JCS 本刊論文 | 從自動化技術到生成式人工智能——技術對勞動者影響的技能異質性研究

Welcome to follow

The Journal of Chinese Sociology

2025年11月17日，The Journal of Chinese Sociology（《中國社會學學刊》）上線文章From automation technology to generative AI: skill heterogeneity in technology’s impact on laborers（《從自動化技術到生成式人工智能——技術對勞動者影響的技能異質性研究》。

| 作者簡介

張詠雪

華中科技大學社會學院講師

主要研究方向：計算社會學、數字社會、AI的社會影響

Abstract

This is a study of skill heterogeneity in technology’s impact on laborers transitioning from physical automation to cognitive automation. The findings indicate that cognitive skills are a crucial determinant of the extent of technological influence. Considering both technological substitution and technological control, high-skilled and low-skilled workers experience limited technological substitution. However, high-skilled workers are not significantly affected by technological control, whereas low-skilled workers are subjected to stronger technological control. A comparison between automation technology and large language models (LLMs) reveals that the former primarily affects the secondary sector, while the latter mainly affects the tertiary industry. Furthermore, LLMs disproportionately influence women, younger demographics, professional skilled laborers, and higher-income groups. In the context of a new technological revolution led by artificial intelligence, analyzing and exploring the impact of technology holds significant theoretical and practical implications.

Keywords

Artificial intelligence; Automation; Cognitive skills; Processing effects; Heterogeneity

Introduction

The report from the 20th National Congress of the Communist Party of China emphasizes that “from now on, the central task of the Communist Party of China is to unite and lead people of all ethnic groups throughout the country to build a powerful modern socialist country comprehensively, achieve the second centenary goal, and comprehensively advance the great rejuvenation of the Chinese nation through Chinese modernization” (Xi 2022: 21). Technological breakthroughs are foundational to social transformation, with technological revolutions historically shaping and inextricably linked with modernization processes across diverse societies. The First Industrial Revolution initiated widespread mechanization and early forms of mechanical power automation. Building on this, the Second Industrial Revolution advanced through electrification and mass production, fostering more complex automated systems. The Third Industrial Revolution propelled society towards digital automation, driven by electronics and information technology, laying the groundwork for the subsequent development of cognitive automation (Jia 2016). Against this historical backdrop, the proliferation of new technologies centered on artificial intelligence has increasingly impacted high-skilled labor.

Since the 1990s, China has been progressively transitioning into a digital society (Qiu 2022). Over the past decade, China’s artificial intelligence (AI) industry has grown rapidly in scale. By 2019, China ranked second globally in the number of AI industry professionals and AI startups, and held the top position worldwide for the number of graduates in STEM (science, technology, engineering, and mathematics) fields (Zhang and Wang 2019). Furthermore, analysis from “The AI Index 2021 Annual Report” indicates that by the time of its publication, China’s AI penetration rate had reached first in the world, with the country also ranking first globally in the volume of AI-related academic papers, patents, and financing (Zhang et al. 2021). Given the rapid advancement of AI within China, the impact of this technology on high-skilled workers has emerged as a research imperative with considerable theoretical and practical salience.

Since 2022, the landscape of AI technology has been reshaped by large language models (LLMs), exemplified by OpenAI’s release of ChatGPT. The revolutionary breakthrough engendered by LLMs lies in their ability to transform AI’s functionality from mere natural language understanding to content generation. The types of tasks they can perform now encompass natural language understanding, natural language generation, the execution of knowledge-intensive tasks, and reasoning (Yang et al. 2024). In contrast to LLMs, early AI technologies broadly fall under the category of narrow or specialized AI (Qiu 2023). They are often called “weak AI” due to their singular, tool-like functions. Prominent examples include autonomous driving, chess-playing algorithms, machine vision (e.g., fingerprint, facial, and retinal recognition), expert systems, and automated planning (Zhai and Peng 2016). Given this context, AI’s impact on the labor market has become a central concern for contemporary scholars.

Literature review

and research question

Technological substitution

and technological control

Historically, technological development has been a primarily contributor to advancements in human productivity, a fundamental dynamic underpinning economic and social change. In neoclassical growth models, exogenous technological progress has been identified as the primary driver of long-term per capita output growth; similarly, in endogenous growth theory, the rate of endogenous technological change is a crucial factor influencing economic growth (Barro and Sala-I-Martin 1997). At the macro level, numerous studies have demonstrated technology’s positive impact on production efficiency (Yan and Wang 2004; Wang et al. 2006; Liu and Zhang 2008; Li and Dace 2020).

However, at the micro level, technological development can negatively affect laborers. Scholarly inquiry into the adverse effects of technology on labor typically delineates two main theoretical streams. The first views technology as a competitor to workers, primarily examining processes of worker substitution. In economics, the “technological substitution” perspective is a popular explanation for technology’s impact on the labor market. The second conceptualizes technology as a mechanism of control, investigating the degree of labor alienation workers experience following technological integration. Centered in sociology, it draws from labor process theory, using a “technological control” perspective to critically analyze workers’ conditions and experiences under technological influence.

In this context, technological substitution refers to the displacement of workers from tasks or jobs when machines can perform those same tasks more efficiently or at a lower cost (Nordhaus 2007). Studies from the United States indicate that as automation technology diffuses across the economy, it contributes to a decline in the labor share of national income and a decrease in the proportion of the employed population (Acemoglu and Restrepo 2018b). In the US manufacturing sector, the introduction of industrial robots has been shown to exert a negative effect on employment and wages (Acemoglu et al. 2022; Acemoglu and Restrepo 2022). In China, research indicates that while the increasing deployment of robots has been associated with gains in both wages and labor productivity, wage growth has lagged behind productivity gains, resulting in a diminished laborers’ income share (Yu et al. 2019). In short, beyond causing unemployment, technological substitution can decrease wages for those who remain employed (Cheng et al. 2018; Acemoglu and Restrepo 2019; Frank et al. 2019).

