Artificial Intelligence, Technological Innovation, and Employment Transformation for Sustainable Development: Evidence from China

Abstract

With the rapid advancement of artificial intelligence (AI) technology, the global employment structure is undergoing profound transformations, significantly impacting social sustainability. This study utilizes panel data from 30 Chinese provinces spanning the years 2010 to 2022 and applies a two-way fixed-effects model to analyze the impact of AI development on the employment skills structure. The findings indicate that advancements in AI technology significantly suppress the demand for low-skilled labor while markedly enhancing the demand for both middle- and high-skilled labor. The threshold effect analysis reveals a nonlinear relationship between AI advancements and the demand for low-skilled workers. Mediation effect tests demonstrate that technological innovation serves as a mediating factor in AI’s impact on low- and middle-skilled labor but has no significant effect on high-skilled labor. The heterogeneity analysis further indicates that AI’s negative impact on low-skilled female employment is more severe than for males, while its positive impact on high-skilled male workers is significant. Additionally, the employment effects of AI are mainly observed in labor-intensive provinces, with minimal influence in capital-intensive areas. This study suggests harnessing AI’s potential to promote employment while proactively mitigating its disruptive effects on the labor market through enhanced research and development support, strengthened employment security, and coordinated regional economic development, thereby advancing sustainable economic and social progress.

1. Introduction

Artificial Intelligence, as a cutting-edge technology in the current field of science and technology, not only plays an essential role in advancing technological progress and driving industrial transformation but also has significant impacts on national economic growth and social development. As its applications continue to deepen, AI is increasingly recognized as a transformative force across multiple industries and a key engine of economic growth. China has emerged as a key global competitor in AI innovation. According to a report by iResearch (Shanghai, China), the scale of China’s AI industry reached CNY 213.7 billion in 2023 and is projected to expand to CNY 811 billion by 2028, with a five-year compound annual growth rate of 30.6% [1]. Amid China’s ongoing economic transformation, AI has become a critical tool for sustaining growth and driving long-term development [2]. As a hallmark of the Fourth Industrial Revolution, AI not only accelerates technological progress and industrial transformation but also has far-reaching effects on employment, sparking widespread concerns regarding job displacement and labor market restructuring.

Artificial intelligence has created new employment opportunities while simultaneously exacerbating social inequalities. In the rapidly evolving economic, technological, and social landscape, ensuring that labor markets provide stable, equitable, and high-quality employment is of critical importance. Particularly in the context of promoting social equity and inclusivity, it is essential to guarantee that disadvantaged groups have equal access to labor markets and opportunities for career advancement [3,4]. The Sustainable Development Goals (SDGs), adopted by the United Nations, in 2015, as part of the 2030 Agenda for Sustainable Development [5], closely link high-quality labor market development to gender equality (SDG 5), decent work and economic growth (SDG 8), and reduced inequalities (SDG 10). Specifically, SDG 5 aims to empower all women and girls by promoting equal rights and opportunities, ensuring their full participation in employment and career advancement. SDG 8 seeks to promote sustained, inclusive, and sustainable economic growth while providing full, productive employment and decent work for all. SDG 10 focuses on narrowing disparities both within and among countries, with an emphasis on improving the socioeconomic conditions of marginalized, minority, and vulnerable groups. In the context of rapid advancements in artificial intelligence, the skill-biased effects it engenders may further widen the employment gaps among workers of differing skill levels, thereby exacerbating issues of income inequality and unequal opportunities. Ensuring stable and inclusive employment opportunities is crucial for driving sustained economic growth and fostering social cohesion [6]. Within the framework of sustainable development, exploring the relationship between AI and employment not only deepens our understanding of how technological change impacts labor markets but also serves as a vital pathway for fostering inclusive growth, reducing employment disparities, and achieving sustainable development goals.

In recent years, the most notable advancements in artificial intelligence have stemmed from the use of algorithms to process and analyze vast amounts of structured data, with machine learning and deep learning playing an indispensable role in applications such as speech recognition, text translation, and image recognition. As these technologies have evolved, generative AI has entered the public spotlight, with large language models (LLMs) sparking widespread discussion [7,8]. The relationship between generative AI and employment is a key factor driving AI-induced labor market transformation. Large language models (LLMs), such as GPT, have already had significant economic, social, and policy implications [9]. Specifically, it is estimated that approximately 80% of the U.S. workforce will be affected by the introduction of LLMs, with particularly pronounced impacts on industries such as law, securities, commodities, and investment [10,11]. Generative AI is fundamentally reshaping managerial and professional functions, altering labor skill demands, and accelerating workforce substitution and restructuring, thereby significantly influencing labor market allocation [12]. Moreover, machine learning and related technologies are increasingly regarded as potential general-purpose technologies (GPTs). Their rapid development not only signals substantial economic value across multiple industries and professions but also suggests far-reaching implications for the labor market [13,14].

The AI-driven transformation of employment structures is accelerating industrial upgrading and technological advancement, injecting new momentum into long-term economic growth [15]. Compared to developed countries, China’s labor market is predominantly composed of low- to middle-skilled workers, with significant disparities in educational attainment between urban and rural areas. These structural differences limit the workforce’s adaptability to the digitalization and intelligent transformation of industries. At this critical juncture, as China transitions between old and new economic drivers, an in-depth analysis of AI’s impact on the country’s employment skills structure can provide a solid foundation for policy making. Such research is essential for effectively addressing the challenges posed by technological change, ensuring a positive cycle of economic development and social stability, and fostering employment equality and sustainable social progress. Beyond its domestic policy implications, this study also offers valuable insights for other countries navigating similar labor market transformations.

A review of the existing literature reveals several limitations in the current research on the impact of artificial intelligence on the structure of employment skills, which call for further exploration and refinement. First, existing studies fall short in explaining the structural characteristics and dynamic evolution of China’s labor market, which is predominantly composed of low- and middle-skilled workers, with high-skilled labor playing a secondary role. In particular, there is a significant gap in in-depth analysis regarding gender and regional disparities. Second, most empirical studies rely on industrial firm-level data, especially indicators such as the density of industrial robots, to measure the level of AI development [16,17,18,19]. However, such indicators typically capture only a narrow dimension of intelligence within the manufacturing sector and fail to reflect the broader penetration of AI technologies across diverse industries. Moreover, certain composite indicator systems, which are often constructed based on specific research samples, tend to incorporate non-AI-related variables, thus compromising the accuracy and explanatory power of AI measurements [20].

In response to these limitations, this study introduces a patent-based AI database as the core data source, thereby expanding both the scope and depth of AI measurement from a methodological perspective. Specifically, the database is constructed based on key AI-related terms derived from recent Chinese policy documents and standard systems, complemented by internationally recognized reports and seminal literature. It compiles AI-related patent application and authorization information across all provinces in China. Compared to traditional indicators such as industrial robot density, patent data provide a wider scope, enabling the capture of AI application trends beyond the manufacturing sector and providing a more comprehensive reflection of regional AI development levels. Analysis using this database enables a more precise assessment of the impact of AI technological advancement on the employment skill structure. This approach offers significant policy implications for sustainable development, particularly in the context of China’s unique labor market structure, and contributes new theoretical perspectives and empirical dimensions to the study of AI and labor market dynamics.

This study constructs an AI patent database spanning the years 2010 to 2022 and employs panel data from 30 Chinese provinces to conduct an empirical analysis using a two-way fixed-effects model, a mediation effect model, and a threshold effect model. The objective is to examine the impact of AI technology on the employment skills structure and explore the underlying mechanisms through which these effects occur. Key research questions include the following: How does the application of AI in China affect employment across different skill levels? Do these employment effects vary by gender? How do AI-driven employment effects differ across regions and skill levels? What are the transmission mechanisms through which AI influences workers with varying skill levels? Addressing these questions provides a clearer understanding of the relationship between AI and employment skill structures in the Chinese context, offering valuable insights into the dynamics of AI-driven labor market transformations.

The structure of this paper is organized as follows: Section 2 presents the theoretical framework and research hypotheses, outlining four key hypotheses proposed in this study. Section 3 introduces the empirical research strategy and data sources, providing a detailed discussion of model construction, variable selection, and data specifications. Section 4 conducts an in-depth analysis of the empirical findings, focusing on the impact of AI on employment across different skill levels, genders, and regions, and examining the underlying transmission mechanisms. Section 5 summarizes the main conclusions of the study. Finally, Section 6 proposes a series of policy recommendations aimed at enhancing employment security systems and improving workers’ skill development, contributing to sustainable social and economic development.

