Artificial intelligence and within-firm wage gaps: Evidence from China’s manufacturing sector

Yanhui YIN; Meiying CHEN; Shasha FANG; Yanhui Yin; Meiying Chen; Shasha Fang

doi:10.16995/ilr.23835

Responsibility for opinions expressed in signed articles rests solely with their authors, and publication does not constitute an endorsement by the ILO.

This article is also available in French, in Revue internationale du Travail 165 (3), and Spanish, in Revista Internacional del Trabajo 145 (3).

1. Introduction

The influence of artificial intelligence (AI) on wage inequality has become a central question in labour economics and organizational studies. While some researchers argue that AI adoption widens wage disparities by disproportionately benefiting high-skilled workers (Autor, Katz and Kearney 2006; Acemoglu and Restrepo 2018), others suggest that productivity gains from AI may improve wage outcomes across the board (Graetz and Michaels 2015; Domini et al. 2022). Another stream of the literature finds little to no consistent effect, emphasizing the role of firm-level heterogeneity and broader institutional settings in shaping the outcomes (Koch, Manuylov and Smolka 2021; Humlum 2022). These divergent findings reflect the complex and context-dependent nature of AI’s impact on wage structures (Acemoglu and Autor 2011; Howcroft and Taylor 2023).

The variability in AI’s distributive effects is particularly evident across different sectors and national contexts. Research shows contrasting results in developed versus developing economies (Goos, Manning and Salomons 2014; Yuan, Sun and Chen 2025) and between industries, such as services (Damioli, van Roy and Vertesy 2021) and manufacturing (Barth et al. 2020). This growing body of evidence points to the need for more granular, sector-specific analysis to unpack the mechanisms through which AI adoption influences wage outcomes.

China presents a particularly compelling case. As one of the world’s largest adopters of AI technologies, China accounted for nearly one third of global industrial robot installations in 2022 (IFR 2023). The country’s manufacturing sector, which has traditionally been labour-intensive, is undergoing a rapid transformation towards more capital-intensive production, driven by rising labour costs, government incentives and global competitiveness pressures (Huang, Ju and Yue 2024a, 2024b). This shift has created fertile ground for AI integration, supported by vast data resources, robust digital infrastructure and strong state-led industrial policy. According to Deloitte China (2020), manufacturing generates more data annually than any other sector, making it a prime candidate for AI-driven productivity enhancements. In its 2020 report, Deloitte China projected that the market for AI in China’s manufacturing would exceed US$ 2 billion by 2025, with annual growth rates exceeding 40 per cent since 2019.

These developments raise important questions about how AI reshapes wage dynamics within firms, particularly through changes in firm productivity, labour demand and organizational structure. This study focuses on the within-firm wage gap, specifically the compensation disparity between top management and ordinary employees, as a key indicator of AI’s distributive impact. This focus reflects growing concerns that AI adoption may reinforce hierarchical wage structures by concentrating rewards among a small subset of strategically positioned workers.

The central research question addressed in this article is: How does AI transformation influence the within-firm wage gap in Chinese manufacturing firms? We focus on the manufacturing sector due to its central role in China’s AI strategy and its ongoing shift toward capital-intensive, digitally integrated production. We explore how AI-induced changes in productivity and labour structure affect wage dispersion and examine how these effects vary across firm types by ownership, capital intensity and regional digital infrastructure.

This study makes several contributions. Empirically, it leverages firm-level textual disclosures to measure not only the degree of AI adoption, but also the strategic intent behind such transformation. Methodologically, it applies text mining techniques to identify AI-related language in corporate reports, offering a more nuanced and scalable measurement of AI exposure than traditional patent- or investment-based metrics.

The remainder of this article is organized as follows. Section 2 reviews the relevant literature and develops the research hypotheses. Section 3 describes the data, variable construction and empirical methodology. Section 4 presents the empirical results, including the baseline findings, endogeneity analysis and robustness checks. Section 5 examines the underlying mechanisms and heterogeneity of AI’s effects on wage disparities. Finally, section 6 concludes and discusses policy implications.

2. Literature review

AI has emerged as one of the most transformative technological advancements in recent decades, functioning as “intelligent agents” that respond dynamically to their environments (Acemoglu and Restrepo 2020). While AI holds the potential to significantly enhance productivity, its broader implications for income distribution and labour market structures, particularly within-firm wage disparity, have sparked growing academic and policy debate.

A substantial body of literature argues that AI adoption exacerbates wage inequality by disproportionately benefiting high-skilled and non-routine workers while displacing low-skilled and routine-task workers. Theoretical frameworks such as Skill-Biased Technological Change (SBTC), Task-Biased Technological Change (TBTC) and Routine-Biased Technological Change (RBTC) offer complementary perspectives on these dynamics.

SBTC posits that technological progress complements high-skilled labour while substituting for low-skilled labour, thereby increasing the relative demand for skilled workers and widening wage disparities (Autor, Katz and Kearney 2006; Autor and Dorn 2013; Goos, Manning and Salomons 2014; Violante 2008). In contrast, TBTC and RBTC shift the analytical focus from skill levels to task characteristics. These frameworks argue that routine tasks – whether cognitive or manual – are more susceptible to automation, while non-routine tasks involving problem-solving, creativity or interpersonal interaction are augmented by technological change (Autor, Levy and Murnane 2003; Acemoglu and Autor 2011; Sharfaei 2024).