Moreover, the technological substitution effect directly contributes to a significant reduction in aggregate work hours. The integration of automation technologies often boosts productivity, which, theoretically, could enable a reduction in labor input and thus shorter work hours (Autor and Salomons 2018; Acemoglu and Restrepo 2022). Quantitative research based on data from 17 countries has revealed that for low-skilled laborers, robot adoption rates influence work hours; specifically, higher robot adoption rates are correlated with fewer work hours (Graetz and Michaels 2018). Furthermore, field observations by Chinese scholars have revealed that while the deployment of automation technology can alleviate workloads, shorten work hours, and reduce work intensity, these changes can paradoxically lead to reduced overtime, diminished income, and even job displacement (Zhang and Zhang 2019).

In contrast, scholars emphasizing technological control focus more on the labor process. Braverman’s (1978: 103–112) influential synthesis posits that technological control over the labor process unfolds through three principal dimensions. First, a key manifestation of this labor control lies in separating worker skills from the labor process. As automation increasingly displaces human tasks, the labor process’s reliance on workers’ core competencies diminishes, intensifying deskilling (Qiu et al. 2019; Xu and Ye 2020). Second, labor control is exemplified by the separation of conception from execution. Algorithms increasingly supplant human judgment in worker task allocation (Chen 2020), while automated equipment concurrently reinforces the established assembly line’s regulation of work pace (Xu and Ye 2020). Third, managers solidify their control over the labor process through the monopolization of knowledge. Automation and digital technologies enhance managers’ capacity for direct control (Xu and Ye 2020). Concurrently, automation-induced deskilling simplifies tasks, reduces wages, and diminishes worker autonomy; this resultant curtailment of flexibility, in turn, indirectly leads to extended work hours (Cai and Shi 2016; Xu and Ye 2020).

In the digital era, this logic of control has evolved into a more subtle and pervasive form driven by data and algorithms. The seamless integration of automated equipment with production management software enables real-time monitoring of efficiency and the optimization of labor rhythms. Thus, the labor process is no longer merely governed by the physical assembly line but by an invisible, algorithmically determined set of performance targets. This “algorithmic pacing” further compresses worker autonomy, compelling employees to adapt to the “optimal” operational efficiency of machines or systems, which often necessitates extending labor time or intensifying effort to meet system demands. From this perspective, automation does not usher in an era of leisure, but instead tethers workers to the logic of capital valorization, with the extension of work hours being a direct and measurable consequence.

Despite their distinct theoretical frameworks, both the technological substitution and technological control perspectives broadly converge on the adverse influence of technology on workers’ wages. However, a key distinction emerges in their implications for work hours: the technological substitution framework anticipates a reduction, while the technological control framework suggests an extension of work hours.

Concurrently, the production effect of automation technology mitigates these aforementioned negative impacts. Automation technology can induce capital accumulation, thereby fostering an expansion of labor demand (Acemoglu and Restrepo 2019). It also simultaneously generates novel specialized tasks and employment opportunities through “creative destruction” (Schumpeter 1999: 149), and facilitates the reassignment of labor to alternative roles (Autor 2015). In contexts where labor supply falls short of demand, the inherent limitations of the technological substitution effect become particularly salient (Yu et al. 2019). This is evident in China, where an aging population and persistent labor shortages have led to its technological substitution effect being characterized as “complementary substitution” (Chen et al. 2018; Song and Zuo 2019). This further suggests that, under conditions of robust labor demand, workers are prone to simultaneously undergo both “deskilling” and “reskilling” processes (Qiu et al. 2019; Xu and Chen 2020).

The dual nature of technological progress and its differentiated impacts across skill groups give rise to the theoretical dialogue underpinning this research. Technology, as a tool of production, can theoretically reduce labor intensity and shorten work hours by replacing repetitive tasks. As a tool of management, it can also be used to intensify control over the labor process, leading to alienation and longer work hours. How has the interplay between these two forces—technological substitution and technological control—unfolded in the era of automation and AI?

Skill heterogeneity

in technological exposure

The intersection of workers’ job tasks and tasks amenable to technological automation has been termed “technological exposure” (Webb 2019). Workers possessing distinct skill sets confront differential risks of technological exposure. Across the literature, laborers are frequently dichotomized into binary skill categories, including, for instance, skilled versus unskilled, highly versus less educated, or cognitive-skilled versus non-cognitive-skilled workers (Liu and Grusky 2013). According to the requisite worker skills, job tasks can be differentiated into four categories: routine manual, routine cognitive, non-routine manual, and non-routine cognitive (Autor et al. 2003). Within the contemporary literature on technological change and the labor market, a prevalent analytical framework, exemplified by Acemoglu and Restrepo (2018a), posits a fundamental division between high-skilled and low-skilled occupations. This distinction is predominantly drawn on the inherent characteristics of tasks: high-skilled roles are typically understood to involve non-routine cognitive functions, demanding advanced capacities for critical judgment, sophisticated problem-solving, and complex analytical synthesis. Conversely, low-skilled roles are defined mainly by their routine, standardized, and comparatively less cognitively demanding nature, often entailing repetitive physical or algorithmic execution.