2. Theoretical Analysis and Hypotheses

Technological advancement has both negative and positive effects on employment, encompassing substitution effects, as well as employment-enhancing mechanisms. The substitution effect refers to the replacement of human labor with artificial intelligence through automation technologies, which may result in the elimination of certain jobs and a reduction in overall labor demand [21]. In contrast to high-skilled and high-income occupations, low-skilled and low-income jobs are more susceptible to displacement by AI-driven automation. However, automation’s advancement also generates positive spillover effects, which can somewhat counterbalance the adverse impacts of substitution and even create new employment opportunities. At the core of these positive effects lies the creation effect, whereby the application of new technologies gives rise to emerging occupational categories and job roles, thereby increasing labor demand [22]. In addition to the creation effect, productivity and reinstatement effects also substantially contribute to employment growth. The productivity effect denotes an improvement in production efficiency that is enabled by new technologies, which expands the scale of production and, in turn, stimulates the demand for labor [23,24]. The reinstatement effect pertains to the emergence of new tasks that leverage the comparative advantages of human labor, leading to a transformation of work content and further boosting labor demand [25]. Although artificial intelligence is unlikely to cause a substantial decline in total employment, its impact on the structure of employment remains considerable and warrants close attention [21]. The overall framework of AI reshaping the employment skill structure is shown in Figure 1.

Autor (2003) was among the first to redefine the production function from a task-based rather than a skill-based perspective, theoretically examining the effects of computer technology on different types of tasks [24]. He argued that computer applications substitute for routine tasks while complementing non-routine tasks. Routine tasks are highly repetitive and follow fixed procedures, whereas non-routine tasks require complex understanding, analysis, and decision making, making them difficult to fully automate. As a result, computerization reduces the demand for workers engaged in routine tasks, thereby affecting their employment prospects. Existing studies widely recognize the significant substitution effect of automation and AI, as evidenced by the declining labor share in value-added income [23,26]. Low-skilled workers, who are primarily engaged in routine tasks, face a higher risk of displacement, increasing unemployment pressure. In contrast, high-skilled workers, whose jobs often involve non-routine tasks, are less susceptible to automation-driven replacement.

Advancements in AI technology not only enhance overall productivity but also reduce production costs, expand market demand, and promote capital accumulation and output growth, thereby contributing to higher employment levels to some extent [27,28,29]. Acemoglu and Restrepo (2019b) proposed a task-based analytical framework, arguing that automation technologies improve productivity through more flexible task allocation—a mechanism known as the productivity effect. Under certain conditions, this effect can increase labor’s share of the total output [25]. Moreover, AI can generate new tasks through both the creation effect and the reinstatement effect, which helps offset some of the negative impacts caused by its substitution effect. As a complementary technology that enhances labor productivity, AI directly increases the marginal productivity of labor. OpenAI’s research team estimates that the application of large language models can improve the efficiency of approximately 15% of work tasks in the United States and predicts that AI will further assist workers in enhancing productivity across more industries in the future [10]. Overall, the net effect of automation on labor demand depends on the dynamic balance between substitution effects and employment-enhancing effects.

However, the current wave of technological change has not sufficiently focused on creating new tasks capable of effectively absorbing labor. This trend may lead to stagnation in labor demand growth, a declining labor share in national income, worsening income inequality, and a slowdown in productivity growth [21]. In addition, the gap between the skills required by new technologies and those workers actually have makes labor market adjustment more difficult and worsens employment inequality. As a result, technological innovation may induce market failures, with distortions in the innovation process potentially weakening or even reversing the welfare benefits of technological progress, thereby constraining economic growth to some extent [30]. The shortage of new tasks and the growing mismatch between job needs and worker capabilities have caused an imbalance between labor supply and demand, ultimately reducing labor demand. This challenge is particularly pronounced for low-skilled workers, who face a heightened risk of unemployment, posing a threat to the sustainable development of employment. In this context, AI technology is actively reshaping how firms assign tasks and structure skills, further affecting labor market dynamics.

Research indicates that firms adopting AI experience a significant increase in skill turnover, characterized by both the displacement of existing skill demands and the introduction of new skills. Simultaneously, these firms reduce hiring for non-AI-related positions [31]. The impact of AI technology is particularly pronounced for high-skilled workers, especially those with higher educational attainment, as their professional skills exhibit greater compatibility with AI applications [32]. Moreover, AI automation is taking over more traditional jobs, which pushes the remaining employment opportunities into smaller groups for highly technical workers [33]. As technological advancements continue, the ability to solve complex problems has become increasingly critical, leading to a rapid rise in demand for cognitive-task-oriented occupations. Consequently, skilled workers hold a competitive advantage over low-skilled laborers. In addition, the financial returns from education have been rising steadily, which further strengthens the position of high-skilled workers in the labor market [34].

The widespread adoption of automation technology has significantly reduced employment opportunities for low-skilled workers by replacing repetitive tasks [35]. AI can efficiently perform simple, repetitive, accuracy-sensitive, and risk-prone tasks, thereby displacing certain traditional job roles. Studies predict that in the coming decades, approximately 54% of jobs in China could be affected by AI substitution [36]. Since low-skilled workers are primarily engaged in routine and easily automated tasks, they face a greater risk of displacement by advanced technology [26]. At the same time, because high-skilled workers typically have higher marginal productivity than low-skilled workers, AI applications indirectly increase the demand for high-skilled labor [37]. The new job opportunities generated by technological advancements are predominantly concentrated in knowledge-intensive and technology-driven fields, such as AI trainers and smart hardware technicians. This trend not only fuels demand for high-skilled labor but also fosters a structural shift in the labor market toward higher skill levels. Skill-biased technological change increases the demand for high-skilled workers and reduces reliance on low-skilled labor, reshaping firms’ internal skill structures [38,39]. Moreover, robotics and automation technologies are accelerating the demand for skilled workers, with labor markets exhibiting a persistent absorption effect for high-skilled labor [40,41]. In the transition of employment structures, only workers with higher education levels or specialized skills can secure reemployment in roles that require greater creativity and complexity [42]. By contrast, low-skilled workers face a higher risk of displacement and may encounter long-term unemployment challenges. Based on this analysis, this study proposes H1 and H2.

H1. 

Artificial intelligence has a significant negative impact on the demand for low-skilled labor.

H2. 

Artificial intelligence has a significant positive impact on the demand for high-skilled labor.

Currently, scholars have not yet reached a consensus on the impact of AI technology on the employment demand for middle-skilled workers. Some studies suggest that job polarization exists in the labor market, meaning that middle-skilled workers engaged in routine manual labor and low-cognitive tasks will be adversely affected by AI, leading to a decline in the number of middle-skilled jobs while employment in both low-skilled and high-skilled positions increases [34,43]. However, other researchers argue that AI has little effect on the employment demand for middle-skilled workers [44,45]. Still, some studies indicate that AI might even increase demand for middle-skilled workers. In technology-intensive and knowledge-intensive industries, as automation and AI technologies advance, the demand for middle-skilled jobs may grow [46]. While automation may reduce middle-skilled jobs that rely heavily on routine tasks, the redistribution of tasks could sustain the demand for middle-skilled labor [47]. Although AI may replace traditional middle-skilled occupations focused on routine tasks, it can also enhance productivity, thereby increasing the demand for middle-skilled workers engaged in non-routine tasks [48]. In AI-assisted work environments, middle-skilled workers can upgrade their adaptive skills to avoid complete displacement and secure positions that complement AI systems [49]. Based on this analysis, this study proposes H3.

H3. 

Artificial intelligence has a significant positive impact on the demand for middle-skilled labor.

The application of artificial intelligence technology imposes higher technical demands on research and development (R&D) departments, compelling them to engage in continuous innovation to adapt to job transformations driven by emerging technologies. According to the theory of technological innovation diffusion, the advancement of AI increases the demand for innovative talent, fostering a clustering effect that establishes talent hubs and attracts high-skilled professionals. AI induces innovation across multiple industries, driving technological advancements and economic transformation [50]. As both an initiator and enabler of innovation, AI plays a pivotal role in accelerating various forms of technological progress [51]. The mechanism through which AI promotes technological innovation operates via several pathways; it accelerates knowledge creation and technological spillovers, enhances learning and absorptive capacities, and boosts investments in R&D and human capital, ultimately fostering innovation [52].

In the short term, technological innovation increases the demand for key production inputs needed to develop new products. However, as process innovation enhances efficiency, labor demand gradually decreases, which may lead to higher unemployment rates. The theory of “creative destruction” suggests that new technologies not only eliminate outdated production models but also create new market opportunities, underscoring the dual role of innovation in economic development [53]. Empirical studies further support this view. Cirillo et al. (2017), based on data from European countries, found that technological innovation drives firms to gradually replace low-skilled jobs with high-skilled positions [54]. Similarly, Vivarelli (2014) highlights that, in the context of labor-saving and skill-biased technological innovations, shortages of skilled labor often result in the displacement of low-skilled workers [55]. As a result, technological innovation makes traditional tasks performed by low-skilled workers more vulnerable to automation, significantly reducing demand for low-skilled labor. At the same time, it generates new tasks and occupations, particularly in middle- and high-skilled roles, thereby fostering employment growth in these categories. Based on the above analysis, this study proposes H4.

H4. 

