. These findings promote upgrading export technical complexity at the enterprise and industry levels.

Keywords

artificial intelligence export technical complexity spillover effects resource reallocation effects China

Introduction

New technologies, production methods, and products represented by Big Data, artificial intelligence, and industrial robots are emerging in large quantities, becoming an essential driving force for a new scientific and technological revolution and industrial change. Artificial intelligence, as a high-level automation technology, highlights integration with other industries with a primary focus on promoting intelligent manufacturing development, the most significant feature of robot participation in production activities. According to the International Federation of Robotics (IFR), since 2003, China’s industrial robot installations have developed rapidly, with an average annual growth rate over 20% (Figure 1), and have ranked first worldwide for several years. From the global industrial robot stock perspective, by 2019, China’s industrial robot stock reached nearly 800,000 units (Figure 2), accounting for approximately 30% of all global industrial robots. In 2016, China surpassed Japan for the first time as the country with the largest stock of industrial robots worldwide.

Figure 1.

Annual growth rate of industrial robots of major countries, 2000 to 2018.

Figure 2.

Stock of industrial robots in major countries, 2000 to 2019.

Artificial intelligence is the critical key link and essential foundation for connecting intelligent manufacturing and industrial applications; thus, the rapid development of artificial intelligence in China will likely profoundly impact the country’s current and future intelligent manufacturing and industrial applications. China’s manufacturing industry has for some time been in the low-cost link of processing and assembly in the global value chain, and there is a risk of being “locked” and “solidified” by the low end of production (D. Li & Zhu, 2019; Song et al., 2022), the level of industrial development is low, and the technical content of products is not high either. The external environment is becoming more complex and uncertain, especially in the face of the triple pressure of domestic and foreign demand contraction, contraction of supply shock, and weakening of market expectations. Accelerating trade and product structure upgrades, improving the technical level of China’s manufacturing industry and export technical complexity, and realizing the transformation of Made in China to “Created in China” may prompt the driving industrial force to promote economic growth and high-quality development in the new era. Therefore, this study examines the impact and action mechanism of artificial intelligence on the export technical complexity in China to provide a policy reference for Chinese manufacturing enterprises to upgrade their technical complexity.

Literature Review

Impact of Artificial Intelligence

Since their inception, artificial intelligence has widely attracted attention from economists, while extant literature regarding the impact of artificial intelligence on the economic field mainly focus on topics such as economic growth (L. Wang et al., 2021), the labor market (Acemoglu & Restrepo, 2020; DeCanio, 2016; Kromann et al., 2011), income distribution (Lankisch et al., 2019), and energy performance (G. Huang et al., 2022; E. Z. Wang et al., 2022; Yin and Zeng, 2023). For example, L. Wang et al. (2021) developed a theoretical framework and explored the effects of artificial intelligence on capital, wages, and economic growth, and found that an increase in artificial intelligence capital results in lower capital prices, a higher wage, and increased productivity. Acemoglu and Restrepo (2020) studied the effects of industrial robots on US labor market. They illustrated theoretically and empirically that robots may reduce employment and wages. Yin and Zeng (2023) examined the impact of industrial intelligence on energy intensity in China, and showed that industrial intelligence reduced energy intensity.

With the in-depth development of artificial intelligence, some scholars have begun to focus on the impact of artificial intelligence on international indivision of labor and trade (Alguacil et al., 2022; Artuc et al., 2018; Destefano et al., 2019; Faber, 2020; Goldfarb & Trefler, 2018; Krenz et al., 2021; Stapleton & Webb, 2020). For example, Artuc et al. (2018) studied the impact of industrial robots on north–south trade from theoretical and empirical aspects on the basis of Ricardo’s two-stage production and trade model, and the results showed that industrial robots promoted import and export growth between developed and developing countries by reshaping comparative advantages. In contrast, Berg et al. (2016), Rodrik (2018), Krenz et al. (2021), Destefano et al. (2019), and Faber (2020) believe that the application of industrial robots has eroded the developing countries’ labor cost advantages, affected the international division of labor, promoted the return of manufacturing to developed countries, reduced their employment opportunities, and caused declines in offshoring and trade shares, ultimately aggravating inequality. However, at the enterprise-level, researchers such as Stapleton and Webb (2020) have also found that the application of industrial robots can actually scale up production, increase productivity, and help increase imports from developing countries and the number of subsidiaries.

Part of the related literature about robots and trade has highlighted on the effects of robots on export performance (Alguacil et al., 2022; Zhang et al., 2023). For example, Alguacil et al. (2022) analyzed robot adoption and export performance in Spain at the firm level and pointed that firms adopting robots have experienced a sharp increase in their export probability, export sales, and share of exports in total output. Zhang et al. (2023) further discussed Chinese industrial firms’ export behavior to state that robot adoption made it easier to win export competition and increase market share just for large-scale firms, while small and medium-sized enterprises (SMEs) did not seem to enjoy any benefits from adoption.

Determinants and Effects of Export Technical Complexity

Meanwhile, extensive research has also been conducted on export technical complexity, positing that the export technical complexity reflected the characteristics of productive knowledge and technology distribution contained in a country’s export products; therefore, it is an effective indicator to measure the upgrading of a country’s trade structure (Hausmann et al., 2007) and high-quality development of trade. After this indicator was first proposed by Hausman et al. (2007), studies in this direction began to proliferate. To date, this research has mainly focused on the determinants and effects of export technical complexity. The relevant literature has also posited that macro-level determinants, such as regional human capital enhancement and government supportive policies (Z. Wang & Wei, 2007), trade liberalization (Nguyen, 2016), FDI (Xu & Lu, 2009), cultural diversity (Fan et al., 2018), improved financial services (Qi & Xiang, 2020), scientific research investment (C. Q. Li & Lu, 2017), factor structure improvement (Yu & Hu, 2015), agricultural technology (Bai, 2020), population aging (Wu et al., 2022), and cleaner production standards (Ju et al., 2022) can facilitate upgrades to a country or region’s export technical complexity. However, lower technological levels, excessive export prices, and corruption will hinder efforts to upgrade export technical complexity (Cabral and Veiga, 2010; Hallak and Schott, 2011; Willem & Pai, 2015). In terms of micro-level determinants, institutional similarities and corporate heterogeneity (Demir & Hu, 2021), labor productivity, capital stock and infrastructure (S. Huang et al., 2014), product density (Ding & Li, 2018), and digital capabilities (Banga, 2022) have demonstrated significant effects on export technical complexity. In addition, the studies on the effects of export technical complexity showed that the upgrading of export technical complexity not only helped to promote the upgrading of export product quality and productivity improvement (Mishra et al., 2011; Zhu & Fu, 2013), which can also significantly promote economic growth (Jarreau & Poncet, 2012).

Artificial Intelligence and Export Product Quality

Closer to the topic of the present research is the impact of artificial intelligence on export product quality. For example, DeStefano and Timmis (2021) primarily discussed the impact of artificial intelligence application, such as industrial robots, on the export quality of developed and underdeveloped countries. In addition, Hong et al. (2022) focused on the impact of industrial robots on upgrading Chinese enterprises’ export products, and examined the impact of industrial robots on export product quality upgrades from the perspective of enterprise cost changes and structural effects of labor allocation. They found that there was a U-shaped relationship between the use of industrial robots and export products’ quality upgrading, most enterprises have not yet realized the quality upgrading of export products. Lin et al. (2022) used the PSM-DID method to examine the impact of industrial robots at the enterprise level on the quality upgrading of Chinese export trade and found that the impact caused by the application of industrial robots on the export product quality is dynamic, which increases initially and then decreases.

Research Gaps

Some research gaps exist in the literature. First, the existing studies on industrial robots and artificial intelligence have focused on economic growth, labor market, energy performance, and international division of labor, leaving its trade performance impact largely ignored. Compared with the few research on the trade scale and quality effects, this research mainly focuses on the impact of artificial intelligence on export structure, addressing the gap in existing research on the impact of trade performance. Second, the previous literature on the determinants and effects of export technical complexity has ignored the link between emerging artificial intelligence technologies and export technical complexity. Finally, few studies have explored the impact of artificial intelligence at the enterprise and industry levels.