Building on conceptualizations of occupational skill differentiation, such as those discussed by Webb (2019), the labor market exhibits clear patterns of task allocation. Occupations predominantly composed of routine, manual, or codified tasks—such as factory and warehouse handlers, installers, and food processing personnel—are often associated with roles requiring minimal formal credentials or complex problem-solving skills. Conversely, occupations characterized by non-routine cognitive functions, demanding advanced analytical capacities and specialized professional expertise—including clinical laboratory technologists, chemical engineers, and optometrists—represent the higher end of the skill spectrum. As AI continues to reshape the division of labor, the analytical salience of cognitive skills has intensified, serving as a critical lens through which to understand the restratification of occupations and the evolving demands placed on human capital.

Within analyses of technological change and labor markets, the composition and specificity of skills emerge as a critical mediator of divergent labor market outcomes stemming from technological advancement. Contrary to earlier predictions regarding automation’s primary impact on routine, low-skilled labor, a growing body of research is now suggesting that advancements in AI are progressively exposing even highly skilled occupations to substantial technological risks. This emerging perspective highlights a robust substitution effect of AI and automation on complex cognitive tasks previously considered exclusive to advanced human capital (Webb 2019; Acemoglu and Restrepo 2018a). Counter to some projections of broad AI-driven transformation, analyses of China’s industrial development reveal that the diffusion of highly autonomous production systems within manufacturing sectors remains constrained.

Furthermore, the deployed automation technology largely operates below the threshold of generalized AI, primarily executing pre-programmed, task-specific functions. Consequently, this empirical reality suggests that the proximate risk of technological unemployment or deskilling continues to disproportionately affect occupations characterized by routine, manual, and replicable tasks, rather than those demanding advanced human judgment and non-codified expertise (Xu and Hui 2019). Nevertheless, irrespective of the primary locus of technological displacement, workers in routine-task occupations remain vulnerable to adverse outcomes. This vulnerability stems from cascading displacement, whereby the displacement of individuals from higher-skilled roles increases competition in segments historically dominated by less credentialed labor, thereby contributing to wage stagnation and job precarity (Acemoglu and Restrepo 2018a).

Grounded in the conceptualization of skill differentiation, managerial and non-managerial roles are affected by differential forms of technological control, reflecting the distinct human capital demands across the organizational division of labor. From a managerial perspective, while automation exerts a discernible deskilling effect on routine cognitive tasks, the increasing complexity of work and evolving management philosophies necessitate a transformation in managerial roles from direct oversight to strategic collaboration (Zhang 2023). This evolving managerial imperative demands heightened social and non-codifiable skills, insulating such roles from the direct substitutional effects of automation. Yet, even as managerial roles transform, the very design of modern technological systems, embodying technical control, paradoxically concentrates and intensifies surveillance and coordination capacities (Edwards 1979: 120). This process disproportionately consolidates power within a limited cadre of managers. The select managerial cadre, often empowered by technical control and algorithmic oversight, skillfully mobilizes pervasive disciplinary micro-powers to shape laboring bodies into conformity with automated processes and to extract greater surplus value through work intensification (Xu and Ye 2020). Acclimating to and navigating novel technologies often demands extended work durations for employees (Zhang 2007). Viewed through this lens, the deployment of automation technology may alleviate managerial burdens, yet it rarely diminishes the actual work hours for the managed workforce.

The impact of technology on both income and work hours exhibits skill-biased heterogeneity. So long as automation technology has not achieved sophisticated autonomous decision-making and self-governance, workers whose core competencies reside in complex cognitive skills are unlikely to face direct substitution, and the degree of technological control exerted over them will remain substantially constrained. Consequently, a second central inquiry of this study addresses the heterogeneous impact of automation technology across the workforce, disaggregated by skill. This overarching question is refined by two sub-questions: To what extent do cognitive skills account for variation in technological outcomes? And, what are the distinct effects of automation technology on occupations requiring advanced skills versus those requiring routine skills?

Generative AI: unpacking

emerging sociological possibilities

With the advancements of AI technology, human society is transitioning from physical automation to cognitive automation. The academic debate concerning AI’s potential for human cognitive labor substitution has deep historical roots. For example, a frequently cited projection posits that as much as 47% of occupations in the US may be susceptible to automation in the near term (Frey and Osborne 2017). Conversely, other scholars, drawing on occupational or industry-level analyses, contend that AI primarily replaces humans in specific tasks, thus far exhibiting no discernible aggregate impact on the broader labor market (Acemoglu et al. 2022). For AI automation technologies to genuinely replace human cognitive labor, they must advance from narrow AI to general AI.

The emergence of LLMs, exemplified by OpenAI’s GPT-4, has transformed the development landscape of AI automation. The generative artificial intelligence (GAI) capabilities demonstrated by LLMs enable, to a discernible extent, knowledge production and the execution of complex cognitive tasks (Yang et al. 2024). This even extends to participation in scholarly endeavors across both the natural and social sciences (Zheng and Lv 2023). One might contend that LLMs demonstrate a closer proximity to GAI. Nevertheless, LLMs currently fall short of the fundamental criteria for GAI, thus remaining within the classification of “weak AI” (Qiu 2023). Nonetheless, the application domains and scenarios for AI technologies are poised for significant expansion, and this transition from traditional automation to GAI portends a non-trivial impact on the labor market.