Technological innovation has a significant mediating effect on the negative impact of AI on the demand for low-skilled labor and the positive impact on the demand for middle- and high-skilled labor.

3. Empirical Strategy and Data Sources

3.1. Construction of the Linear Model

The model is specified as follows:

$$Skil{l}_{i,t}=a_0+{\alpha }_{1}Paten{t}_{i,t}+\delta C_{i,t}+{\mu }_t+{\omega }_i+{\epsilon }_{i,t}$$

(1)

In this model, i represents the province, t represents time, the dependent variable Skill_i,_t is the employment skill structure, the core explanatory variable Patent_i,_t is the provincial-level artificial intelligence technology development index, C_i,_t represents a series of provincial-level control variables, μ_t denotes time-fixed effects, ω_i represents provincial-fixed effects, ε_i,_t is the random error term. The coefficient α₁ is of primary interest in this study, as it reflects the impact of AI development on the regional employment skill structure.

To examine whether there exists a transmission mechanism of technological innovation in the impact of provincial-level AI development on the employment skill structure, this study follows the approach of Wen et al. [56] and constructs the mediation effect test models (2) and (3).

$$Y_{i,t}=b_0+b_{1}Paten{t}_{i,t}+\delta C_{i,t}+{\mu }_t+{\omega }_i+{\epsilon }_{i,t}$$

(2)

$$Skil{l}_{i,t}=c_0+c_{1}Paten{t}_{i,t}+c_2{Y}_{i,t}+\delta C_{i,t}+{\mu }_t+{\omega }_i+{\epsilon }_{i,t}$$

(3)

In model (2), the effect of AI development on the mediator variable, Y_i,_t, is tested, and in model (3), the impact of AI development and the mediator variable on the employment skill structure is examined to determine the existence of a mediation effect. The design of other variables follows that of model (1). The mediation effect testing procedure is as follows:

Step 1: Test the coefficient α₁ in model (1). If α₁ is significant, proceed; if α₁ is not significant, this indicates that the explanatory variable and the dependent variable are not related.

Step 2: Sequentially test the coefficients b₁ in model (2) and c₂ in model (3). If both are significant, proceed; if at least one is not significant, use the bootstrap method for further testing. If significant, continue testing; otherwise, stop.

Step 3: Test the coefficient c₁. If c₁ is not significant, this indicates a full mediation effect; if c₁ is significant, it indicates a partial mediation effect. In the case of partial mediation, α₁ reflects the total effect of AI development on the employment skill structure, c₁ reflects the direct effect of AI on the employment skill structure, and b₁ × c₂ reflects the indirect effect of AI on the employment skill structure through the mediator variable, i.e., the mediation effect.

3.2. Construction of the Threshold Model

Under the framework of the static threshold model, we investigate the nonlinear relationship between AI development levels and the employment skill structure. This model assumes that when the AI level (i.e., the threshold variable) reaches a specific value, the impact of AI on the employment skill structure will change significantly. The basic expression of the model is as follows:

$$Skil{l}_{i,t}=d_0+d_{1}Paten{t}_{i,t}+\gamma D\left(q_{i,t}\le \theta \right)+\delta C_{i,t}+{\mu }_t+{\omega }_i+{\epsilon }_{i,t}$$

(4)

Parameter d₁ is the regression coefficient of the explanatory variable, and D(q_i·_t ≤ θ) is an indicator function used to represent the threshold effect. This model introduces the threshold variable q_i,_t to examine the varying impacts of AI levels on the employment skill structure, with the assumption that the effect will change significantly when the threshold value θ is reached. The primary advantage of the static panel model lies in its simplicity and efficiency. Since it does not involve the lagged dependent variable, the static threshold model is effective in capturing the nonlinear relationships among variables.

3.3. Variable Selection and Measurement

3.3.1. Dependent Variable

Employment Skill Structure (Skill): The proportion of educational levels is used as an indicator to measure the labor force’s employment skill structure. Drawing from the methodology of Wang et al. [19] on the classification of workers by skill level, the share of the total employed population with different educational backgrounds is defined as the proportion of low-skilled labor employment (Low skill), middle-skilled labor employment (Middle skill), and high-skilled labor employment (High skill), corresponding to individuals with education levels of middle school or below, high school, and college or above, respectively.

3.3.2. Independent Variable

Level of Artificial Intelligence Development (Patent): This study utilizes data on AI patent applications to measure the development level of AI technology across different provinces [57,58]. The number of patent applications and granted patents jointly reflects a region’s or industry’s R&D investment in the field of artificial intelligence. Patent applications indicate the level of technological innovation activity, while granted patents, to some extent, reflect the feasibility and recognition of the technology. To ensure the robustness of our research findings, we conducted a robustness check by substituting AI patent grants for patent applications as the core explanatory variable in our analysis.

For the selection of AI-related keywords, this study draws on the methodologies of Wang (2022) and Yin (2023) [59,60]. Specifically, AI keywords were identified based on official policy documents, including the Guidelines for the Construction of the National New Generation AI Standard System, the Three-Year Action Plan for Promoting the Development of the New Generation AI Industry (2018–2020), the Guidelines for the Comprehensive Standardization System of the National AI Industry (2024 Edition), and the New Generation AI Development Plan. Additionally, references were made to Stanford University’s Artificial Intelligence Index Report 2021 and other relevant literature. In the preceding analysis, we systematically reviewed the core application scenarios and subfields of artificial intelligence technologies, and preliminarily extracted relevant technical terms and keywords. To ensure representativeness and general applicability, we conducted cross-validation across multiple sources and retained only those keywords that appeared in at least two authoritative references. This process enabled the construction of an accurate and comprehensive AI keyword lexicon.

To further improve the accuracy of keyword identification and minimize false positives, this study adopted the following control measures. First, keyword searches were confined to the title, abstract, and claims sections of patent documents to ensure the extracted keywords are closely related to the core technical content of the patents. Second, a keyword combination matching rule was applied, as follows: a patent was classified as AI-related only if it included at least one term from the Core Technologies category (e.g., core sensors and big data) and one from the Application Domains category (e.g., smart education and smart cities). Third, a manual verification procedure was conducted by randomly sampling 500 patents from the collected dataset. The technical content of each sampled patent was manually compared against the keyword matching results. The observed false positive rate was below 5%, confirming the robustness and effectiveness of the keyword identification strategy.

This study employed Python3.8 to extract patent data from the China National Knowledge Infrastructure (CNKI) patent database using AI-related keywords [61]. The dataset was constructed based on the number of AI patent applications and grants, providing the foundation for the research analysis. The patent retrieval process focused on key elements, such as titles, abstracts, and claims, ensuring that patents containing relevant AI keywords are classified accordingly. Drawing on policy documents and prior academic research, this study develops a comprehensive AI keyword framework, covering both core technologies and application domains, as follows:

-

Core Technologies: core sensors, image recognition, speech recognition, big data, machine learning, deep learning, cloud computing, intelligent chips, computer vision, human–computer interaction, knowledge graphs, and natural language processing;

-

Application Domains: autonomous driving, smart security, smart cities, smart education, smart finance, smart agriculture, smart healthcare, intelligent robotics, smart homes, smart transportation, intelligent devices, smart logistics, and smart manufacturing.

Additionally, we extracted key information for each patent, including the application date, publication date, grant announcement date, applicant, and application address. The application date is used to determine the patent’s corresponding year, while the publication and grant announcement dates provide insight into its authorization status. The application address allows for the identification of the province to which the patent belongs. Based on this data, we aggregate the number of AI patents in each province and use it as an indicator to measure the level of AI technology development across regions.

3.3.3. Mediator Variable

Technological Innovation (Innovation): Technological innovation is measured through the following two indicators: investment in innovation funding and the concentration of innovative talent. Specifically, innovation funding is measured by internal R&D expenditures, while the concentration of innovative talent is quantified using the full-time equivalent of R&D personnel.

3.3.4. Control Variables

Based on previous classical literature, the following control variables were selected: The logarithm of the per capita regional GDP (Pergdp); level of urbanization (Urban), measured as the ratio of the urban population to the total population; degree of openness (Open), measured by the ratio of the total goods imported and exported to the regional GDP; level of technology market development (Tech), calculated as the ratio of the technology market transaction volume to the regional GDP. To eliminate heteroscedasticity, the following were used: logarithm of technology market development; level of transportation infrastructure (Transportation), measured by the logarithm of the road mileage; level of social consumption (Consumption), measured by the ratio of the total retail sales of social consumer goods to the regional GDP.

3.4. Data Sources

Since 2010, third-generation artificial intelligence technology, driven primarily by machine learning algorithms, has undergone rapid evolution and widespread integration across various sectors worldwide. As a general-purpose technology, AI has emerged as a key driver of the latest wave of technological and industrial transformation. This study utilizes panel data from 2010 to 2022, covering 30 provinces in China (excluding Hong Kong, Macau, Taiwan, and Tibet) to examine the impact of regional AI development on employment skill structures. The data sources include the China Labor Statistical Yearbook, China Statistical Yearbook, National Bureau of Statistics, and various provincial statistical yearbooks. AI development levels are measured using data from the China Patent Database. Missing data points are addressed through linear interpolation. Descriptive statistics for the key variables are presented in

Table 1. Descriptive statistics results.