Based on the above, the present study contributes to the literature in several ways. First, to the best of our knowledge, our study is the first to investigate the impact and action mechanism of artificial intelligence on manufacturing enterprises’ export technical complexity. To this end, we use the latest matched data from the China Customs and Chinese Industrial Enterprises databases from 2000 to 2015 to identify the impact of artificial intelligence. Artificial intelligence technologies, such as industrial robots, are widely used and popular in developing countries, for example, in China, according to data from the IFR, the number of industrial robots has rapidly increased since 2003. Their average annual growth rate has exceeded 20% for nearly 20 consecutive years, ranking first globally. By 2019, the stock of industrial robots in China accounted for almost 30% of the global stock, reaching nearly 800,000 units, surpassing Japan for the first time in 2016, and becoming the country with the largest stock of industrial robots worldwide. However, previous studies have mainly concentrated on developed countries. The present study focuses on China, providing new empirical evidence on emerging developing countries. Second, unlike most previous research, this study empirically addresses the enterprise and industry levels and examines the impact of artificial intelligence on export technical complexity at the industry level through spillover and resource reallocation effects. Finally, this study used the propensity score matching (PSM) and difference-in-difference (DID) methods for baseline estimation and the instrumental variable method for robustness testing. After controlling for endogenous problems, such as two-way causality and missing variables, the impact of artificial intelligence on Chinese manufacturing enterprises’ export technical complexity is more accurately identified.

The remainder of this paper is organized as follows. Section “Theoretical Mechanisms” provides a theoretical analysis of the impact mechanism of artificial intelligence on export technical complexity. Section “Materials and Methods” describes the research design. Section “Results” presents the results of empirical analysis, while Sections “Enterprise-Level Further Discussion” and “Industry-Level Further Discussion” provide the enterprise-level analysis and industry-level expansion analysis, respectively. Finally, Section “Conclusions and Implications” offers the conclusion and policy implications.

Theoretical Mechanisms

Based on a review of relevant literature, this study mainly analyses the mechanism of artificial intelligence affecting the export technical complexity upgrading of Chinese manufacturing enterprises from the perspective of labor productivity and innovation.

Labor Productivity

Artificial intelligence can promote labor productivity by deepening capital and improving human capital. Regarding capital deepening, in the Abel-Blanchard model (Madsen, 2010), technological advances increase the expected return on capital stock, which leads to capital deepening. As typical representatives of technological advancement, artificial intelligence has dual roles in both technological progress and capital deepening. Technological progress and capital deepening are mutually integrated and inseparable; thus, capital investment increases when technological advances lead to new production equipment. According to neoclassical growth theory, capital deepening changes in the same direction as labor productivity, and since every capital good embodies the latest technology at the moment of its construction (Phelps, 1962), robotic automated production and artificial intelligence promote upgraded labor productivity (Graetz & Michaels, 2015). Therefore, with a large investment in new equipment, enterprises achieve technological developments at an accelerated speed, thereby improving labor productivity.

In terms of human capital improvement, the advancement of artificial intelligence increases the related demand for highly skilled labor, creating higher requirements for the labor force’s skills. Hémous and Olsen (2022) have found that artificial intelligence can reduce the demand for low-skilled labor while increasing labor productivity and thereby increasing the demand for highly skilled labor, their results indicate complementary and substitution effects between artificial intelligence and labor. In addition, the integration and development of artificial intelligence in different industries lead to new formats and models, which, in turn, promote the transformation and upgrading of industrial structure. Such transformations and upgrades to the industrial structure not only promote the flow of human capital but also advance higher requirements for human capital, inevitably increasing the demand for highly skilled labor (Michaels et al., 2014), thereby directly promoting improvements in human capital. Moreover, export technical complexity reflects a country’s export products’ technical content and production efficiency (Rodrik, 2006). Thus, we propose hypothesis 1 (H1).

H1. Artificial intelligence can promote export technical complexity through improved labor productivity.

Enterprise Innovation

The enterprise innovation effect generated by artificial intelligence may help Chinese manufacturing enterprises upgrade their export technical complexity. First, the penetration and transformation of artificial intelligence have greatly improved enterprises’ information-sharing and processing capabilities, allowing enterprises to break through the restrictions of industry and geography on “cognitive distance,” and significantly enhance the efficiency of enterprise information searches. These improvements may reduce uncertainty in the innovation process, promote new knowledge discovery, and contribute to improved innovation ability. Second, the productivity effect of artificial intelligence may reduce production and operation costs (Javaid et al., 2021), which saves enterprise funds. Therefore, an enterprise may invest more capital in complex and high-end technology research and development (R&D), alleviating the financing constraints in the R&D process and improving the enterprise’s level innovation. Third, artificial intelligence induces an innovative effect of “learning by doing.”Arrow (1962) has found that investment in fixed assets in the production field can generate knowledge and experience through “learning by doing,” which leads to improved production efficiency. As a physical form of innovative results relative to general fixed asset investment, industrial robots are easier to apply, imitate, and learn in practice due to their science and technology characteristics, integration, and intelligence. The resulting technological progress and labor productivity promotion are generally significant. Moreover, enterprise innovation is an essential factor that affects export technical complexity (Zhu and Fu, 2013). Hence, we propose Hypothesis 2 (H2).

H2. Artificial intelligence may help enterprises upgrade their export technical complexity through the effect of enterprise innovation.

Materials and Methods

Data

Considering the IFR data is only industry-level data, and this study investigates the impact of artificial intelligence on the export technical complexity at the enterprise level, hence, we use the imported robots data from China Customs database to characterize artificial intelligence, however, the latest data from China Customs Database and Chinese Industrial Enterprises Database are up to 2015, therefore, this study uses the above databases from 2000 to 2015 based on data availability. For the Chinese Industrial Enterprises Database (2000–2015), we refer to Cai and Liu (2009) data processing method by excluding the following: (a) samples missing variables such as employees, total assets, fixed assets, sales, and industrial output value; (b) samples with fewer than eight employees; (c) samples where the total assets were less than the current assets or the total fixed assets were greater than the total assets; (d) samples with paid-in capital less than or equal to 0; and (e) samples with enterprise age is less than 0 years or greater than 100 years.

Using the China Customs Database (2000–2015), we could calculate the export technical complexity of Chinese manufacturing enterprises. We processed the data according to the following steps: (a) exclude the samples with missing information such as enterprise names, export location names, and product names; (b) exclude the samples with a single trade size of less than $50 or a quantity unit of less than 1; (c) add monthly data to annual data; (d) exclude the samples of trading intermediary firms; (e) only keep the samples with the largest number of counting units under the same product code; (f) add Harmonized System (HS) 8-digit level data to HS 6-digit level data; (g) convert the product data into HS07 level data; (h) exclude the non-manufacturing samples; (i) exclude the primary and resource products; and (j) exclude the homogeneous product samples.

Finally, we initially matched the robot import data and export data of export technical complexity by enterprise name code. Then, we matched the data from both databases by enterprise name, and matched the remaining data by each enterprise’s postal code and the last seven digits of the telephone number. We obtained 307,622 observations, which included 86,366 enterprises, of which 20,688 were robot import enterprises, and 65,678 were non-imported robot enterprises.

Baseline Model

To alleviate the endogenous problems caused by selective bias and missing variables, we initially matched the treatment group to a similar control group using the PSM method proposed by Heckman et al. (1997), and then adopted the multi-period DID method for estimation. Regarding the specific PSM method, first, we divided the samples into two groups: robot importing enterprises (treatment group) and robot non-importing enterprises (control group). Second, to meet the principle of conditional independence, we selected the variables that affected the export technical complexity of enterprises and whether to import robots as covariates for matching. The final covariates included the following: whether it was a general trading enterprise, the total factor productivity of enterprise, and the enterprise size. The higher-order terms of some variables were used as matching variables. Then, we selected the caliper K neighbor method for matching, a matching ratio of 1:5, and a caliper choice of 0.05. Finally, after matching, we obtained 257,044 observations, 85,628 observations in the treatment group, and 171,416 observations in the control group. Among them, the number of enterprises in the treatment group was 20,548, and the number in the control group was 57,922.