Scholars have aligned human capabilities with AI-executable tasks to calculate the “AI occupational exposure” (AIOE) for different occupations (Felten et al. 2021), and subsequently have further calculated the exposure rates of these occupations to LLMs in AI (Felten et al. 2023). They have found that telemarketers and various higher education teachers, such as those of English language and literature, foreign language literature, and history, are the occupations with the highest exposure rates. Meanwhile, the industries with the highest exposure rates are legal services, securities, and commodity investment. Based on this theoretical framework, the OpenAI team utilized the O*NET occupational information database to compute LLM technological exposure rates across various occupations. By integrating manual annotation with GPT-4 classification, their analysis revealed that AI could impact approximately 80% of the US workforce, affecting at least 10% of their work tasks, while around 19% of the workforce might experience more severe disruption, involving at least 50% of their work tasks (Eloundou et al. 2023). These studies indicate that occupations with higher socioeconomic status tend to have higher exposure rates to AI technology. China’s occupational structure diverges significantly from that of the US, and this raises a critical question: What distinct implications do varying levels of AIOE hold within China’s occupational structure? Given that the actual impact outcomes are not yet empirically discernible, I adopt an exploratory approach to address this study’s third research question: What potential impacts might GAI exert on the workforce within China’s occupational structure?

Research design

Data sources

To address the initial two questions requiring confirmatory analysis, I use data from the 2018 China Labor-force Dynamics Survey (CLDS2018) and integrate it with the O*NET occupational information database. For the third question, which necessitates exploratory analysis, I match data from the 2021 Chinese Social Survey (CSS2021) with the AIOE index (Felten et al. 2023).

The CLDS2018, administered by the Center for Social Science Survey at Sun Yat-sen University, has incorporated several questions pertaining to the application of automation technology. This survey encompassed all laborers aged 15 to 64 residing within the sampled households. After excluding any observations with missing values for key variables, a final sample of 2,799 cases remained for inclusion in the study. The O*NET occupational information database is a comprehensive occupational information system developed by the US Department of Labor’s Employment and Training Administration. GAI technology had yet to be widely or formally adopted in 2021. Given that the empirical effects of GAI on the labor market were not yet widely observable in 2021, I adopt an exploratory approach for this study’s third research question. Using this approach, I examine the potential susceptibility of China’s workforce to GAI based on occupational exposure frameworks, rather than empirically observed labor market outcomes. This acknowledges that the nascent stage of GAI’s integration is a foundational step for future empirical investigations as more longitudinal data become available. The CSS2021, conducted by the Institute of Sociology at the Chinese Academy of Social Sciences, provides research findings generalizable to the national household population aged 18 to 69. Its data represent the latest publicly available large-scale survey data, offering the most current insight. After excluding any samples with relevant missing values, 4,149 samples remained for inclusion in the study.

The O*NET occupational information network database is an occupational information system developed by the US Department of Labor’s Employment and Training Administration. Drawing on this database, I quantified occupational cognitive skill levels and computed the AIOE. Recognizing the inherent limitations of directly applying a US-centric O*NET to the Chinese context, which may exhibit distinct occupational structures and skill definitions, I proceeded with careful consideration. Without a comparable comprehensive occupational skills database specifically for China, and given the fundamental commonalities of general cognitive abilities across industrialized economies, I use O*NET data, employing cross-referencing through occupational codes to improve consistency. I acknowledge this methodological constraint, and highlight the need for future research using context-specific Chinese occupational data to validate these findings.

Variable description

The dependent variables in this study are hourly wage and work hours. An hourly wage is defined as the ratio of pre-tax annual income to annual work hours. To accommodate model assumptions, the hourly wage variable was subsequently logarithmically transformed. Work hours were operationalized as weekly work hours, measured in hours. Consistent with labor process theory and emerging scholarship on algorithmic management, longer weekly work hours are examined as a potential manifestation of intensified technological control. At the same time, work duration can also be influenced by a multiplicity of other organizational and individual factors.

The independent variable is whether automation technology is used in the laborer’s workplace. The specific question in the questionnaire corresponding to this is: “Is your enterprise/unit currently using highly automated technologies, robots, or artificial intelligence (e.g., autonomous driving, machine translation, industrial robots)?”.

The control variables are gender (male?=?1), age, age squared, years of education, self-rated health score (ranging from 1 to 5), residential location (urban?=?1), regional location (East, Central, or West), hierarchical position (enterprise/unit head?=?4, middle management?=?3, general cadre?=?2, general staff?=?1), whether the job requires training, possession of certificates, job-required experience score, socioeconomic status (SES) score, type of enterprise/unit (within the system?=?1), size of the enterprise/unit (number of employees after logarithmic transformation), industry type, and sector type. The descriptive statistics for each variable are presented in Table 1.

To identify the heterogeneous effects of technology on high- and low-skilled workers, I operationalize skill level by differentiating groups based on cognitive abilities. The differential effects of technology on high- and low-skilled laborers are shaped by cognitive skill levels. Traditional studies often differentiate high- and low-skilled laborers based on university enrollment, a method that may confound educational attainment with underlying skills. To achieve greater precision, I quantify cognitive skills using a cognitive ability score based on indicators from the O*NET occupational information database. As conceptualized within the O*NET occupational information database, cognitive ability refers to the capacity for knowledge acquisition and application in problem-solving. This construct covers diverse abilities such as attention, creativity and reasoning, memory, perception, quantitative reasoning, spatial manipulation, and verbal comprehension. The O*NET occupational information database provides scores for 873 occupations on various abilities, systematically coded under the Standard Occupational Classification (SOC) system. A cognitive ability score for each occupation was constructed by summing the scores across six distinct cognitive abilities, with the exception of spatial manipulation. Occupational socioeconomic status (SES) was determined using the International Socioeconomic Index of Occupational Status (ISEI).