Variable	N	Mean	Standard Deviation	Minimum	Maximum
log_Patent	390	9.614	1.728	4.615	13.061
Low skill (%)	390	63.309	12.411	19.5	86.468
Middle skill (%)	390	15.077	3.998	5.633	25.7
High skill (%)	390	19.727	10.428	6.489	65.341
log_Pergdp	390	10.821	0.488	9.464	12.156
Urban	390	0.595	0.124	0.338	0.896
Open	390	0.276	0.290	0.008	1.464
log_Tech	390	−4.941	1.375	−8.577	−1.654
Transportation	390	2.456	0.077	2.240	2.558
Consumption	390	0.385	0.058	0.180	0.504

.

4. Results

4.1. Baseline Results

The baseline regression results of Equation (1) are presented in

Table 2. Baseline regression results.

Variable	Low Skill	Middle Skill	High Skill
	(1)	(2)	(3)
log_Patent	−2.047 ***	1.192 ***	1.174 **
	(−3.61)	(3.37)	(2.61)
log_Pergdp	4.152	8.038 ***	−14.44 ***
	(1.79)	(5.57)	(−7.73)
Urban	−0.626	0.122	−0.0917
	(−0.32)	(0.10)	(−0.06)
Open	−2.208	−1.910	4.516 ***
	(−1.32)	(−1.83)	(3.66)
log_Tech	−0.181	−0.409 *	0.716 ***
	(−0.66)	(−2.40)	(3.35)
Transportation	1.506	0.459	−1.454
	(1.45)	(0.71)	(−1.83)
Consumption	35.71 ***	4.229	−41.80 ***
	(3.60)	(0.68)	(−5.41)
Constant	24.62	−73.43 ***	163.4 ***
	(1.14)	(−5.45)	(9.53)
Time-fixed effect	YES	YES	YES
Province-fixed effect	YES	YES	YES
N	390	390	390
R2	0.718	0.490	0.752

T-values are in parentheses. ***, **, and * indicate significance at the 0.1%, 1%, and 5% levels, respectively. The same applies to subsequent tables.

. Column (1) shows that artificial intelligence has a significantly negative effect on the demand for low-skilled labor, indicating that AI reduces the need for low-skilled workers and, thus, confirms H1. Column (2) presents the regression results for middle-skilled labor, showing a significantly positive coefficient, which suggests that AI significantly increases the demand for middle-skilled workers, validating H3. Column (3) displays the regression results for high-skilled labor, where the significantly positive coefficient indicates that AI significantly enhances the demand for high-skilled workers, supporting H2. All three regressions in Columns (1) to (3) include both time-fixed effects and province-fixed effects.

Low-skilled workers typically engage in highly repetitive, well-defined, and easily standardized tasks, which are most susceptible to automation driven by artificial intelligence. As a result, the adoption of AI significantly reduces the demand for low-skilled labor. On the other hand, according to the theories of productivity effects and creation effects, AI enhances the overall production efficiency while also contributing to employment growth. Specifically, AI boosts the productivity of middle-skilled workers with complementary skills, helping to expand employment in this group and creating new roles such as technical support and data analysis. In addition, the development, maintenance, and management of AI systems require advanced knowledge and technical skills, which leads to strong growth in high-skilled employment and greater demand for highly skilled labor.

4.2. Threshold Effect Test

The impact of artificial intelligence on the employment skill structure is multidimensional and may exhibit phase-specific variations depending on the level of AI technological development. To explore the potential existence of a nonlinear relationship and assess whether technological advances in AI help optimize the employment structure, this study conducts a threshold effect test.

Table 3. Threshold effect test results.

Labor Force	Threshold Model	Threshold Value	F-Statistic	p-Value	Critical Value at Different Significance Levels			Confidence Interval
					1%	5%	10%
Low skill	Single threshold	9.834	31.820	0.027	35.900	26.945	23.047	[9.794, 9.854]
Middle skill	Single threshold	9.913	28.840	0.090	43.896	33.091	32.896	[9.822, 9.980]
High skill	Single threshold	9.702	14.770	0.080	19.843	16.491	14.300	[9.554, 9.790]

presents the threshold effect test results for labor with different skill levels. Using the logarithmic value of AI technological development as the threshold variable, the results indicate a single threshold effect for low-skilled labor, with a threshold value of 9.834. However, no threshold effect is observed for both middle-skilled and high-skilled labor. Figure 2 shows the likelihood ratio curve from the single-threshold model for low-skilled labor, which helps pinpoint the exact value of the optimal threshold.

The threshold regression results in

Table 4. Threshold effect regression results.

Variable	Low Skill	Middle Skill	High Skill
	(1)	(2)	(3)
log_Patent(log_Patent ≤ 9.834)	−2.133 *
	(−2.20)
log_Patent(log_Patent > 9.834)	−2.545 **
	(−2.84)
log_Patent(log_Patent ≤ 9.913)		0.782
		(1.55)
log_Patent(log_Patent > 9.913)		1.041 *
		(2.15)
log_Patent(log_Patent ≤ 9.702)			1.203
			(1.56)
log_Patent(log_Patent > 9.702)			1.415
			(1.82)
log_Pergdp	−6.060 **	1.498	4.747 **
	(−3.30)	(1.50)	(2.98)
Urban	3.483	−6.258 *	2.207
	(1.73)	(−2.06)	(0.51)
Open	−0.409	−1.962	2.450
	(−0.15)	(−0.93)	(0.71)
log_Tech	−1.361 **	0.223	1.261 **
	(−2.98)	(0.79)	(2.89)
Transportation	6.770 ***	−2.355 ***	−4.300 ***
	(5.92)	(−3.86)	(−4.68)
Consumption	−15.18	5.408	10.72
	(−1.33)	(0.75)	(1.12)
Constant	72.43 ***	22.73 **	3.211
	(7.17)	(3.40)	(0.36)
Time-fixed effect	YES	YES	YES
Province-fixed effect	YES	YES	YES
R2	0.557	0.238	0.585
F-statistic	25.410	9.820	18.240

indicate that for low-skilled labor, the regression coefficient is −2.133 before reaching the threshold value and −2.545 after surpassing it. This suggests that as AI technology advances, its substitution effect on low-skilled labor intensifies. Advanced algorithms and automation technologies now enable AI to take over repetitive manual tasks and increasingly move into areas such as clerical work, customer service, and logistics—shifting from a supportive tool to a fully autonomous system. Additionally, declining AI costs and rising labor costs encourage firms to adopt automation technologies to enhance efficiency. Moreover, the weaker adaptability of low-skilled workers further exacerbates their risk of being replaced. Without skill upgrading to transition into more complex roles, this group faces increasing employment challenges.

AI exhibits no threshold effect on middle-skilled labor, indicating that its impact remains steady across different levels of technological advancement, without significant shifts or turning points. This stability can be attributed to the adaptability of middle-skilled positions, which can gradually integrate AI technologies. Likewise, no threshold effect is found for high-skilled labor, as these workers are central to AI development and continue to be in strong demand, supporting a stable human–AI collaboration within enterprises.

4.3. Mediation Effect of Technological Innovation

This study examines the mediating role of technological innovation in the relationship between AI and employment skill structures. Technological innovation is measured through the following two key indicators: innovation talent concentration and innovation funding investment, represented by the logarithm of full-time equivalent R&D personnel and the logarithm of internal R&D expenditure, respectively.

Table 5. The impact of artificial intelligence on technological innovation.

Variable	RD_erf	RD_pfe
	(1)	(2)
log_Patent	0.0722 *	0.187 ***
	(2.35)	(5.43)
log_Pergdp	0.498 ***	−0.121
	(3.88)	(−0.84)
Urban	0.180	0.145
	(1.63)	(1.17)
Open	0.0904	−0.163
	(0.93)	(−1.49)
log_Tech	−0.0928 ***	−0.0992 ***
	(−5.97)	(−5.67)
Transportation	−0.00287	−0.165 **
	(−0.05)	(−2.60)
Consumption	0.243	−0.594
	(0.44)	(−0.96)
Constant	6.394 ***	11.23 ***
	(5.44)	(8.49)
Time-fixed effect	YES	YES
Province-fixed effect	YES	YES
N	390	390
R2	0.669	0.655

presents the regression results of AI on technological innovation. Column (1) reports the impact of AI on innovation talent concentration, showing a significant positive effect. This suggests that the widespread application of AI effectively attracts and integrates innovation talent, creating more opportunities for professional growth and enhancing talent concentration effects. Column (2) shows that AI significantly increases R&D investment, highlighting its role in boosting economic returns, building an innovation ecosystem, and improving resource allocation to support economic transformation.