Although the PSM method can control for observable differences in firm characteristics between treatment and control groups, the unobservable differences between groups and missing variables may lead to estimation biases. Therefore, referring to Angrist and Pischke (2014), we also used multi-period DID to build a measurement model to investigate how artificial intelligence affects the export technical complexity upgrading of Chinese manufacturing enterprises. We use robot import as a proxy variable for artificial intelligence. The econometric model used for multi-period DID is as follows:

\begin{matrix} \ln {etc}_{it} = α_{1} + α_{2} {treat}_{it} \times {post}_{it} + α_{3} x_{it} + α_{4} z_{jt} \\ + v_{i} + v_{t} + ε_{ijt} \end{matrix}

(1)

where $i$ denotes enterprise, $j$ is four-digit industry code in the national economic industry classification (CIC), $lnetc$ is the export technical complexity at the enterprise level, and $treat$ is a dummy variable representing robot-importing enterprises. If enterprise $i$ imports robots in the sample period, the enterprise belongs to the treatment group, and the dummy variable equals one, otherwise it is zero. $post$ is a dummy variable representing the period variable of whether robots have been imported; if enterprise $i$ imported robots in the current year, the value of the current year and subsequent years would be one, and zero otherwise. $x$ represents the enterprise-level control variables, $z$ represents the industry-level control variables, $v_{i}, v_{t}$ represent the fixed effects at the enterprise and year levels, respectively, and $ε_{ijt}$ is the random error term. To eliminate heteroscedasticity and autocorrelation problems as much as possible, we clustered all regressions at the enterprise level.

Variables

Export Technical Complexity

In view of this study’s aim to examine the impact of artificial intelligence on the export technical complexity upgrading of Chinese manufacturing enterprises, comparing export technical complexity across borders was unnecessary. Therefore, referring to the calculation methods of Xu and Lu (2009), we measured Chinese manufacturing enterprises’ export technical complexity using two steps. First, we measured the technical complexity of the Chinese product level, using the following equation:

t c_{l} = \sum_{p = 1}^{n} \frac{x_{pl} / X_{p}}{\sum_{p} x_{pl} / X_{p}} Y_{p}

(2)

where $tc$ is technical complexity at the product level, $x_{pl}$ is the export value of product $l$ in province $p$ , $X_{p}$ is the export value of province $p$ , and $x_{pl} / X_{p}$ is the export share of product $l$ in province $p$ . $\sum_{p} x_{pl} / X_{p}$ is the sum of export shares of products $l$ from all provinces in China. $Y_{p}$ is the real GDP per capita of province $p$ . The export technical complexity of product $l$ is a weighted average of the real per capita GDP of each province weighted by the revealed competitive advantage (RCA) index of the product by each province.

Second, we summed the export technical complexity at the product level obtained in the first step at the enterprise level. The specific equation is as follows:

{etc}_{i} = \sum_{l} (x_{il} / X_{i}) t c_{l}

(3)

where $x_{il}$ represents the export value of product $l$ of enterprise $i$ , $X_{i}$ is the total export volume of enterprise $i$ , and $x_{il} / X_{i}$ is the export share of product $l$ of enterprise $i$ . We summed the export technical complexity at the enterprise level, and then processed naturally logarithmically.

Considering economic development levels vary greatly among provinces in China, the difference in the quality of the same export product between different provinces will lead to differences in the export technical complexity of the product between the provinces. Therefore, following Xu (2010), we first added product quality adjustments to calculate the export technical complexity at the product level in the first step. The specific practice was initially to calculate the export quality of each province at the product level, according to the China Customs Database. Second, we multipled $λ$ -power of export quality at the product level for each province with export technical complexity obtained in the first step. Then, we could obtain the adjusted export technical complexity at the product level for each province. Finally, we summed the adjusted export technical complexity at the product level according to the weight of export share of the products at the enterprise level, and then could obtain the adjustment value of export technical complexity at the enterprise level. The formula for measuring and adjusting export quality at the product level for each province is shown in Equations (4) and (5), as follows:

qualit y_{pl} = pric e_{pl} / \sum_{p = 1}^{n} (x_{pl} / X_{l} \times pric e_{pl})

(4)

{tc}_{pl}^{adj} = (qualit y_{pl})^{λ} \times t c_{l}

(5)

In Equation (4), $qualit y_{pl}$ indicates the product quality of the product $l$ in province $p$ , $pric e_{pl}$ indicates the export price of the product $l$ in province $p$ , and $x_{pl} / X_{l}$ is the export proportion of product $l$ in province $p$ accounting for the whole country; that is, we used the relative export price of product $l$ of province $p$ to indicate its quality level. In Equation (5), ${tc}_{pl}^{adj}$ indicates the export technical complexity of product $l$ of province $p$ after product quality adjustment, and $λ$ takes a value of 0.2, with reference to Xu (2007). We further summed the export technical complexity at the product level of each province after adjustment to the enterprise level, and the weight was consistent with Equation (3). Finally, we obtained the adjustment value of the export technical complexity of Chinese manufacturing enterprises for robustness testing. The equation is as follows:

{etc}_{i}^{adj} = \sum_{l} (x_{il} / X_{i}) {tc}_{pl}^{adj}

(6)

Artificial Intelligence

Artificial intelligence is the core explanatory variable for this study. Since industrial robots participation in production activities is one of the most significant features of artificial intelligence application, we take industrial robots application to represent artificial intelligence. According to data provided by the International Federation of Robotics, by 2015, over 70% of domestic industrial robots relied on imports, indicating that the import of industrial robots reflects industrial robots application to a certain extent. Moreover, various countries only provide industrial robot stock data at the industry level, and domestic industrial robots use data at the enterprise level has been temporarily unavailable. Therefore, following Acemoglu and Restrepo (2020), we used the import data of industrial robots as a proxy variable for artificial intelligence.

Industrial robots are mainly divided into seven categories: IC factory special automatic handling robots (HS848640), multi-functional robots (HS847950), laser welding robots (HS851580), handling robots (HS842890), spraying robots (HS842489), arc welding robots (HS851531), and resistance welding robots (HS851521). The first three are high-tech complexity industrial robots, and the last four are medium- and low-tech complexity industrial robots. In the sample period, if an enterprise imported any type of industrial robot in year $t$ , then the value of $t r e a t_{i t} \times p o s t_i t$ during the current year and subsequent years is one, otherwise it is zero. When an enterprise had several years of robot-importing behavior during the sample period, the year the enterprise first imported industrial robots and the subsequent years are selected and take a value of one. In view of the need for robustness testing, we also counted the accumulative import amount ( $lnrov$ ) and accumulative import volume ( $lnroq$ ) of industrial robots for enterprises using logarithmic processing.

Control Variables

Referring to previous literature on the influencing factors of export technical complexity, we selected eight variables as control variables at the enterprise level and industry levels. First, enterprise size ( $labor$ ) was characterized by the natural logarithm of the average annual number of employees at an enterprise. Second, enterprise age ( $age$ ), was characterized by the natural logarithm of the difference between the current year and the year the enterprise was opened. Third, we determined whether an enterprise was a foreign funded ( $type$ ); for foreign-funded enterprises, the value is one, otherwise it is zero. Fourth, we determined whether a firm was a general trade enterprise ( $tradetype$ ); for general trade enterprises, the value is one, otherwise it is zero. Fifth, we included the total factor productivity of enterprise ( $tfp$ ). We referred to Head and Ries (2003) and used the estimation method $tfp = \ln (y / l) - s \ln (k / l)$ to estimate the total factor productivity of an enterprise, where $y$ is characterized by the gross industrial output value, $l$ is characterized by the average annual number of employees, $k$ is characterized by the total value of fixed assets, and $s$ is the contribution rate of capital in the production function, which is valued at one-third. Sixth was enterprise capital intensity ( $inten$ ), characterized by the ratio of fixed assets to the number of employees divided by 100. Seventh was enterprise liquidity ( $constr$ ), characterized by the proportion of the difference between current enterprise assets and current enterprise liabilities to total enterprise assets. Eighth, to measure industry competition ( $hhi$ ), we used the Herfindahl-Hirschman Index, which is the sum of the squares of proportion of enterprise sales in CIC four-digit codes at the industry level. To eliminate the effects of price movements, we carried out the corresponding price reductions for all value variables in this study. Further, to eliminate the effect of outliers, we winsorized all continuous variables by 1% before and after. Table 1 shows the descriptive statistics of the main variables. The distribution of the value and the adjustment value of export technical complexity are concentrated. There are relatively few observations after importing robots from the mean, maximum, and minimum values of ${treat}_{it} \times {post}_{it}$ .

Table 1.

Descriptive Statistics of the Main Variables.