Analytical methods

Validating the impact of automation technology: multivariate linear regression and propensity score matching

In this study, I employed multivariate linear regression to investigate the impact of automation technology on workers’ hourly wages and weekly work hours. Hourly wage and weekly work hours are the dependent variables, with exposure to new technology as the key independent variable. A robust model is constructed after incorporating all relevant control variables. The baseline model is specified as follows:

The inherent differences in labor environments (e.g., industry, sector, firm type, firm size) between technology-exposed and unexposed groups introduce a significant risk of selection bias. Consequently, observed technological impacts might reflect these underlying environmental variations rather than the technology itself. I address this challenge by utilizing propensity score matching. Propensity scores for introducing new technologies are estimated using relevant labor environment variables, including organizational scale, organizational type, industry type, sector type, regional location, and geographic area. Nearest-neighbor matching with replacement is then performed based on these scores. Subsequently, the net effect of technology’s impact is estimated, and bootstrapping is employed for significance testing.

The Importance of cognitive skills: a causal random forest model

To explore the extent to which cognitive skills moderate the treatment effect of automation technology, I employed the causal random forest (CRF) method for data-driven exploration. As a non-parametric machine learning approach, CRF addresses a key limitation of traditional linear models by capturing heterogeneous causal effects across different subgroups. Based on the framework of random forest, CRF recursively partitions data to identify distinct subgroups with similar treatment effects. The unique structural advantage enables the estimation of the treatment variable’s heterogeneous causal effects across diverse subgroups (Jia 2022). This data-driven process is crucial in social science research where complex, non-linear relationships and high-dimensional data are common, as it enables the estimation of heterogeneous effects without relying on pre-specified functional forms or arbitrary assumptions about model specification.

Furthermore, these models provide a significant methodological advantage by measuring feature importance. This capability offers insights into which control variables are most influential in estimating causal relationships, thereby enhancing the model's reliability. As Hu et al. (2021) detail, this method can mitigate some of the influence of arbitrary human assumptions on model specification. By allowing us to construct a more reliable counterfactual, this approach enables us to more accurately isolate the true causal impact of automation on individuals with varying cognitive skills.

Differentiating high- and low-skilled workers: an unsupervised clustering algorithm

To objectively differentiate workers based on their cognitive skills and prevent researcher subjectivity, I employed an unsupervised clustering algorithm for group classification. Clustering algorithms learn the inherent distribution structure of data to identify any underlying properties or patterns, partitioning samples into disjoint subsets. These subsets, known as “clusters,” exhibit high within-cluster similarity and low between-cluster similarity (Liang and Jia 2022). In this study, I applied an unsupervised clustering algorithm to cognitive ability data across six dimensions to distinguish between two worker typologies. The most robust clustering solution derived from this process is adopted as the definitional basis for these worker groups. These six dimensions are: attention capacity, creativity and reasoning ability, memory capacity, perceptual ability, quantitative ability, and verbal ability.

Automation technology’s impact

on workers

Automation technology’s impact on worker wages depends on the interplay between its substitution and productivity effects. When the productivity gains are insufficient to compensate for the displacement arising from automation, we anticipate a negative effect on worker wages. Conversely, if the productivity effect offsets the substitution effect, automation may exert a neutral or beneficial effect on wages. These theoretical propositions underlie competing Hypotheses 1.1 and 1.2.

Hypothesis 1.1: If the substitution effect predominates, the introduction of automation technology will decrease workers’ wages.

Hypothesis 1.2: If the productivity effect predominates, the introduction of automation technology will have no negative effect on workers’ wages, or even a positive effect.

The operationalized measure for wages is the hourly wage. This is because using total annual income as a wage measure would have obscured the distinct effects on work hours and hourly wages. For instance, an increase in a worker’s total annual income could be the result of either extended work hours or a rise in the hourly rate, demonstrating that changes in these two components might be misaligned. From the perspective of technology substitution theory, the introduction of automation technology reduces workers’ hours. Conversely, technology control theory suggests that automation technology would extend work hours. This theoretical divergence underpins competing Hypotheses 2.1 and 2.2 in this study.

Hypothesis 2.1: If the technology substitution theory holds, the introduction of automation technology will reduce work hours.

Hypothesis 2.2: If the technology control theory holds, the introduction of automation technology will extend work hours.

The multivariate linear regression models constructed based on the above hypotheses are presented in Table 2. Among these, Models 1 and 3 serve as baseline models. The analysis reveals that the introduction of technology significantly reduced hourly wages by 0.15 units (Model 2) and significantly extended weekly work hours by 3.86 units (Model 4). However, propensity score matching results reveal a more nuanced picture. Once the propensity for technology adoption was controlled for—yielding substantial balance between the control (without automation technology) and experimental (with automation technology) groups across key labor environment attributes, including organizational scale, organizational type, industry type, sector type, regional location, and geographic area—there was no significant difference between these two matched groups. The results also indicate no significant direct difference in hourly wages attributable to the introduction of technology after controlling for observable variables. This suggests that the observed wage reduction might have been more related to workers’ pre-existing labor environment characteristics. When workers operate in comparable labor environments (e.g., the same industry or sector), whether their respective firms have adopted automation technology might not significantly differentiate their wages. This implies that automation technology’s negative effect on hourly wages primarily manifests at macro levels, such as industry and sector. At the micro level, automation technology’s negative effect on hourly wages appears to be circumscribed within specific industries or sectors. This does not imply that technology has no effect, but rather that its negative substitution effect on wages may be counteracted by other forces (such as productivity effects or bargaining power from labor shortages) in the current Chinese labor market. In contrast, automation technology’s effect on work hours is unrelated to the labor environment’s propensity for technology adoption; even within environments sharing the same propensity for technology adoption, automation technology’s extending effect on work hours remains significant. These findings are consistent with Hypotheses 1.2 and 2.2. They demonstrate the limitations of the technology substitution effect on wages, as well as the robust explanatory power of labor process theory (or technology control theory) regarding changes in work hours under technological influence.