Table 6. Mediation effect test of technological innovation.

Variable	Low Skill	Middle Skill	High Skill	Low Skill	Middle Skill	High Skill
	(3)	(4)	(5)	(6)	(7)	(8)
log_Patent	−1.944 ***	1.102 **	1.150 *	−1.806 **	0.955 **	1.144 *
	(−3.41)	(3.11)	(2.53)	(−3.08)	(2.63)	(2.46)
RD_erf	−1.448 **	1.261 *	0.256
	(−2.48)	(2.13)	(0.33)
RD_pfe				−1.364 **	1.339 *	0.180
				(−2.57)	(2.49)	(0.26)
log_Pergdp	4.907 *	7.381 ***	−14.55 ***	4.022	8.166 ***	−14.44 ***
	(2.08)	(5.03)	(−7.65)	(1.74)	(5.70)	(−7.71)
Urban	−0.391	−0.0827	−0.119	−0.417	−0.0829	−0.128
	(−0.20)	(−0.07)	(−0.08)	(−0.21)	(−0.07)	(−0.08)
Open	−2.165	−1.947	4.514 ***	−2.531	−1.593	4.573 ***
	(−1.30)	(−1.88)	(3.65)	(−1.51)	(−1.53)	(3.64)
log_Tech	−0.325	−0.283	0.742 **	−0.309	−0.284	0.731 **
	(−1.13)	(−1.58)	(3.25)	(−1.08)	(−1.61)	(3.30)
Transportation	1.499	0.466	−1.455	1.249	0.711	−1.419
	(1.44)	(0.72)	(−1.83)	(1.19)	(1.09)	(−1.76)
Consumption	36.32 ***	3.698	−41.91 ***	35.09 ***	4.843	−41.82 ***
	(3.66)	(0.60)	(−5.41)	(3.54)	(0.79)	(−5.40)
Constant	33.58	−81.23 ***	161.6 ***	40.07	−88.60 ***	161.5 ***
	(1.50)	(−5.85)	(9.01)	(1.69)	(−6.03)	(8.64)
Time-fixed effect	YES	YES	YES	YES	YES	YES
Province-fixed effect	YES	YES	YES	YES	YES	YES
N	390	390	390	390	390	390
R2	0.720	0.497	0.752	0.721	0.500	0.752

presents the regression results of AI and technological innovation on employment skill structures. Column (3) reports the regression results of AI and innovation talent concentration on low-skilled labor demand, indicating a partial mediating effect of innovation talent concentration in the relationship between AI and low-skilled labor demand. Column (4) examines AI and the innovation talent concentration in relation to middle-skilled labor demand, also revealing a partial mediating effect. Column (5) assesses AI and innovation talent concentration in relation to high-skilled labor demand, showing no significant mediating effect. Column (6) reports the regression of AI and innovation funding investment on low-skilled labor demand, suggesting a partial mediating effect. Column (7) examines AI and innovation funding investment in relation to middle-skilled labor demand, also revealing a partial mediating effect. Column (8) assesses AI and innovation funding investment in relation to high-skilled labor demand, showing no mediating effect. These findings suggest that technological innovation significantly mediates AI’s impact on low- and middle-skilled labor demand but has no observable effect on high-skilled labor, thus largely confirming H4. This disparity may arise from the complexity and diversity inherent in high-skilled tasks. High-skilled occupations often involve unstructured activities such as problem-solving, systems thinking, and interdisciplinary collaboration. These tasks are less dependent on technological innovation but are more influenced by the advancements in AI technologies themselves.

4.4. Heterogeneity Analysis

4.4.1. Gender Heterogeneity Test

The impact of artificial intelligence on labor demand may vary by gender. Conducting a gender heterogeneity analysis helps uncover AI’s differential effects on employment opportunities for different genders, providing a scientific basis for formulating more inclusive and equitable policies. Column (1) and Column (4) of

Table 7. Gender heterogeneity test.

Variable	Male			Female
	Low Skill	Middle Skill	High Skill	Low Skill	Middle Skill	High Skill
	(1)	(2)	(3)	(4)	(5)	(6)
log_Patent	−1.925 ***	1.118 **	0.989 *	−2.149 ***	1.292 **	0.648
	(−3.34)	(3.29)	(2.38)	(−3.73)	(3.19)	(1.22)
log_Pergdp	5.690 *	5.028 ***	−11.74 ***	2.163	12.15 ***	−16.80 ***
	(2.42)	(3.63)	(−6.70)	(0.92)	(7.34)	(−8.13)
Urban	0.0787	0.694	−0.940	−1.560	−0.676	0.580
	(0.04)	(0.59)	(−0.67)	(−0.78)	(−0.48)	(0.32)
Open	−3.456 *	−0.707	4.243 ***	−0.282	−3.608 **	3.798 **
	(−2.03)	(−0.71)	(3.53)	(−0.17)	(−3.02)	(2.79)
log_Tech	−0.305	−0.335 *	0.723 ***	0.0262	−0.533 **	0.721 **
	(−1.10)	(−2.05)	(3.53)	(0.09)	(−2.72)	(2.99)
Transportation	1.098	0.730	−1.487	1.675	0.357	−0.932
	(1.04)	(1.17)	(−1.94)	(1.59)	(0.48)	(−1.06)
Consumption	35.54 ***	1.237	−39.56 ***	34.85 ***	8.822	−43.14 ***
	(3.52)	(0.21)	(−5.31)	(3.46)	(1.24)	(−4.98)
Constant	12.04	−46.62 ***	140.7 ***	44.75 *	−113.0 ***	181.0 ***
	(0.55)	(−3.60)	(8.73)	(2.04)	(−7.31)	(9.26)
Time-fixed effect	YES	YES	YES	YES	YES	YES
Province-fixed effect	YES	YES	YES	YES	YES	YES
N	390	390	390	390	390	390
R2	0.704	0.484	0.725	0.719	0.483	0.786

present the regression results of AI on the demand for low-skilled male and female labor, respectively. The results indicate that AI has a significant negative impact on the demand for both low-skilled male and female workers, with a more pronounced negative effect on low-skilled female workers. This disparity may be explained by occupational segregation and gender-based divisions of labor, where women are more often employed in clerical, service, and assembly positions that involve repetitive tasks highly susceptible to automation. In contrast, low-skilled male workers are more commonly employed in physically intensive, on-site roles, such as those in construction, transportation, and logistics. These occupations are relatively more resistant to automation in the short term, resulting in a comparatively smaller impact. Column (2) and Column (5) of Table 7 report the regression results of AI on the demand for middle-skilled male and female labor, revealing that AI has a significant positive impact on both groups.

Columns (3) and (6) of Table 7 present the regression results of AI on the demand for high-skilled male and female labor. The findings indicate that AI has a significant positive impact on high-skilled male labor demand, whereas its effect on high-skilled female labor is not statistically significant. This disparity may reflect the underrepresentation of women in high-skilled fields, particularly in STEM, where AI-related positions such as algorithm design and system development are expanding. Furthermore, gender disparities in access to technology and career advancement opportunities may limit the extent to which high-skilled women can benefit from the expansion of AI-related employment.

4.4.2. Factor Intensity Heterogeneity Test

Factor intensity is measured by the ratio of the total fixed asset investment to the number of employed persons at the end of the year. A province is classified as capital-intensive if its factor intensity exceeds the national average in a given year; otherwise, it is classified as labor-intensive. Columns (1) to (3) of

Table 8. Factor intensity heterogeneity test.

Variable	Capital-Intensive Provinces			Labor-Intensive Provinces
	Low Skill	Middle Skill	High Skill	Low Skill	Middle Skill	High Skill
	(1)	(2)	(3)	(4)	(5)	(6)
log_Patent	−1.454	0.922	0.849	−3.590 **	1.082	2.530 *
	(−1.68)	(1.75)	(1.43)	(−2.82)	(1.55)	(2.54)
log_Pergdp	10.29 *	2.885	−15.20 ***	3.168	7.226 ***	−11.62 ***
	(2.02)	(0.93)	(−4.58)	(0.95)	(3.95)	(−4.42)
Urban	4.157	−7.142 **	2.822	−0.475	2.258	−2.016
	(1.17)	(−3.29)	(1.27)	(−0.13)	(1.08)	(−0.57)
Open	−3.776	3.396 *	0.759	0.543	−4.562 **	4.147
	(−1.41)	(2.09)	(0.46)	(0.19)	(−2.90)	(1.95)
log_Tech	−1.532 *	0.893 *	0.830	0.272	−0.632 ***	0.475
	(−2.27)	(2.18)	(1.93)	(0.82)	(−3.46)	(1.80)
Transportation	0.696	0.298	−0.492	1.822	0.294	−1.824
	(0.35)	(0.25)	(−0.40)	(1.18)	(0.35)	(−1.52)
Consumption	31.30	10.51	−43.64 ***	55.99 ***	−13.64	−48.82 ***
	(1.92)	(1.06)	(−4.21)	(3.49)	(−1.55)	(−3.78)
Constant	−32.95	−18.23	164.3 ***	33.32	−58.04 **	136.1 ***
	(−0.69)	(−0.63)	(5.37)	(0.98)	(−3.09)	(5.05)
Time-fixed effect	YES	YES	YES	YES	YES	YES
Province-fixed effect	YES	YES	YES	YES	YES	YES
N	92	92	92	298	298	298
R2	0.734	0.514	0.806	0.702	0.567	0.716

present the regression results for capital-intensive provinces, showing that AI does not have a significant impact on the demand for low-, middle-, or high-skilled labor in these regions. Columns (4) to (6) display the regression results for labor-intensive provinces, indicating that AI has a significant negative impact on low-skilled labor demand, no significant effect on middle-skilled labor, and a significant positive impact on high-skilled labor. The lack of a significant effect on middle-skilled labor demand can be attributed to the competing substitution and complementarity effects within these occupations, where the opposing forces balance out each other.