Explanatory variables	Variable meaning	Observations	Mean	Maximum	Minimum
$lnetc$	Export technical complexity	257,044	10.865	12.253	8.971
$lnet c^{adj}$	Adjustment value for export technical complexity	257,044	10.807	12.239	8.864
$treat_post$	The intersection of the group variable and the period variable of artificial intelligence	257,044	0.333	1	0
$lnrov$	Robot import amount	257,044	0.217	4.709	0
$lnroq$	Robot import quantity	257,044	0.175	3.584	0
$type$	Foreign funded enterprise	257,044	0.340	1	0
$tradetype$	General trade enterprise	257,044	0.341	1	0
$labor$	Enterprise size	257,044	5.742	8.59	3.045
$tfp$	Total factor productivity of enterprise	257,044	4.315	6.815	2.277
$age$	Enterprise age	257,044	2.189	3.871	0.693
$inten$	Enterprise capital intensity	257,044	3.760	6.997	0.137
$constr$	Enterprise liquidity	257,044	0.071	0.785	−1.765
$hhi$	Industry competition	257,044	0.016	0.139	0.0009

Results

Matching Balance Test Results

$PSM$ is actually a counterfactual analysis method. We could not observe the characteristics when the enterprises in the experimental group did not import industrial robot in reality; thus, it is necessary to use $PSM$ to select the samples closest to the imported industrial robot enterprises as the control group among the enterprises that have never imported industrial robots, to take the control group as a reference to obtain the impact of industrial robot importing on enterprises’ export technical complexity. To ensure the effectiveness of the matching of $PSM$ and estimation of $DID$ , we conducted a matching balance test for the matching of the treatment group and the control group (Table 2). From the differences in the means of each covariate between the treatment and control groups, the difference after matching compared with pre-matching was greatly reduced, and the t-test results showed that the p-value of each covariate was zero prior to matching. This indicated that the covariates significantly differed between the treatment and control groups; however, after matching, each covariate had a p-value greater than .1, indicating there were no significant differences between the treatment and control groups. Thus, the overall matching effect was ideal.

Table 2.

Matching Balance Test Results.

Explanatory variables	Before/after matching	Mean		Differences between treatment group and control group (%)	t-Test
Explanatory variables	Before/after matching	Treatment group	Control group	Differences between treatment group and control group (%)	t-Value	p-Value
$tradetype$	Before matching	0.378	0.302	16.1	42.19	.000
$tradetype$	After matching	0.378	0.377	0.1	0.31	.758
$tfp$	Before matching	4.371	4.311	6.6	17.30	.000
$tfp$	After matching	2.174	2.176	−0.5	−1.16	.245
$labor$	Before matching	5.965	5.527	39.3	104.38	.000
$labor$	After matching	5.964	5.962	0.2	0.37	.708

Note. To save space, Table 2 only lists the matching balance test results for univariate variables, and the higher-order term test results for other covariates are conclusive.

Baseline Results

In Table 3, the results are reported in Columns (1) to (2) based on the whole sample $DID$ , while the results in Columns (3) to (4) are based on $PSM - DID$ . Columns (1) and (3) do not control for any variables, Columns (2) and (4) control for enterprise-level variables and industry-level variables; each column controls for the fixed effects of the enterprise and year.

Table 3.

Baseline Results.

Explanatory variables	$DID$		$PSM - DID$
Explanatory variables	(1)	(2)	(3)	(4)
$treat_post$	0.0150*** (0.0029)	0.0134*** (0.0030)	0.0156*** (0.0033)	0.0137*** (0.0031)
$type$		0.0034*** (0.0013)		0.0026* (0.0014)
$tradetype$		−0.0016* (0.0009)		−0.0028** (0.0011)
$labor$		0.0085*** (0.0010)		0.0092*** (0.0012)
$tfp$		0.0062*** (0.0009)		0.0069*** (0.0009)
$age$		−0.0049*** (0.0017)		−0.0055*** (0.0019)
$inten$		0.450*** (0.0635)		0.564*** (0.0793)
$constr$		−0.0021 (0.0014)		−0.0021 (0.0016)
$hhi$		0.0208 (0.0284)		0.0214 (0.0311)
$constant$	10.83*** (0.0030)	10.75*** (0.0101)		10.75*** (0.0115)
Observations	279,224	279,224	228,904	228,904
$R^{2}$	.984	.982	.983	.983

Note. Robust standard errors are in parentheses. We controlled for firm and year fixed effects, which is the same as in the following tables. In addition, it is worth noting that 28,398 observations on the estimation of $DID$ and 28,140 observations on the estimation of $PSM - DID$ have been automatically excluded by the command $reghdfe$ respectively; thus, the observations of regressions are inconsistent with the observations of the previous statistics.

, **, and *** indicate significance at the 10%, 5%, and 1% levels, respectively.

The results showed that the impact of artificial intelligence was significantly positive at the 1% level, regardless of whether the estimation was based on $DID$ or on $PSM - DID$ . Further, regardless of whether the control variables were added, the significance of the estimated coefficients did not change. The results presented in Column (4), indicated that the export technical complexity of industrial robot importing enterprises had significantly improved, with an average increase of approximately 1.37% when compared with enterprises that have not imported industrial robots. The reason may be that the application of industrial robots contributes to upgrading export technological complexity by improving labor productivity and promoting innovation. This result is consistent with Graetz and Michaels (2015) and L. Wang et al. (2023).

Parallel Trend Test Results

A prerequisite for the validity of the regression results of $DID$ and $PSM - DID$ depends on the assumption of the parallel trend condition. That is, if the treatment group did not import robot, the trends for the export technical complexity of the treatment and control group should be parallel. If the treatment group has a certain difference with the control group before the import, then the result is not likely to be the net effect of industrial robot importing. In view of this, we further used the event study method to test parallel trends and the dynamic effect of the impact of industrial robots importing on export technical complexity over time. The regression equation constructed is as follows:

l n e t c_{i t} = β_{1} + β_{2} t r e a t_{i t}^{- 5} \times p o s t_{i t}^{- 5} + + β_{11} t r e a t_{i t}^{5} \times p o s t_{i t}^{5} + β_{12} x_{i t} + β_{13} z_{j t} + v_{i} + v_{t} + ε_i j t

(7)

where ${treat}_{it}^{n_{-}^{+}} \times {post}_{it}^{n_{-}^{+}}$ are a series of dummy variables. In n years before the robot was imported, the value of ${treat}_{it}^{- n} \times {post}_{it}^{- n}$ is one, otherwise it is zero. In n years after the robot was imported, the value of ${treat}_{it}^{+ n} \times {post}_{it}^{+ n}$ is one, otherwise it is zero. When the enterprise imported industrial robots in the current period, the value of ${treat}_{it} \times {post}_{it}$ is one, otherwise it is zero. In addition, we used the year before industrial robot importing as the reference group, and the coefficient of ${treat}_{it}^{n_{-}^{+}} \times {post}_{it}^{n_{-}^{+}}$ in the results is the significant difference of the export technical complexity trend between the treatment and control groups in the nth year before and after industrial robots were imported. To more intuitively show the trends of the coefficients before and after industrial robots were imported, we plotted the coefficients of Equation (7) into a parallel trend chart, as shown in Figure 3.

Figure 3.

Parallel trend test results.

In Figure 3, the horizontal axis represents the year before and year after industrial robots were imported, and the vertical axis represents the estimated coefficient of each year. It shows that the estimated coefficient of each year before the current industrial robot import period is within the 95% confidence interval, which is not significantly different with zero. This indicates that, before industrial robots were imported, there were no significant differences in the change trends for export technical complexity between the treatment and control groups. Thus, the parallel trend hypothesis was supported. When n was 0, …, 5, the estimated coefficient of each year displayed a trend of increasing first and then decreasing, showing an “inverted U-shape,” which was significantly positive at the 5% level. Thus, this showed, to a certain extent, that the impact of industrial robot importing on the export technical complexity is dynamic, which is specifically manifested as a continuous promotion effect. The estimated coefficient reached the maximum (0.0254) in the second period after industrial robots were imported, and then gradually decreased, but the decline was smaller, resulting in the estimated coefficient of the fifth year after industrial robots were imported still being large (0.0165). This may be related to the long service life of imported industrial robots and the lag impact of robot importing on intermediary variables, such as labor productivity. This result is consistent with the study of Zhang et al. (2023) and Lin et al. (2022) on the dynamic effect of industrial robot applications on export performance and the quality upgrade of export products.