Skill heterogeneity

in automation technology’s effect

on workers

Cognitive skills represent a crucial dimension for distinguishing high-skilled from low-skilled workers. This section explores cognitive skills’ role in mediating the effect of technology and the heterogeneous effects of automation technology on high-skilled and low-skilled workers.

Cognitive skills’ role

in mediating technological impact

To investigate the importance of cognitive skills in mediating technological impact, I employ the causal random forest (CRF) model (Athey et al. 2019) to estimate variable feature importance. Following the parameter tuning approach described by Hu Anning et al. (2021), I utilized bootstrapping with replacement sampling, generating a new dataset of 100,000 samples based on 2,799 samples from the CLDS2018 data. Building on this, four CRF models were established: 2,000 trees (using the honest algorithm), 1,500 trees (using the honest algorithm), 1,000 trees (using the honest algorithm), and 1,000 trees (not using the honest algorithm). Within each CRF model, the training sample proportions are set to 0.6, 0.5, and 0.2, respectively, generating a total of 12 distinct models. Each model estimates the conditional average treatment effect (CATE) of new technology on individual hourly wages and weekly work hours. The correlation in the estimation results across models indicated that the model with a training sample proportion of 0.2, parameters set to 1,500 trees, and employing the honest estimation algorithm, exhibited the best generalization performance and the most robust estimation results. Therefore, I adopted this model to estimate the CATE for the samples.

Figure 1 displays the feature importance for each variable, where a higher degree of importance signifies a greater role in explaining the technological effect’s heterogeneity. The ranking of feature importance shows that for individual hourly wages, the top three influencing factors for the technology treatment effect’s heterogeneity are organizational scale, self-rated health, and cognitive ability score. For individual weekly work hours, the top three factors influencing the technology treatment effect’s heterogeneity are SES score, cognitive ability score, and age. Our analysis also reveals that cognitive ability scores play a significant role in the technology treatment effect’s heterogeneity across hourly wages and weekly work hours. Other variables related to occupational ability, such as years of education, required work experience, bureaucratic position, whether training was received, and whether a certificate was obtained, all exhibit low degrees of feature importance. This demonstrates that the occupational cognitive ability score has significant explanatory power for the technological effect’s heterogeneity. Moreover, it indicates that the cognitive ability scores obtained through matching with the O*NET occupational information database have important analytical significance for distinguishing between high-skilled and low-skilled workers.

So, what role do cognitive skills play in mediating technological impact? As shown in Figure 2, examining individual CATE trends as cognitive ability scores change reveals that as cognitive ability scores increase, the technological treatment effect on hourly wages shifts from negative to positive, with a significant Pearson correlation coefficient of 0.25. This indicates that the technological substitution effect workers experience diminishes as their cognitive ability scores rise. For the group with the highest cognitive ability scores, the technological treatment effect on hourly wages is even positive, suggesting that the application of automation technology yields them positive returns. Simultaneously, the technological treatment effect on weekly work hours gradually approaches zero, with a significant Pearson correlation coefficient of ?0.13, indicating that the technological control workers experience also diminishes as cognitive ability scores increase. Overall, as cognitive ability scores improve, workers increasingly benefit from technology.

Heterogeneity in

the technological impact

under cognitive skill differentiation

The preceding analysis indicates that cognitive skills, as measured by cognitive ability scores, serve as an effective indicator for distinguishing between high-skilled and low-skilled workers. Therefore, these two types of workers are classified based on cognitive ability scores using a clustering algorithm and various percentage-based division methods. Considering the performance of various clustering models on the silhouette coefficient, CH score (Calinski-Harabasz score), and DBI index (Davies-Bouldin index), I selected the K-means clustering algorithm to group the workers into two categories. The K-means clustering algorithm partitions high-skilled and low-skilled workers into two disjoint groups across all dimensions of cognitive ability, with high-skilled workers consistently scoring significantly higher than low-skilled workers across all metrics.

To mitigate the influence of different classification methods for high-skilled and low-skilled workers on the study’s conclusions, high-skilled and low-skilled workers are also classified based on cognitive ability score rankings using three different cutoff points: the top 30% (Classification Method 1), the top 40% (Classification Method 2), and the top 50% (Classification Method 3). Including the K-means clustering, a total of four classification methods are employed, with separate regression models established for each to estimate the technological effect.

As shown in Figure 3, low-skilled workers are concentrated in Industries 1–8. These include agriculture, forestry, animal husbandry, and fishery; mining; manufacturing; electricity, gas, and water production and supply; and construction. In contrast, high-skilled workers are concentrated in Industries 9–15. These include finance and insurance; real estate; health, sports, and social welfare; and scientific research and comprehensive technical services.

Skill heterogeneity is inevitable in technology exposure. If AI automation technology does not truly replace cognitive labor, but instead, as observed by Chinese scholars (Xu and Hui 2019; Xu and Ye 2020), it disproportionately replaces low-skilled labor, then low-skilled workers will experience stronger technological replacement and control. Meanwhile, high-skilled workers will be less affected. This leads to the following hypotheses.

Hypothesis 3.1: Compared to low-skilled workers, high-skilled workers’ hourly wages will be less affected by technology, or even unaffected.

Hypothesis 3.2: Compared to low-skilled workers, high-skilled workers’ work hours will be less affected by technology, or even unaffected.

The results of the multivariate linear regression models constructed based on the above hypotheses are presented in Table 3. Regardless of the classification method, the technological effect on high-skilled workers’ hourly wages was not statistically significant. In contrast, low-skilled workers’ hourly wages were consistently and negatively affected by automation technology, decreasing by at least 0.17 units. This robustly supports Hypothesis 3.1, indicating that automation technology does not significantly impact high-skilled workers’ hourly wages, while low-skilled workers remain the primary group experiencing its substitution effect.