The above analysis indicates that the impact of artificial intelligence on employment skill structures is more evident in labor-intensive provinces. The primary reason for this disparity lies in the differences in industrial and labor structures across regions, which lead to distinct paths of technological substitution and upgrading. Labor-intensive regions are generally dominated by low-value-added sectors, such as manufacturing and light industry, which heavily depend on low-skilled labor. The introduction of artificial intelligence directly substitutes a large number of standardized and repetitive jobs, thereby triggering significant adjustments in the employment structure. In contrast, capital-intensive regions, where AI is more advanced, are typically dominated by high-tech industries like finance, information services, and advanced manufacturing. These regions have a larger proportion of high-skilled labor, meaning that the impact of AI integration on low-skilled employment is less pronounced. Additionally, the organizational structures and management mechanisms of enterprises in these regions are more developed, with stronger capabilities for technological absorption and mechanisms to buffer employment transitions, facilitating a smoother structural shift during technological substitution.

4.5. Endogeneity Treatment

When directly regressing AI development levels on the employment skill structure across provinces, there may be endogeneity bias due to potential reverse causality between AI adoption and employment skill composition. To address this concern and reduce the endogeneity arising from the bidirectional relationship between AI adoption and labor demand, this study uses the instrumental variable (IV) approach. Specifically, we select the one-period-lagged AI patent count for each province as the instrument and apply the two-stage least squares (2SLS) method. The regression results are presented in

Table 9. Endogeneity test.

Variable	Low Skill		Middle Skill		High Skill
	1SLS	2SLS	1SLS	2SLS	1SLS	2SLS
	(1)	(2)	(3)	(4)	(5)	(6)
log_Patent	0.978 ***	−1.355 ***	0.978 ***	0.509	0.976 ***	1.528 ***
	(71.21)	(−2.87)	(71.21)	(1.32)	(65.66)	(3.43)
rk LM p-value		0.000		0.000		0.000
rk F-statistic		507		507		507
log_Pergdp	0.067	−6.711 *	0.067	5.261 **	0.071	2.034
	(1.31)	(−2.56)	(1.31)	(2.76)	(1.45)	(1.09)
Urban	−0.086	−7.972	−0.086	1.049	−0.072	6.893 *
	(−0.69)	(−1.63)	(−0.69)	(0.25)	(−0.53)	(2.55)
Open	0.081	−1.948	0.081	−0.0724	0.073	1.635
	(1.80)	(−0.66)	(1.80)	(−0.04)	(1.65)	(0.63)
log_Tech	−0.004	−1.428 *	−0.004	0.0773	−0.003	1.414 *
	(−0.37)	(−2.08)	(−0.37)	(0.23)	(−0.20)	(2.41)
Transportation	0.059 *	5.452 ***	0.059 *	−2.163 ***	0.058 ***	−3.220 ***
	(2.42)	(4.87)	(2.42)	(−3.33)	(2.34)	(−3.64)
Consumption	0.567	3.787	0.567	0.689	0.537	−2.596
	(1.93)	(0.21)	(1.93)	(0.08)	1.71	(−0.18)
Constant	−1.219 *	80.78 **	−1.219 *	−12.89	−1.179 ***	25.34
	(−2.17)	(2.78)	(−2.17)	(−0.67)	(−2.22)	(1.34)
Time-fixed effect	YES	YES	YES	YES	YES	YES
Province-fixed effect	YES	YES	YES	YES	YES	YES
N	390	390	390	390	390	390
R2		0.831		0.731		0.771

. Columns (1) and (2) report the estimates for low-skilled labor demand, columns (3) and (4) for middle-skilled labor demand, and columns (5) and (6) for high-skilled labor demand.

The results show that the rk F-statistics exceed the critical value, indicating a strong correlation between the instrumental variable and the key explanatory variable, thereby ruling out the weak instrument problem. Additionally, the rk LM test yields a p-value of 0, rejecting the null hypothesis and confirming that the model does not suffer from under-identification. The instrumental variable estimates reveal that AI significantly reduces the demand for low-skilled labor while increasing the demand for high-skilled labor, with no significant effect on middle-skilled labor. These findings provide robust support for the study’s hypotheses, reinforcing the conclusion that AI reshapes employment structures by displacing low-skilled jobs and fostering high-skilled employment opportunities.

4.6. Robustness Check

To enhance the credibility of the model’s estimation results, this study conducted robustness checks by employing core explanatory variable substitution, placebo tests, and the inclusion of additional control variables.

4.6.1. Core Explanatory Variable Substitution

In this study, AI patent applications were used as the core explanatory variable to measure the level of AI technology development. However, while the number of patent applications primarily reflects the intensity of the technological innovation, the granting process involves rigorous evaluation, making the number of granted patents a stronger indicator of technological feasibility and impact. To further validate the robustness of the research findings and examine the impact of AI technology on employment structures from different perspectives, we conducted a robustness check by using AI patent grants as an alternative variable.

Table 10. Replacing the core independent variable.

Variable	Low Skill	Middle Skill	High Skill
	(1)	(2)	(3)
log_patent_grant	−1.845 ***	1.058 ***	1.012 **
	(−3.50)	(3.60)	(2.80)
log_pergdp	7.436 ***	1.024	−6.994 ***
	(3.58)	(0.76)	(−4.63)
Urban	−57.99 ***	67.33 ***	−5.808
	(−5.34)	(9.59)	(−0.74)
Open	4.029 **	3.456 ***	−4.982 ***
	(2.60)	(3.45)	(−4.43)
log_tech	0.345	−0.348	−0.0646
	(1.24)	(−1.93)	(−0.32)
Transportation	−0.399	0.151	0.308
	(−0.26)	(0.15)	(0.28)
Consumption	16.37 **	8.893 *	−10.21 *
	(2.70)	(2.27)	(−2.31)
Constant	33.61	−32.16 *	76.89 ***
	(1.51)	(−2.23)	(4.75)
Time-fixed effect	YES	YES	YES
Province-fixed effect	YES	YES	YES
N	390	390	390
R2	0.894	0.708	0.887

presents the effects of AI patent grants on the employment skill structure. The empirical results indicate that AI technology development exerts a significant negative impact on the demand for low-skilled labor while having a significant positive effect on the demand for middle- and high-skilled labor. These findings are consistent with the baseline regression results, further confirming the robustness of the research conclusions.

4.6.2. Placebo Test

To further validate the robustness of the model, a placebo test was conducted by examining whether future AI development levels influence past employment skill structures. Specifically, the study constructed a five-year-lagged AI patent dataset and restricted the analysis to the period from 2010 to 2015. This test aims to determine whether future AI levels have a significant impact on prior employment skill structures, thereby assessing the validity and robustness of the model. By employing this approach, potential reverse causality between the dependent and independent variables over time can be effectively ruled out, ensuring the reliability of the causal inference. The placebo test results, as presented in

Table 11. Placebo test.

Variable	Low Skill	Middle Skill	High Skill
	(1)	(2)	(3)
log_Patent_future	−0.128	−0.533	1.517
	(−0.09)	(−0.63)	(1.43)
log_Pergdp	10.31	−0.392	−11.53 *
	(1.59)	(−0.10)	(−2.42)
Urban	−7.779	25.01	−24.09
	(−0.27)	(1.43)	(−1.23)
Open	3.642	−0.978	−3.548
	(1.32)	(−0.58)	(−1.98)
log_Tech	−0.545	−0.0530	0.718
	(−0.87)	(−0.14)	(1.58)
Transportation	−0.266	1.135	−0.507
	(−0.05)	(0.37)	(−0.15)
Consumption	10.14	3.069	−15.67
	(0.43)	(0.22)	(−0.97)
Constant	−21.05	−3.252	136.0 *
	(−0.26)	(−0.07)	(2.48)
Time-fixed effect	YES	YES	YES
Province-fixed effect	YES	YES	YES
N	180	180	180
R2	0.274	0.263	0.425

, indicate that AI development levels have no significant effect on the demand for low-skilled, middle-skilled, or high-skilled labor. These findings confirm the robustness of the baseline model and reinforce the credibility of the study’s empirical conclusions.