Robustness Tests

Other Endogenous Issues

Although the estimation based on $PSM - DID$ could alleviate the endogenous problems caused by observable and unobservable variables, we still could not avoid other endogenous problems caused by missing variables and bidirectional causation. Therefore, we used the instrumental variable method for robustness testing. Referring to L. Li et al. (2021) and Hasan et al. (2009), we selected the robot import intensity in the industry in the current year as the instrumental variable for a two-stage least squares estimation. We used the proportion of industrial robot import enterprises in the current industry to represent the robot import intensity. The reason for this selection was that the robot import intensity in the industry can affect the decision-making of industrial robots import of enterprise through demonstration effects and competition effects, and then improve the possibility an enterprise will import industrial robots (L. Li et al., 2021), however, the explained variable of this study is the export technical complexity at the enterprise level, thus, the export technology complexity at the enterprise level will not affect the robot import intensity in the industry. (2) In line with Lin et al. (2022), we used the lag period of the accumulative amount of imported robots as the instrumental variable. The reason for this choice is that the lag period of import robots can affect the current import of industrial robots; however, the export technical complexity of the current period does not affect the lag period of import robots. The results for the instrumental variable method are shown in Columns (1) and (2) of Table 4. The first stage of the estimation results indicates that the instrumental variables are positively and significantly correlated with the industrial robot import of enterprise; thus, it meets the correlation condition of the instrumental variables and explanatory variable. The second stage estimation results showed that the impact of artificial intelligence on the export technical complexity of Chinese enterprises under the instrumental variable method remained significantly positive, and Kleibergen-Paap rk LM and Kleibergen-Paap rk Wald F-value results showed that the null hypothesis of the insufficient and weak recognition of instrumental variables was rejected at the 1% level, indicating that the instrumental variables were valid. After considering endogenous problems using the two-stage least squares method, the results remained robust.

Table 4.

Robustness Test Results.

Explanatory variables	Estimations using the instrumental variable method		Replacing the explanatory variable		Replacing the explained variable
Explanatory variables	(1)	(2)	(3)	(4)	(5)
$treat_post$	1.478*** (0.234)	0.149*** (0.022)			0.0119*** (0.0038)
$\ln rov / \ln roq$			0.0072*** (0.0009)	0.0045*** (0.0007)
Kleibergen-Paap rk LM (p-value)	48.266*** (0.000)	1,235.95*** (0.000)
Kleibergen-Paap rk Wald F	49.336	1,471.5
Observations	228,895	191,328	228,904	228,904	228,904
$R^{2}$			.983	.983	.972

Note. To save space, Table 4 reports only the estimated coefficients of the core explanatory variables, as do Tables 5 to 8.

Replacing Explanatory Variables

Borrowing from Acemoglu and Restrepo (2022) and Aaronson and Phelan (2019), we used the import amount and import volume of industrial robots as proxy variables for artificial intelligence, and examined their impact on export technical complexity after logarithmic processing. Columns (3) and (4) of Table 4 show the estimation results. The estimated coefficients of the two alternative explanatory variables remained significantly positive at the 1% level; thus, the overall results were consistent.

Replacing the Explained Variable

Variations in the quality of the same export product between provinces may lead to differences in the export technical complexity of the same export product in different provinces. Thus, borrowing from the correction method of Xu (2010) for export product quality, we re-measured the export technical complexity of the enterprises, then obtained its adjustment value for robustness testing, adopting Equation (1). The results are shown in Column (5) of Table 4. The estimated coefficient was 0.011, which did not change substantially, indicating that the results of benchmark regression were still valid after changing the measurement method for export technical complexity.

Enterprise-Level Further Discussion

Heterogeneity Test

Technical Complexity of Imported Robots

Based on the technical complexity, we divided the industrial robots into the categories of high-tech complexity and medium- and low-tech complexity. The types of industrial robots imported may have had a heterogeneous impact on the technical spillover effect, substitution of low-skilled labor, and improvement of labor productivity, and then may have had differential impacts on the export technical complexity of an enterprise. Accordingly, we further examined the heterogeneity effects of high-tech complexity industrial robots and medium- and low-tech complexity industrial robots on an enterprise’s export technical complexity. As the results in Columns (1) and (2) of Table 5 show, it is clear that the import of industrial robots under both classifications had a significant role in promoting the improvement of export technical complexity. Further, importing high-tech complexity robots have had a greater role in promoting an enterprise’s export technical complexity. This is likely because imported high-tech complexity robots can more obviously replace low-skilled labor or promote technological innovation, thereby greatly improving an enterprise’s export technical complexity. Moreover, high-tech complexity industrial robots are more conducive to generating technological spillover effects, thereby promoting the improvement of an enterprise’s export technical complexity.

Table 5.

Heterogeneity Test Results.

Explanatory variables	Technical complexity of imported robots		Types of enterprise ownership		Region
Explanatory variables	(1) High technical complexity	(2) Medium and low technical complexity	(3) Foreign-funded enterprises	(4) Local enterprises	(5) East	(6) Middle	(7) West
$treat_post$	0.0144*** (0.0030)	0.0099*** (0.0036)	−0.0088 (0.0061)	0.0198*** (0.0033)	0.0124*** (0.0032)	0.0111 (0.0149)	0.0098 (0.0157)
Observations	221,159	210,200	67,838	145,375	215,060	8,974	4,870
$R^{2}$	.983	.984	.979	.983	.983	.988	.987

Enterprise Ownership

Owing to the difference in technical level and resource allocation efficiency of foreign-funded enterprises and local enterprises, there will also be differences in the impact of industrial robots import on the export technical complexity of different ownership enterprises. Therefore, we classified the sample into foreign-funded and local enterprises, then examined the heterogeneity impacts. As shown by the results in Columns (3) and (4) of Table 5, the importing of industrial robots by local enterprises had a significant role in promoting the improvement of export technical complexity; however, there was no significant impact on export technical complexity for foreign-funded enterprises. To explain the reason for these results, it could be because the productivity level of foreign-funded enterprises is higher compared with that of local enterprises, the spillover effects brought by industrial robot imports may be greater; however, compared with foreign-funded enterprises, local enterprises are in inferior positions in terms of technical level, human capital, and productivity. Therefore, local enterprises are more likely to obtain technology spillover through foreign enterprises, which, in turn, can promote the export technical complexity to a greater extent.

Region

According to the statistics on the import of industrial robots at the province level in the China Customs Database, we explored the possibility that there could have been more industrial imported robots in Shanghai, Jiangsu, Guangdong, and Beijing, which are mainly distributed in the eastern region of China. Moreover, compared with the central and western regions, the eastern region has a higher level of economic development, higher institutional quality, and a greater highly skilled labor force, thus, the import of industrial robots may have had a greater role in promoting export technical complexity. Accordingly, we divided the samples into eastern, central, and western samples for grouped regression, the results of which are shown in Columns (5)-(7) of Table 5. We found that the import of industrial robots by enterprises in the eastern region had a significant role in promoting export technical complexity improvement; however, it had no significant impact on the export technical complexity of enterprises in the central and western regions, in line with Dong et al. (2022). Regarding the possible reason for this finding, the number of enterprises that imported industrial robots in the central region during the sample period accounted for less than 4% of the total number of imported industrial robot enterprises, and even fewer enterprises imported industrial robots in the western region, accounting for only approximately 2%. Moreover, the average import quantity of industrial robots in the central and western regions is only about one-twentieth of the average import quantity of eastern enterprises. Thus, it is difficult to form a scale effect in the central and western regions. In addition, the technical spillover effect after the import of industrial robots depends on the absorption capacity of local enterprises. Thus, the low level of economic development, backward system construction, and lack of highly skilled labor in the central and western regions will inhibit their technology spillover effects, which, in turn, will restrict the improvement of enterprises’ export technical complexity.

Mechanism Test

The basic conclusion of this study is that artificial intelligence can significantly improve an enterprise’s export technical complexity. Here, based on the above theoretical analysis, we present a constructed model of mediation effects to further verify how artificial intelligence can promote improved export technical complexity among manufacturing enterprises. The mediation effect models are as follows:

{lnetc}_{it} = λ_{1} + λ_{2} {treat}_{it} \times {post}_{it} + λ_{3} x_{it} + λ_{4} z_{jt} + v_{i} + v_{t} + ε_{ijt}

(8)

{mv}_{it} = η_{1} + η_{2} {treat}_{it} \times {post}_{it} + η_{3} x_{it} + η_{4} z_{jt} + v_{i} + v_{t} + ε_{ijt}

(9)

\begin{matrix} {lnetc}_{it} = δ_{1} + δ_{2} {treat}_{it} \times {post}_{it} + δ_{3} m v_{it} + δ_{4} x_{it} + δ_{5} z_{jt} \\ + v_{i} + v_{t} + ε_{ijt} \end{matrix}

(10)

where $mv_it$ represents the mediation variables for this study, including labor productivity ( $lp$ ) and the innovation output of enterprise ( $ty$ ). Referring to Haltiwanger et al. (1999), we characterized the labor productivity by the natural logarithm of the total industrial output of a firm divided by the number of employees, and the total industrial output is deflated by the factory price index. We denoted an enterprise’s innovation output by the sum of the number of patents of industrial enterprises, which was derived from the patent database of industrial enterprises.