Simultaneously, except for Classification Method 2, the technological effect on high-skilled workers’ weekly work hours is not significant across any other classification methods, while low-skilled workers’ weekly work hours are consistently significantly affected, being extended by at least 4.48 h. This finding initially supports Hypothesis 3.2, indicating that automation technology similarly does not significantly affect high-skilled workers’ weekly work hours, while low-skilled workers remain the most impacted group.

However, consistent with the overall observation, propensity score matching reveals no significant difference in hourly wages between the two groups of low-skilled workers (those exposed to automation technology and those who are not). This finding suggests that the technological substitution effect low-skilled workers experience is also circumscribed. This outcome could stem from two potential mechanisms: (1) the counteracting productivity effect of automation technology, implying that low-skilled workers may have adapted to “new tasks” through “reskilling”; (2) even if situated in positions with a higher degree of “deskilling”, low-skilled workers might still gain greater bargaining power due to a tight labor supply.

According to the “Most In-Demand Occupations”Footnote6 ranking released by China’s Ministry of Human Resources and Social Security, frontline employees and some specialized technical personnel in labor-intensive industries (including salespersons, restaurant servers, cashiers, packers, and automobile production line operators) are in short supply. In this context, even if these in-demand occupations have high technology exposure rates, the technological substitution effect remains insignificant. The application of automation technology in China is driven by a structural environment characterized by a labor supply shortage, suggesting that the productivity effect should be more substantial. Considering the industry distribution of low-skilled workers, market demand is robust for them, and the substitution effect of technology is limited.

Concurrently, the introduction of automation technology indeed leads to an extension of work hours for low-skilled workers. This aligns with labor process theory, which posits that automation intensifies capital control over the production process. Productivity is perpetually insufficient for capital expansion, despite its rapid increase (Braverman 1978: 185). Low-skilled workers are compelled to accommodate machine demands to ensure the continuous operation and maximized efficiency of automated equipment, and their diminished autonomy forces them to extend their work hours. Moreover, automation boosts production efficiency, potentially prompting enterprises to expand their scale. The resulting increase in work tasks and intensity contributes to longer work hours for low-skilled workers. Furthermore, the introduction of automation technology intensifies competition among workers, leading to lower expectations of job stability among low-skilled workers. To retain their jobs or increase their incomes, low-skilled workers are compelled to engage in competitive overtime (Yang and Qiu 2020).

This seemingly contradictory finding—a muted wage substitution effect alongside a potent work-hour control effect—reveals the complex negotiation among capital, technology, and labor within China’s specific labor market context. In a macro-environment characterized by labor shortages, capital’s ability to suppress wages via technology is constrained. Even after introducing robots, firms must still compete for the human workers necessary to operate, supervise, and maintain the robotic systems. This grants labor a degree of bargaining power that counteracts technology’s wage-depressing potential. However, where capital’s power is constrained in a wage setting, it becomes amplified in the control over the production process. For the firm, substantial investment in automation creates a powerful incentive to maximize its utilization to ensure a swift return. The “downtime” of expensive machinery becomes a critical cost to be minimized, which, in turn, compels firms to extend human work hours to maintain continuous system operation. Therefore, the impact of technology is not a simple act of substitution, but rather a negotiated outcome: workers may retain their nominal wages amidst labor market tightness, yet they pay a price in the form of longer hours and diminished autonomy, reflecting capital’s reassertion of control within the factory walls.

Synthesizing the above discussions leads to the conclusion that as cognitive ability scores increase, the negative impact of automation technology on workers gradually diminishes, while its positive effect progressively strengthens. For high-skilled workers, automation technology has no significant impact; for low-skilled workers, automation technology extends work hours, but does not significantly affect hourly wages. Thus, under the counteracting influence of productivity effects, the substitution effect of automation technology appears circumscribed, yet its control effect remains significant. Furthermore, this control effect of automation technology exhibits skill heterogeneity, centered on cognitive skills.

Exploring the impact of GAI

To explore the potential impact of cognitive automation, particularly its newest form, GAI, on workers, I first compared industry differences in the impact of automation technology versus GAI, and then investigated the specific populations affected by GAI. Given the current absence of observable social consequences stemming from GAI’s impact on workers in China, I adopted an exploratory approach to address this emerging question.

Given the inherent commonalities in occupational skill requirements across countries and the high correlation between technology exposure and skills, the technology exposure rate for a specific occupation in the US is likely similar to that in China. However, it is important to acknowledge that the occupational structures and demographic compositions of the workforce may differ between the two countries. For instance, a given occupation could be predominantly held by young women in one country, while primarily by middle-aged men in another. By matching the estimated results of AI occupational exposure (AIOE) (Felten et al. 2023) with the CSS2021 sample using occupational codes, I have derived the distribution of AI LLM exposure rates in China. Given the inherent timeliness of the occupational structure captured by the CSS2021, I have used the occupational structure measured by CSS2021 to represent it when GAI begins to exert its influence.

Comparison of

automation technology

and GAI impacts

Analysis of the CLDS2018 survey data reveals that the secondary sector, particularly manufacturing, exhibits a stronger propensity for automation technology adoption. As Figure 4 illustrates, manufacturing (Industry 3), electricity, gas, and water production and supply (Industry 4), and scientific research and comprehensive technical services (Industry 14) all show higher automation technology adoption propensity values than the overall average (indicated by the dashed line in Figure 4).

Figure 5 presents the industry distribution of AI LLM exposure rates based on CSS2021 data. Among these, wholesale and retail trade, food services (Industry 8), finance and insurance (Industry 9), real estate (Industry 10), social services (Industry 11), education, culture, arts, radio, film, and television (Industry 13), scientific research and comprehensive technical services (Industry 14), and state organs, party and government agencies, and social organizations (Industry 15) all exhibit AI LLM exposure rates exceeding the overall average (indicated by the dashed line in Figure 5).