4.6.3. Additional Control Variables

To enhance the robustness of the research findings, this study incorporates the following four additional control variables: (1) R&D intensity (R&D); (2) financial development level (Financial), measured by the ratio of the sum of deposits and loans of financial institutions to the regional GDP; (3) population density (Population), calculated as the ratio of the province’s resident population at the end of the year to its land area; and (4) unemployment rate (Unemployment).

Table 12. Adding control variables.

Variable	Low Skill	Middle Skill	High Skill
	(1)	(2)	(3)
log_Patent	−2.019 ***	1.273 ***	1.018 *
	(−3.54)	(3.88)	(2.46)
log_Pergdp	3.104	1.263	−5.587 **
	(1.11)	(0.78)	(−2.66)
Urban	−1.138	4.544 ***	−4.262 **
	(−0.51)	(3.50)	(−2.61)
Open	−2.350	−1.926	4.546 ***
	(−1.36)	(−1.93)	(3.87)
log_Tech	−0.0562	−0.361 *	0.508 *
	(−0.20)	(−2.26)	(2.54)
Transportation	1.249	0.491	−1.484 *
	(1.20)	(0.82)	(−2.03)
Consumption	39.49 ***	−3.201	−37.17 ***
	(3.84)	(−0.54)	(−5.05)
Research	11.52	9.017	−17.12
	(0.23)	(0.31)	(−0.47)
Banking	−1.757 *	0.0398	1.681 **
	(−2.19)	(0.09)	(2.99)
Population	−0.000638	−0.0123 ***	0.0130 ***
	(−0.23)	(−7.80)	(6.87)
Unemployment	−0.978 *	0.286	0.730 *
	(−2.09)	(1.06)	(2.09)
Constant	43.87	−8.474	71.86 ***
	(1.65)	(−0.55)	(3.66)
Time-fixed effect	YES	YES	YES
Province-fixed effect	YES	YES	YES
N	390	390	390
R2	0.725	0.573	0.793

displays the results of the impact of AI development on the employment skill structure after adding control variables. Columns (1), (2), and (3) respectively show the impact of AI on the demand for low-skilled, middle-skilled, and high-skilled labor. The results indicate that AI development has a significant negative impact on the demand for low-skilled labor, and a significant positive impact on the demand for both middle-skilled and high-skilled labor. These findings are consistent with the baseline regression results, further confirming the robustness of the research conclusions.

5. Conclusions

The key to achieving sustainable economic development lies in ensuring decent employment opportunities. This study aims to analyze the impact of AI adoption on employment across different skill levels (low-skilled, middle-skilled, and high-skilled), genders, and regional labor markets, with a particular focus on safeguarding workers’ rights, reducing poverty, and mitigating inequality amid technological transformation. Additionally, it underscores the importance of granting women equal employment rights and career development opportunities while improving the employment prospects and socioeconomic conditions of impoverished, minority, and other disadvantaged groups. This research is closely aligned with the United Nations Sustainable Development Goals, particularly in promoting inclusive growth and social equity.

This study employs panel data from 30 provinces in China and utilizes a two-way fixed-effects model, a mediation effect model, and a threshold effect model to empirically analyze the impact of AI development on China’s employment skill structure. The research findings indicate that artificial intelligence technology significantly suppresses the market demand for low-skilled labor. As AI technology continues to advance, the substitution pressure on low-skilled jobs steadily increases. The empirical results show that for every one logarithmic unit increase in the AI technology level, the demand for low-skilled labor decreases by approximately 2.047 percentage points on average, highlighting the strong substitution effect of AI on repetitive and procedural work, particularly among low-skilled groups. From a sustainable development perspective, the negative impact of AI on low-skilled workers hinders the achievement of SDG 8, “Decent Work and Economic Growth”, and reveals the unequal distribution of technological benefits. In the absence of effective skills upgrading and inclusive support policies, technological progress may further exacerbate the socioeconomic vulnerability of low-skilled groups, thus impeding the realization of SDG 10, which is “Reduced Inequality” by empowering vulnerable populations and promoting socioeconomic inclusion [5].

In contrast, the demand for high-skilled labor exhibits a significant positive relationship with the advancement of artificial intelligence technology. Specifically, for every one logarithmic unit increase in the AI technology level, the demand for high-skilled labor increases by approximately 1.174 percentage points. This finding aligns with Bessen’s (2019) conclusion, further confirming that the application of AI in highly specialized fields both replaces traditional jobs and creates new job opportunities [62].

In addition, the study also finds that the development of artificial intelligence technology has a significant positive impact on the demand for middle-skilled labor. Specifically, when the level of AI technology increases by one logarithmic unit, the demand for middle-skilled labor rises by approximately 1.192 percentage points. This finding differs from the “job polarization” hypothesis proposed by Autor et al. (2003), which suggests that middle-skilled jobs decline because of automation [26]. Similarly, Acemoglu and Restrepo (2018) argue that automation reduces the demand for middle-skilled labor [21]. This discrepancy reflects structural differences between the labor markets of the United States and China and is closely related to China’s relatively lower level of automation and the greater demand for workers with complementary technical skills.

From the perspective of the threshold effect, AI exhibits a threshold effect on the demand for low-skilled labor, whereas no such effect is observed for middle-skilled and high-skilled labor. Specifically, the negative impact of AI on low-skilled labor demand intensifies as AI technology advances, while its influence on middle-skilled and high-skilled labor remains stable.

Furthermore, this study examines the mediating role of technological innovation in the impact of AI on employment skill structures by conducting a mediation effect test, using innovation talent concentration and innovation funding investment as indicators. The findings reveal that technological innovation partially mediates the negative impact of AI on low-skilled labor demand and also plays a certain mediating role in the positive impact on middle-skilled labor demand. However, no significant mediating effect was observed for high-skilled labor demand. This mediation effect further underscores the severe employment shocks faced by low-skilled workers in the context of widespread AI adoption, with significantly increased unemployment risks. Arntz et al. (2016) previously discussed the challenges workers face in adapting to AI and automation, highlighting the urgency of addressing rising inequality and emphasizing the need to enhance job security for low-skilled workers [63]. This insight provides critical implications for this study, suggesting that future policy measures should focus on strengthening workforce retraining and adaptive support.

Regarding the heterogeneity tests based on gender and factor intensity, the study reveals significant disparities in the employment impact of AI across different demographic groups. Specifically, AI exerts a significant negative effect on the demand for both low-skilled male and low-skilled female workers, with the negative impact being more pronounced for low-skilled females. Moreover, AI positively influences the demand for middle-skilled males, middle-skilled females, and high-skilled males, whereas its effect on high-skilled females is not statistically significant. Manasi et al. (2022) highlighted issues of gender inequality and discrimination in AI applications [64]. The findings of this study align with their conclusions, suggesting that the widespread adoption of AI has substantial implications for female employment. The substitution effects of automation and intelligent technologies further exacerbate employment instability for women, potentially intensifying challenges related to their status and opportunities in the labor market. This trend may further deepen the existing gender segregation in the labor market, placing women at a more disadvantaged position in the ongoing technological transformation, thereby presenting a potential challenge to the realization of SDG 5, which aims to “achieve gender equality and empower all women and girls”. Therefore, in the process of advancing technological progress and economic structural transformation, it is essential to strengthen the identification and response to gender-related impacts, encourage the formulation of inclusive policies, and enhance women’s participation and sense of empowerment within the digital economy. These measures will contribute to the realization of gender equality objectives and promote the social sustainability of technological transitions.

Furthermore, the impact of AI on employment varies significantly between labor-intensive and capital-intensive provinces. In labor-intensive provinces, AI has a significant negative effect on the demand for low-skilled labor, a significant positive effect on the demand for middle-skilled labor, and no significant impact on high-skilled labor. In contrast, these effects are not statistically significant in capital-intensive provinces. This indicates that technological change not only has differentiated effects on regional economic structures and industrial forms but also poses new challenges to the economic and social sustainability at the regional level. Labor-intensive regions, which are heavily dependent on low-skilled labor, may experience the substitution effects of AI undermining their traditional advantages, thereby intensifying imbalances in employment structures and widening economic disparities. In contrast, the increase in middle-skilled labor demand in certain regions, fueled by technological progress, creates new opportunities for optimizing labor structures and upgrading industries. Therefore, a critical focus will be how to effectively integrate artificial intelligence with local industries and enhance the transformation capabilities of low-skilled workers, which will be key to fostering inclusive growth and sustainable development at the regional level. In the process of promoting AI technology, regions should give due attention to its effects on social equity, skill transformation, and regional coordinated development, ensuring that technological benefits do not exacerbate inequalities.