Table 6 shows the results of the mechanism test. Column (1) and (4) present the estimation results from using Equation (8). Columns (2) and (5) show the results of the estimation adopting Equation (9), which estimated the effects of artificial intelligence on mediation variables, including labor productivity and innovation output. Columns (3) and (6) show the estimated results of Equation (10). In Column (3), the estimated coefficient of the artificial intelligence variable is significantly smaller than what is shown in Column (1), indicating that labor productivity is the mediating channel through which artificial intelligence promotes improvement in an enterprise’s export technical complexity (Damioli et al., 2021; Graetz & Michaels, 2015; W. M. Zhang et al., 2023). This result supports H1.

Table 6.

Mechanism Test Results.

Explanatory variables	Labor productivity			Innovation
Explanatory variables	(1) $lnetc$	(2) $lp$	(3) $lnetc$	(4) $lnetc$	(5) $ty$	(6) $lnetc$
$treat_post$	0.0140*** (0.0032)	0.0414*** (0.0039)	0.0134*** (0.0032)	0.0140*** (0.0031)	0.0628*** (0.0116)	0.0139*** (0.0031)
$lp$			0.0153*** (0.0026)
$ty$						0.0015** (0.0007)
Observations	210,522	210,522	210,522	232,674	232,674	232,674
$R^{2}$	.981	.983	.981	.983	.673	.983

The results in Column (5) show that artificial intelligence significantly increases innovation output, in line with L. Wang et al. (2023), and has a further significant impact on export technical complexity, as shown in Column (6). Moreover, the estimated coefficient of artificial intelligence in Column (6) is smaller than what is shown in Column (4); therefore, it indicates that artificial intelligence improves export technical complexity through innovation to support H2. The reason may be as follows. Artificial intelligence have greatly improved the efficiency of enterprise information searches, which can contribute to improved innovation ability. Artificial intelligence can also reduce production and operation costs (Javaid et al., 2021), which make an enterprise have more capital to invest in more complex and high-end technology R&D. In addition, artificial intelligence will also have the innovative effect of “learning by doing,” which leads to improved production efficiency. Furthermore, enterprise innovation is an important factor that affects export technical complexity (Zhu and Fu, 2013).

Industry-Level Further Discussion

The above analyses examined the impact of artificial intelligence in promoting export technology complexity at the enterprise level. In view of the potential spillover effect of technological progress, we further explored the impact of artificial intelligence on export technical complexity at the industry level. We verified the impact of artificial intelligence on export technical complexity at the industry level from two perspectives. First, we examined the impact of artificial intelligence intensity on the export technical complexity of other non-imported enterprises at the regional-industry level, that is, we examined the spillover effect of artificial intelligence on other enterprises. Second, we examined the impact of artificial intelligence intensity on overall export technical complexity at the regional-industry level. On this basis, we further analyzed the resource reallocation effect. The econometric model built under the first way is shown in Equation (11):

\begin{matrix} lnet c_{it} = μ_{1} + μ_{2} ROBO T_{ljt} + μ_{3} x_{it} + μ_{4} z_{jt} + v_{j} + v_{t} \\ + v_{lt} + ε_{iljt} \end{matrix}

(11)

where $l$ is the city code at the regional level, $i$ is the enterprises that have not imported industrial robots at the regional-industry level, $j$ indicates the industry, ${lnetc}_{it}$ is the export technical complexity of other non-imported enterprises, and $ROBO T_{ljt}$ is the application intensity of artificial intelligence at the regional-industry level. We specifically used the proportion of enterprises using industrial robots ( $rate_ent$ ) at the regional-industry level, the proportion of the total sales output value of enterprises using industrial robots ( $rate_sale$ ) at the regional-industry level, and the natural logarithm of the total amount of industrial robots used by enterprises ( $tvalue$ ) at the regional-industry level, to characterize the application intensity of artificial intelligence.

The control variables at the enterprise and industry levels were consistent with Equation (1), while controlling for the fixed effects of enterprise, year, and regional-industry level, and clustering at the enterprise level. The results of the regression are reported in Columns (1) to (3) of Table 7. The application intensity of artificial intelligence as measured by three different variables played a significant role in promoting the improvement of the export technical complexity of other enterprises at the regional-industry level. This indicates that artificial intelligence has created spillover effects on the export technical complexity of other enterprises at the regional-industry level, possibly because when more enterprises apply artificial intelligence, the improvement of their labor productivity will increase the intensity of industry competition, thus forcing other enterprises to improve their labor productivity, which, in turn, promotes export technical complexity.

Table 7.

Results of the Effects on Export Technical Complexity at the Industry Level.

Explanatory variables	Export technical complexity of other enterprises			Export technical complexity at the industry level
Explanatory variables	(1) $rate_ent$	(2) $rate_sale$	(3) $tvalue$	(4) $rate_ent$	(5) $rate_sale$	(6) $tvalue$
$ROBOT$	0.0285*** (0.0048)	0.0146*** (0.0028)	0.0005*** (0.0001)	0.0269*** (0.0026)	0.0284*** (0.0022)	0.0022*** (0.0001)
Observations	251,170	251,170	251,177	132,685	132,676	132,685
$R^{2}$	.984	.984	.984	.962	.962	.962

The specific approach of the second idea is as follows. First, we measured the overall export technical complexity at the regional-industry level. Second, we examined the impact of the application intensity of artificial intelligence on overall export technical complexity at the regional-industry level, of which the application intensity was consistent with the Equation (11). The equation for the overall export technical complexity at the regional-industry level is shown in Equation (12):

{lnetc}_{ljt}^{G} = \sum_{i \in l, j} θ_{it} \times lnet c_{it}

(12)

where $l$ is the city code at the regional level, $i$ represents the enterprise, $j$ represents the industry, and $i \in l, j$ represents the enterprises in region $l$ and industry $j$ . $θ_{it}$ indicates the export share of enterprise $i$ in region $l$ -industry $j$ , and ${lnetc}_{ljt}^{G}$ denotes the overall export technical complexity at the regional-industry level. Referring to Olley and Pakes (1996) method of decomposing productivity, we broke down the overall export technical complexity at the regional-industry level into identities to examine the effects of resource reallocation. An identity consists of a simple average export technical complexity for all firms at the regional-industry level, and a covariance between the export technical complexity of an enterprise and its export share. Specifically, this is shown in Equation (13):

{lnetc}_{ljt}^{G} = {lnetc}_{ljt}^{AVE} + \sum_{i \in l, j} (θ_{it} - \bar{θ_{ljt}}) (lnet c_{it} - {lnetc}_{ljt}^{AVE})

(13)

where ${lnetc}_{ljt}^{AVE}$ represents the simple average export technical complexity at the regional-industry level, $\bar{θ_{ljt}}$ indicates the average export share of all enterprises in region $l$ -industry $j$ , and the second item on the right side of Equation (13) is the covariance between an enterprise’s export technical complexity and its export share, which is recorded as ${lnetc}_{ljt}^{COV}$ , and used to examine the redistributive effects of export share among enterprises with different levels of export technical complexity in the regional industry. The larger the export technical complexity covariance terms, indicating that enterprises with higher export technical complexity within the region-industry have gained a higher share of exports. Accordingly, we constructed a model to further investigate the impact of artificial intelligence intensity on overall export technical complexity and decomposition items at the regional-industry level:

lnet c_{ljt} = ϕ_{1} + ϕ_{2} ROBO T_{ljt} + ϕ_{3} z_{jt} + v_{j} + v_{t} + v_{lt} + ε_{ljt}

(14)

where $lnet c_{ljt}$ represents the overall export technical complexity, simple average export technical complexity and the covariance at the region-industry level. $z_{jt}$ is the degree of competition at the industry level which is consistent with Equation (1), and we controlled for fixed effects at the enterprise, year, and regional-industry levels. The measurement of the intensity of artificial intelligence at the regional-industry level ( $ROBO T_{ljt}$ ) was consistent with Equation (11). Columns (4) to (6) of Table 7 report the impact of artificial intelligence intensity on overall export technical complexity at the regional-industry level. It can be seen that, no matter what method is used to measure the application intensity of artificial intelligence, it shows a significant positive correlation with overall export technical complexity. This indicates that the application of artificial intelligence significantly improves overall export technical complexity at the regional-industry level.