The population affected by GAI

Figure 6 illustrates the distribution of AI LLM exposure rates across various demographic and occupational strata, including gender, age, professional skill requirements, and socioeconomic status. A higher AIOE score indicates a greater likelihood that these social groups are affected by AI LLMs, though it does not specify the valence of these effects (i.e., positive or negative). Compared to men, women have a higher AI LLM exposure rate. As age increases, the AI LLM exposure rate decreases. Workers whose occupations require professional skills have a higher exposure rate compared to those whose occupations do not require professional skills. Overall, groups with higher socioeconomic status have higher AI LLM exposure rates, with the group scoring between 60 and 70 on socioeconomic status having the highest technology exposure rate.

My exploratory analysis reveals that occupational exposure to GAI in China exhibits clear demographic patterns, primarily affecting women, younger workers, professionals, and those with higher socioeconomic status. Interpreting this “exposure” is crucial. A direct inference is that the cognitive skills held by these groups (e.g., writing, information synthesis, coding) are at risk of replacement by AI, potentially leading to a “deskilling” or devaluation of their labor market value. However, an equally important and competing theoretical possibility is “human-AI complementarity.” In high-exposure occupations, GAI can function as an “enabling tool,” freeing professionals from repetitive cognitive tasks to focus on more creative, strategic, and critical work. In this scenario, professionals who can master and direct AI systems could see their productivity and skill value amplified as never before. Therefore, “high exposure” is a neutral concept indicating profound occupational reshaping. However, whether it leads to skill “devaluation” or “appreciation” depends on institutional contexts, organizational strategies, and workers’ adaptive learning. As such, it is an empirical question that warrants future research.

Through this exploration of the impact of GAI, it can be observed that GAI’s influence on workers is concentrated in the tertiary sector, and predominantly affects women, younger demographics, workers with professional skills, and groups with higher socioeconomic status. Does this then imply that GAI will replace these groups? Qiu (2023) argues that GAI, exemplified by ChatGPT, will not completely replace any human occupation, but will undertake portions of human work instead. Concurrently, unlike traditional automation technologies, the groups affected by GAI are professional skill groups with higher socioeconomic status. The occupations in which these workers engage typically have entry barriers or “elite barriers” (Liu and Grusky 2013), and these institutional occupational thresholds curtail the substitution effect of technology. Therefore, despite high-skilled workers with higher socioeconomic status being the primary targets of cognitive automation technology, the negative impact they experience may also be circumscribed.

Conclusion and discussion

Human society is undergoing a profound transformation from physical automation to cognitive automation, with cognitive skills emerging as crucial for workers hoping to navigate this change. Workers possessing cognitive skills have historically withstood the impact of physical automation, but at the crossroads of the burgeoning development of GAI, they face considerable uncertainty. At this juncture, this study explicates this transformative process through a confirmatory analysis of the impact of existing automation technologies and an exploratory analysis of the impact of GAI.

The analysis in this study yields three primary findings. First, the substitution effect of technology is circumscribed, as evidenced by hourly wages not being significantly affected by the introduction of automation technology. However, the control effect of technology remains significant, leading to extended work hours for laborers. These findings suggest that the substitution effect of technology is contingent, while its control effect is structural. The effect on wages, as a direct outcome of labor market bargaining, is highly contingent on macroeconomic conditions and labor’s bargaining power. In contrast, the drive to intensify control over the labor process to maximize efficiency is a structural imperative of capital, making the effect on work hours more robust and pervasive.

Second, this technological control effect exhibits skill heterogeneity centered on cognitive skills, where high-skilled workers are not significantly controlled by technology, whereas low-skilled workers experience stronger technological control. Mastering cognitive skills mitigates the negative aspects of technological impact, while cognitive skills function as a “technological buffer zone” in the era of physical automation. Their value lies not only in boosting productivity, but also in providing a defense against technological control. The non-routine and complex nature of high-skilled work makes it difficult to standardize and control via existing automation, thereby preserving worker autonomy. This highlights that a key dimension of skill’s value is its capacity to maintain human subjectivity against the incursions of managerial control.

Third, an exploratory analysis of GAI’s impact reveals that its influence on workers is concentrated in the tertiary sector, predominantly affecting women, younger demographics, workers with professional skills, and groups with higher socioeconomic status. The rise of GAI heralds a potential shift in the logic of labor stratification, moving from a division between manual and mental labor to one between routine cognitive and creative/critical cognitive labor. As such, many tasks within traditional high-skilled professions (e.g., drafting legal documents, writing code, creating marketing content) are becoming susceptible to automation.

How should we interpret these technological impacts? While theories of technological substitution and technological control often hypothesize a uniformly negative impact of technology, technological change is a long-term, multi-layered process, the effects of which are not entirely detrimental. Using national-level panel data, Graetz and Michaels (2018) find that industrial robot adoption promotes productivity and increases hourly wages, suggesting a long-term positive effect of automation technology on wages. Likewise, through long-term data research at the industry level, Klenert et al. (2023) have found that the use of robots has not reduced employment, and even has a positive effect on overall employment. Furthermore, utilizing micro-level data from the French manufacturing industry, Aghion et al. (2020) demonstrate a positive effect of automation technology on employment.

Qiu and Qiao (2021) argue that technological change has an inherent screening function, whereby only individuals capable of crossing the threshold of technological transformation will ultimately benefit. A new round of technological revolution has begun, with GAI representing a distinct trend. From the perspective o...