To verify the robustness of the research findings, this study conducted an endogeneity test, which confirmed that AI has a significant negative impact on the demand for low-skilled labor, no significant effect on middle-skilled labor, and a significant positive impact on high-skilled labor. These results further validate the reliability of the study’s conclusions. Additionally, robustness checks, including core explanatory variable substitution, placebo tests, and the inclusion of additional control variables, consistently support the conclusions of this study.

Based on the above analysis, the rapid advancement of AI technology has, indeed, intensified employment inequality across skill levels, gender, and regions, posing challenges to sustainable employment development. This conclusion aligns with the findings of Vivarelli (2014) and Cirillo et al. (2017), who highlighted the role of automation in exacerbating labor market disparities [54,55]. To effectively address the employment inequalities brought about by artificial intelligence and to ensure that low-skilled workers, female labor, and regional labor groups have equitable access to employment opportunities, this study puts forward a series of proposed measures. These measures aim to promote the sustainable development of employment while preserving inclusiveness and fairness in the labor market in the context of technological advancement.

6. Discussion

The impact of artificial intelligence on the employment skill structure is not only critical to employment and the sustainable development of society but also closely tied to social equity and stability. Based on the findings of this study, this section will examine the policy implications, practical significance, research limitations, and future research directions, aiming to offer valuable insights for policy making, industrial practices, and academic research.

6.1. Policy Implications

First, relevant government agencies should formulate targeted measures to promote the sustainable development of artificial intelligence technologies. This includes strengthening support for research and development in key and core areas of next-generation information technologies, thus enhancing the country’s overall capacity for technological innovation and independent R&D. Such efforts are essential to ensure that technological progress continues to play a sustained and positive role in driving socioeconomic development. Given that the application of AI may lead to employment restructuring and labor substitution, government support is widely regarded as a key mechanism for mitigating these challenges [65]. Therefore, differentiated and technology-coordinated development strategies should be adopted for various types of AI technologies. These strategies should avoid overemphasizing the substitution effects of AI on labor, and instead focus on fostering a virtuous cycle between technological evolution and the optimization of employment structures [23].

Second, the social security system should be improved, especially in unemployment benefits and transition support, by developing mechanisms that adapt to the intelligent economy, ensuring fairness and resilience during the transformation [33]. Measures such as unemployment insurance, vocational training subsidies, and minimum income guarantees can provide essential support for labor groups most affected by AI, helping them to enhance their re-employment capabilities and strengthening their adaptability and sense of security during periods of transition. In addition, policy frameworks should encourage cross-organizational collaboration and resource sharing, promoting coordinated efforts among enterprises, universities, and training institutions in technology integration and skills development [66]. This would improve workers’ adaptability to new technologies, reducing the risks of skill substitution and structural unemployment caused by AI, and strengthening the sustainability of labor policies in digital transformation.

Third, the orientation of talent cultivation and the structure of academic disciplines in higher and vocational education should be dynamically optimized. Efforts should be made to adjust and upgrade the higher education system, particularly by accelerating the development of interdisciplinary programs centered on intelligent manufacturing and artificial intelligence. By establishing emerging majors closely related to AI technologies and systematically training talent with integrated and cross-disciplinary skills, the education system can better respond to the evolving demands of new occupational forms and skill structures in the AI era.

Moreover, it is essential to improve a lifelong vocational skills training system that aligns with technological advancement and systematically enhances workers’ capabilities to learn and apply AI technologies. Special attention should be given to low-skilled labor groups who are at high risk of unemployment due to AI-driven substitution effects. In response to their employment vulnerability, a comprehensive set of intervention measures should be implemented, including targeted retraining programs, transitional employment support, career counseling, and job-matching services, in order to strengthen their transition capacity and competitiveness in the labor market [67,68,69]. Targeted retraining programs can improve low-skilled workers’ adaptability and reemployment chances through demand-driven curricula, flexible learning, and precise job matching. Transitional support offers economic security and a buffer period for displaced workers by creating short-term jobs, public initiatives, and temporary subsidies, reducing long-term unemployment risks. Career counseling and job-matching services, powered by AI and big data technologies, help accurately identify the alignment between labor skill profiles and job market demands, significantly improving job search efficiency and match quality.

At the same time, female workers—particularly those concentrated in traditional low-skilled occupations—face greater employment barriers and gender discrimination than their male counterparts amid the AI-driven economic transformation. Therefore, it is vital to pay close attention to the challenges and development prospects of women during employment structure adjustments. Through policy support and diversified intervention strategies—such as skill enhancement programs and flexible employment mechanisms—more high-quality and sustainable job opportunities should be created for women [70]. Skill enhancement programs should focus on providing women with ongoing learning and career development support, including digital skills training, leadership development, and cross-sectoral transition courses, aiming to improve their ability to adapt to industrial upgrading under intelligent and automated production systems. Flexible employment mechanisms, on the other hand, can enhance female labor market participation by optimizing labor regulations, improving social security systems, and promoting remote work, flexible working hours, and other adaptable employment arrangements.

Finally, the development of artificial intelligence and automation technologies has significantly heterogeneous impacts across different regions. Labor-intensive provinces are more vulnerable to the substitution effects of technology, whereas employment groups in capital-intensive provinces are relatively less affected. Accordingly, region-specific policies should be adopted. In labor-intensive provinces in particular, it is essential to strengthen vocational training and promote industrial upgrading, guiding the workforce toward technology-intensive or service-oriented sectors in an orderly manner. At the same time, the government should create differentiated development strategies, allocate innovation resources equitably, and reduce regional disparities in AI development and application. Balancing the regional differences in employment structures induced by technological change is essential for fostering a more inclusive and sustainable socioeconomic system [71].

6.2. Practical Implications

For enterprise managers, it is essential to proactively respond to the structural transformations induced by artificial intelligence by formulating strategic actions in both innovation and human resource management. First, enterprises should continuously enhance their technological innovation capabilities and increase R&D investment to promote the deep integration of intelligent technologies into production and management processes. By optimizing data-driven decision-making mechanisms, advancing intelligent automation, and fostering human–machine collaboration, firms can significantly improve productivity and market competitiveness. Particular attention should be paid to emerging intelligent technologies, including generative AI, which hold immense potential in areas such as knowledge generation, content creation, and automated analytics [72]. Leveraging these technologies can enhance firms’ innovation and technological capacity while contributing to sustainable development and societal progress.

Second, managers should enhance institutional arrangements and optimize resource allocation to better support low-skilled and female workers. In terms of skills development, enterprises can establish a capability-oriented training system that integrates job rotation, internal skill certification, and personalized learning pathways to improve the adaptability and upward mobility of low-skilled employees. Regarding gender equality, firms should work to create inclusive workplace environments by promoting flexible working arrangements, establishing clear promotion channels for women, and fostering supportive professional networks. These measures can effectively increase female participation and advancement in the labor market, thereby contributing to greater diversity and inclusion within the organization.

For workers, it is imperative to actively adapt to the structural changes in the labor market driven by artificial intelligence while proactively enhancing their skills and career development awareness. First and foremost, workers should adopt a lifelong learning mindset, continually updating and expanding their skillsets to better meet the new demands of human capital in the age of artificial intelligence [73]. Workers at different skill levels should adopt differentiated development strategies, as follows: low-skilled workers should focus on improving digital literacy and basic technical skills to better adapt to employment changes and mitigate the risk of being replaced by automation; middle-skilled workers should strengthen their cross-disciplinary capabilities and their ability to operate and apply intelligent technologies, thereby enhancing their career resilience and potential for transitioning to higher-value-added positions; high-skilled workers should deepen their professional knowledge, enhance their innovative capabilities and strategic thinking, and actively participate in the development, management, and optimization of AI technologies. Furthermore, workers should cultivate a strong sense of career planning, closely monitor industry trends and job requirements, and actively explore career development paths. Specifically, they can improve their mobility and future employment sustainability by participating in internal promotion mechanisms, skills retraining programs, and entrepreneurial and innovative practices, among other strategies.

6.3. Research Limitations and Future Research Directions

Firstly, this study primarily explored the overall impact of artificial intelligence on the employment skill structure from a macro perspective without delving into its heterogeneous effects across various industries. Given that the application pathways and technological substitution models of AI vary significantly between sectors such as manufacturing and services, the reshaping of the labor force skill structure may exhibit diverse characteristics. Therefore, the applicability of the research findings at the industry level remains limited. Future studies could fill this gap by using industry-specific data and conducting targeted empirical analyses, enhancing the research’s explanatory power and policy relevance.

Secondly, because of the limitations in the scope and research focus, this paper used a skills-based classification as the foundation for analysis and did not systematically introduce the “task perspective” to explore the mechanisms through which artificial intelligence impacts the labor market. However, in practical work settings, AI directly affects the specific tasks embedded in jobs, and an analysis framework based solely on skill levels may fail to fully reveal the dynamic adjustment mechanisms of the employment structure. Therefore, future research could combine task classification and substitutability methods to analyze AI’s effects on substitution, augmentation, and reorganization across tasks, offering deeper insight into how technological advancements impact employment structure.