Table 8 shows the estimated results of artificial intelligence on the effect of simple average export technical complexity and export technical complexity covariance terms at the region-industry level. Columns (1) to (3) present the results of the effects of the application intensity of artificial intelligence using different agent variables on the average export technical complexity at the regional-industry level. We explored the significant role of the application of artificial intelligence in promoting the improvement of the average export technical complexity of the region-industry. Columns (4) to (6) show the results of the effects of the application intensity of artificial intelligence on the covariance term of export technical complexity characterizing the resource reallocation effect.

Table 8.

Results of the Effects on Items at the Industry Level.

Explanatory variables	Simple average export technical complexity			Covariance of export technical complexity
Explanatory variables	(1) $rate_ent$	(2) $rate_sale$	(3) $tvalue$	(4) $rate_ent$	(5) $rate_sale$	(6) $tvalue$
$ROBOT$	0.0238*** (0.0026)	0.0223*** (0.0021)	0.0016*** (0.0001)	0.003*** (0.0005)	0.0065*** (0.0006)	0.0007*** (0.0000)
Observations	132,685	132,676	132,685	132,685	132,676	132,685
$R^{2}$	.965	.965	.965	.034	.035	.036

We also found that the estimated coefficient of the application intensity of artificial intelligence was significantly positive, indicating that the application of artificial intelligence significantly promoted the improvement of the export technical complexity covariance term. This indicates that the redistributive effect of the export share is an important channel for applying artificial intelligence to improve overall export technology complexity at the regional-industry level. A possible reason for this is that the application of artificial intelligence will lead to a shift in export share from enterprises with lower export technical complexity to enterprises with higher export technical complexity. Comparing the estimated coefficients of the average export technical complexity and export technical complexity covariance term under the corresponding artificial intelligence application intensity agent variable in Table 8, it can be seen that the sum of the estimated coefficients of the two is equal to the estimated coefficient of the overall export technical complexity at the regional-industry level (e.g., 0.0238 + 0.003 = 0.0268 ≈ 0.0269), which again verifies that the redistributive effect of export share promotes the effects of artificial intelligence on the improvement of export technical complexity at the regional-industry level.

Conclusions and Implications

Conclusions and Policy Implications

Based on matched data of the China Customs and Chinese Industrial Enterprises Databases from 2000 to 2015, this study examined the impact of artificial intelligence on upgrading the export technical complexity of Chinese manufacturing enterprises at the enterprise and industry levels. First, the results show that artificial intelligence significantly promotes upgrading export technical complexity among Chinese manufacturing enterprises. Second, the impact of artificial intelligence on upgrading export technical complexity is dynamic, showing an “inverted U-shape.” Third, artificial intelligence has a greater effect on export technical complexity upgrading if artificial intelligence features high-tech complexity or enterprises are local or in the eastern region. Fourth, artificial intelligence significantly promotes labor productivity and innovation to upgrade the technical complexity of exporting Chinese manufacturing enterprises. Finally, the application intensity of artificial intelligence significantly impacts the average export technical complexity upgrading of Chinese manufacturing enterprises through spillover and resource reallocation effects.

Based on the above conclusions, this study provides some policy recommendations. First, the Chinese government should encourage enterprises to import industrial robots for transformation and upgrading, especially high-tech complexity robots, to promote the upgrading of export technical complexity. Second, the government should actively support local enterprises that make use of artificial intelligence to encourage upgrading export technical complexity through fiscal and tax preferential and rent reduction policies. Third, the education department must actively promote the construction of professional disciplines, such as robot technology and application, and strengthen the cultivation of high-end talent in related emerging technologies in China. However, it is also necessary to strengthen coordination and cooperation with policies to promote industrial and trade structure upgrading and encourage universities and enterprises to jointly cultivate highly skilled talent, urgently needed in emerging technology fields, to strengthen vocational skills training for employees and continuously improve worker quality. Fourth, enterprises should increase domestic R&D investment in artificial intelligence to achieve independent domestic supply. Finally, the government should address the balance between different regions when promoting the development of artificial intelligence, especially improving supporting measures for the application of artificial intelligence in the central and western regions.

Limitations and Future Research

Due to the data availability, this study used the China Customs Database and Chinese Industrial Enterprises Database from 2000 to 2015. Therefore, future studies can investigate the impact of the new generation of artificial intelligence on upgrading export technical complexity in recent years and compare the heterogeneous impacts with traditional artificial intelligence. In addition, we only focused on the effect of imported artificial intelligence on upgrading export technical complexity. Future studies should compare the heterogeneous impacts between imported artificial intelligence and local artificial intelligence.

Footnotes

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This study was supported by Natural Science Foundation of Hunan Province (Grant No. 2023JJ30218), Zhuzhou Social Science Research Project (Grant No. ZZSK2024135), and Key Scientific Research Project of Hunan Provincial Department of Education (Grant No. 22A0407, 23A0431).

ORCID iD

Changqing Lin

Data Availability Statement

Data sharing not applicable to this article as no datasets were generated or analyzed during the current study.

References

Aaronson

Phelan

B. J.

(2019). Wage shocks and the technological substitution of low-wage jobs. The Economic Journal, 129(617), 1–34.

Acemoglu

Restrepo

(2020). Robots and jobs: Evidence from U. S. labor markets. Journal of Political Economy, 128(6), 2188–2244.

Acemoglu

Restrepo

(2022). Demographics and automation. The Review of Economic Studies, 89(1), 1–44.

Alguacil

Turco

A. L.

Martínez-Zarzoso

(2022). Robot adoption and export performance: Firm-level evidence from Spain. Economic Modelling, 114, 105912.

Angrist

J. D.

Pischke

(2014). Mastering metrics: The path from cause to effect. Princeton University Press.

Arrow

K. J.

(1962). The economic implications of learning by doing. The Review of Economic Studies, 29(3), 155–173.

Artuc

Bastos

Rijkers

(2018). Robots tasks and trade (Working Paper No. 8674). World Bank Policy Research.

Bai

(2020). Research on the measurement and its influencing factors of the technical complexity of agricultural products export in China [Conference session]. Recent Trends in Decision Science and Management: Proceedings of ICDSM 2019. SpringerSingapore.

Banga

(2022). Digital technologies and product upgrading in global value chains: Empirical evidence from Indian manufacturing firms. The European Journal of Development Research, 34(1), 77–102.

10.

Berg

Buffie

E. F.

Zanna

L. F.

(2016). Robots, growth, and inequality. Finance & Development, 53(3), 10–13.

11.

Cabral

Veiga

(2010). Determinants of export diversification and sophistication in Sub-Saharan Africa (FEUNL Working Paper Series No: 550). Social Science Research Network.

12.

Cai

Liu

(2009). Competition and corporate tax avoidance: Evidence from Chinese industrial firms. The Economic Journal, 119(537), 764–795.

13.

Damioli

Van Roy

Vertesy

(2021). The impact of artificial intelligence on labor productivity. Eurasian Business Review, 11, 1–25.

14.

DeCanio

S. J.

(2016). Robots and humans-complements or substitutes? Journal of Macroeconomics, 49, 280–291.

15.

Demir

(2021). Institutional similarity, firm heterogeneity and export sophistication. The World Economy, 29(9), 1–47.

16.

DeStefano

De Backer

Suh

J. R.

(2019). Industrial robotics and product quality (Working Papers, No. 7). OECD Science, Technology and Industry.

17.

DeStefano

Timmis

(2021). Robots and export quality (Working Papers No. 9678). OECD Science, Technology and Industry Policy Research.

18.

Ding

(2018). Product space, potential comparative advantages and export technological complexity. Australian Economic Papers, 57(3), 218–237.

19.

Dong

Guo

(2022). Intellectual property rights protection and export product quality: Evidence from China. International Review of Economics & Finance, 77, 143–158.

20.

Faber

(2020). Robots and reshoring: Evidence from Mexican labor markets. Journal of International Economics, 127, 103384.

21.

Fan

Anwar

Huang

(2018). Cultural diversity and export sophistication. International Review of Economics & Finance, 58, 508–522.

22.

Goldfarb

Trefler

(2018). Artificial intelligence and international trade[M]//The economics of artificial intelligence: An agenda (pp. 463–492). University of Chicago Press.

23.

Graetz

Michaels

(2015). Robots at work: The impact on productivity and jobs. Centre for Economic Performance LSE 447.

24.

Hallak

J. C.

Schott

P. K.

(2011). Estimating cross-country differences in product quality. The Quarterly Journal of Economics, 126, 417–474.

25.

Haltiwanger

J. C.

Lane

J. I.

Spletzer

(1999). Productivity differences across employers: The roles of employer size, age, and human capital. American Economic Review, 89(2), 94–98.

26.

Hasan

Wachtel

Zhou

(2009). Institutional development, financial deepening and economic growth: Evidence from China. Journal of Banking & Finance, 33, 157–170.

27.

Hausmann

Hwang

Rodrik

(2007). What you export matters. Journal of Economic Growth, 12(1), 1–25.

28.

Head

Ries

(2003). Heterogeneity and the FDI versus export decision of Japanese manufacturers. Journal of the Japanese and International Economies, 17(4), 448–467.

29.

Heckman

J. J.

Ichimura

Todd

P. E.

(1997). Matching as an econometric evaluation estimator: Evidence from evaluating a job training programme. Review of Economic Studies, 64(4), 605–654.

30.

Hémous

Olsen

(2022). The rise of the machines: Automation, horizontal innovation, and income inequality. American Economic Journal: Macroeconomics, 14, 179–223.

31.

Hong

L. Y.

Liu

X. J.

Zhan

H. W.

Han

(2022). Use of industrial robots and Chinese enterprises’ export quality upgrading: Evidence from China. The Journal of International Trade & Economic Development, 31(6), 860–875.

32.

Huang

L. Y.

Lin

(2022). Robot adoption and energy performance: Evidence from Chinese industrial firms. Energy Economics, 107, 105837.

33.

Huang

Chen

Liu

(2014). What accounts for the export sophistication of China’s indigenous enterprises in the perspective of vertical specification [Conference session] [pp. 680–686]. International Conference on Management Science & Engineering 21th Annual Conference Proceedings, Helsinki, Finland. IEEE.

34.

Jarreau

Poncet

(2012). Export sophistication and economic growth: Evidence from China. Journal of Development Economics, 97(2), 281–292.

35.

Javaid

Haleem

Singh

R. P.

Suman

(2021). Substantial capabilities of robotics in enhancing industry 4.0 implementation. Cognitive Robotics, 1, 58–75.

36.

Yao

(2022). Cleaner production standards and increased technical complexity for textile exporters. Mathematical Problems in Engineering, 2022(1), 3646949.

37.

Krenz

Prettner

Strulik

(2021). Robots, reshoring, and the lot of low-skilled workers. European Economic Review, 136, 103744.

38.

Kromann

Skaksen

J. R.

Sørensen

(2011). Automation, labor productivity and employment: A cross country comparison (Copenhagen Business School Working Paper). CEBR.

39.

Lankisch

Prettner

Prskawetz

(2019). How can robots affect wage inequality? Economic Modelling, 81, 161–169.

40.

C. Q.

(2017). R&D, financing constraints and export green-sophistication: Evidence from China. China Economic Review, 47, 234–244.

41.

Zhu

J. S.

(2019). The role of environmental regulation and technological innovation in the employment of manufacturing enterprises: Evidence from China. Sustainability, 11(10), 2982.

42.

Wang

Bao

(2021). Employment effect of robots: Mechanisms and Chinese experiences. Management World, 37(9), 104–119.

43.

Lin

Xiao

Yin

(2022). How do industrial robots applications affect the quality upgrade of Chinese export trade? Telecommunications Policy, 46(10), 102425.

44.

Madsen

J. B.

(2010). Growth and capital deepening since 1870: Is it all technological progress? Journal of Macroeconomics, 32(2), 641–656.

45.

Michaels

Natraj

Reenen

J. V.

(2014). Has ICT polarized skill demand? Evidence from eleven countries over twenty-five years. Review of Economics and Statistics, 96(1), 60–77.

46.

Mishra

Lundström

Anand

(2011). Service export sophistication and economic growth (Working Paper No. 5606). World Bank Policy Research.

47.

Nguyen

D. X.

(2016). Trade liberalization and export sophistication in Vietnam. The Journal of International Trade & Economic Development, 25(8), 1071–1089.

48.

Olley

G. S.

Pakes

(1996). The dynamics of productivity in the telecommunications equipment industry. Econometrica, 64(6), 1263–1297.

49.

Phelps

E. S.

(1962).The new view of investment: A neoclassical analysis. The Quarterly Journal of Economics, 76(4), 548–567.

50.

J. Y.

Xiang

G. L.

(2020). Opening up of financial service sector, external financial dependence and the export sophistication of manufacturing sector—an empirical analysis based on different policies of services trade restrictiveness. Journal of International Business and Economics, 6, 78–91.

51.

Rodrik

(2006). What’s so special about China’s export. China & World Economy, 14(5), 1–19.

52.

Rodrik

(2018). New technologies, global value chains, and developing economies (NBER Working Paper No. 25164). https://doi.org/10.3386/w25164.

53.

Song

Y. G.

Hao

X. Z.

Zheng

(2022). Intermediate import, independent innovation and export sophistication of Chinese manufacturing enterprises. Structural Change and Economic Dynamics, 60, 126–140.

54.

Stapleton

Webb

(2020). Automation, trade and multinational activity: Micro evidence from Spain (CSAE Working Paper SeriesWPS/2020–16), University of Oxford.

55.

Wang

E. Z.

Lee

C. C.

(2022). Assessing the impact of industrial robots on manufacturing energy intensity in 38 countries. Energy Economics, 105, 105748.

56.

Wang

Sarker

Alam

Sumon

(2021). Artificial intelligence and economic growth: A theoretical framework. Scientific Annals of Economics and Business, 68(4), 421–443.

57.

Wang

Zhou

Chiao

(2023). Robots and firm innovation: Evidence from Chinese manufacturing. Journal of Business Research, 162, 113878.

58.

Wang

Wei

S. J.

(2007). The rising sophistication in China’s exports: Assessing the roles of processing trade, foreign invested firms, human capital and government policies [Conference session]. Working Paper for the NBER Conference on China’s Growing Role in World Trade, Columbia University, New York.

59.

Willem

Pai

H. K.

(2015). The sophistication of East Asian exports. Journal of the Asia Pacific Economy, 20(4), 658–678.

60.

Tang

Zhang

(2022). Population aging, industrial intelligence and export technology complexity. Sustainability, 14(20), 13600.

61.

(2007). Measuring China’s export sophistication (Working Paper, 10 [pp. 1–39]). China Europe International Business School.

62.

(2010). The sophistication of exports: Is China special? China Economic Review, 21(3), 482–493.

63.

J. Y.

(2009). Foreign direct investment, processing trade, and the sophistication of China’s exports. China Economic Review, 20(3), 425–439.

64.

Yin

Z. H.

Zeng

W. P.

(2023). The effects of industrial intelligence on China’s energy intensity: The role of technology absorptive capacity. Technological Forecasting and Social Change, 191, 122506.

65.

X. J.

(2015). Sophistication of China’s manufactured exports and determinants. Transnational Corporations Review, 7(2), 169–189.

66.

Zhang

F. F.

Zhang

Q. N.

(2023). Robot adoption and export performance: evidence from Chinese industrial firms. Journal of Manufacturing Technology Management, 34(6), 896–916.

67.

Zhang

W. M.

Peng

J. C.

Zhang

(2023). Disruptive displacement: The impacts of industrial robots on the energy industry’s international division of labor from a technological complexity view. Energies, 16(8), 3349.

68.

Zhu

S. J.

X. L.

(2013). Drivers of export upgrading. World Development, 51, 221–233.

Does Artificial Intelligence Improve Export Technical Complexity Upgrade of Manufacturing Enterprises? Evidence from China

Abstract

Keywords

Introduction

Literature Review

Impact of Artificial Intelligence

Determinants and Effects of Export Technical Complexity

Artificial Intelligence and Export Product Quality

Research Gaps

Theoretical Mechanisms

Labor Productivity

Enterprise Innovation

Materials and Methods

Data

Baseline Model

Variables

Export Technical Complexity

Artificial Intelligence

Control Variables

Results

Matching Balance Test Results

Baseline Results

Parallel Trend Test Results

Robustness Tests

Other Endogenous Issues

Replacing Explanatory Variables

Replacing the Explained Variable

Enterprise-Level Further Discussion

Heterogeneity Test

Technical Complexity of Imported Robots

Enterprise Ownership

Region

Mechanism Test

Industry-Level Further Discussion

Conclusions and Implications

Conclusions and Policy Implications

Limitations and Future Research

Footnotes

Declaration of Conflicting Interests

Funding

ORCID iD

Data Availability Statement

References