Notably, TBTC and RBTC suggest that technological change may benefit both high-skilled and low-skilled workers, while displacing middle-skilled workers concentrated in routine occupations. This mechanism has contributed to the phenomenon of job polarization, whereby employment expands at the top and bottom of the skill distribution but contracts in the middle (Autor, Katz and Kearney 2006; Hunt and Nunn 2019). This framework provides a more nuanced understanding of how technology impacts labour markets by accounting for task characteristics rather than solely relying on skill levels.

These frameworks provide a theoretical basis for understanding the rise in within-firm wage disparity, particularly as AI reshapes compensation hierarchies. Empirical studies support this perspective. For example, Barth et al. (2020) found that robot adoption in Norwegian manufacturing firms disproportionately raised wages for high-skilled and managerial employees. Similarly, Mandelman and Zlate (2022) observed that automation in the United States reduced wages and employment for middle-skilled workers, while increasing earnings for high-skilled personnel. Violante (2008) and Hémous and Olsen (2022) emphasize that AI increases demand for high-skilled labour while substituting low-skilled jobs, amplifying wage disparities.

Historical evidence further illustrates this trend. Gordon (1996) documented the displacement of routine jobs during the computerization boom of the 1980s, and Humlum (2022) showed that AI adoption led to rising wages among technical staff and declining compensation among production workers. These studies suggest that AI alters internal wage structures by increasing the value and pay of roles associated with strategic oversight and technical innovation.

However, the empirical relationship between AI and wage inequality remains contested. Some studies indicate that the redistributive effects of AI are not uniform across the labour market and may be concentrated within specific firm structures. For instance, Koch, Manuylov and Smolka (2021) and Humlum (2022) find limited changes in overall wage inequality across wage percentiles following AI adoption. In the Chinese context, Yuan, Sun and Chen (2025) report that AI adoption, particularly among technology-intensive A-share listed firms, is associated with reduced within-firm wage disparity. Meanwhile, Freeman, Ganguli and Handel (2020) and Ozgul et al. (2024) emphasize that wage gains from AI adoption can vary substantially within occupations, with senior technical roles benefiting more than entry-level ones.

Taken together, these findings highlight that the effects of AI on wage structures are highly context-dependent. This variation underscores the importance of conducting a more granular analysis of within-firm wage disparity. In this regard, the following hypotheses are proposed:

H1-1: AI transformation may widen within-firm wage disparity.

H1-2: AI transformation may narrow within-firm wage disparity.

AI technologies have the potential to substantially enhance firm-level productivity. However, the implications of these productivity gains for within-firm wage disparity remain empirically contested.

A growing body of literature suggests that improvements in firm-level productivity can have positive distributional effects, including wage increases and reductions in inequality. Jayachandran (2006) demonstrates that enhanced productivity can directly raise wages and narrow wage gaps by increasing the marginal value of labour. Similarly, Porter and Heppelmann (2014) highlight how the integration of technologies, such as industrial robots and the Internet of Things, can significantly boost productivity, leading not only to higher profitability, but also to more generous compensation packages for employees. Using panel data from Chinese listed companies between 2014 and 2022, Wu et al. (2024) provide empirical support for this mechanism in the context of AI application, showing that firm-level adoption of AI technologies can enhance productivity and contribute to more equitable wage outcomes. In a broader cross-country analysis, Graetz and Michaels (2015) find that, under certain institutional and market conditions, technological advances can reduce wage disparities by improving firm performance and enabling a more equitable distribution of productivity gains.

Conversely, other studies suggest that the benefits of productivity growth from AI are not evenly distributed within firms. Instead, they may disproportionately accrue to skilled and non-routine workers whose tasks complement AI technologies. Jiang, Wang and Liu (2024), for example, show that in China’s manufacturing sector, productivity improvements have primarily raised the wages of technologically complementary workers, thereby amplifying wage inequality. This suggests that the effect of productivity gains on wage structure may depend heavily on occupational roles and firms’ technological adoption strategies.

Cross-country variation further complicates the relationship between productivity and wage inequality. The Organisation for Economic Co-operation and Development notes that in contexts where AI exposure leads to more homogeneous productivity improvements, particularly among workers in similar job roles, the resulting wage structures may become more compressed, thus reducing inequality (Georgieff 2024). Based on this debate, the following hypotheses are proposed:

H2-1: AI transformation increases productivity and thereby widens the within-firm wage gap.

H2-2: AI transformation increases productivity and thereby narrows the within-firm wage gap.

A related but distinct strand of literature explores the role of AI in reshaping firms’ labour demand structures. Rather than improving equity through wage increases alone, AI may affect inequality by reconfiguring workforce composition. According to Domini et al. (2022), AI adoption often leads firms to hire more high-skilled workers to exploit technological complementarities, thereby changing wage structures without directly affecting existing employees. In contrast, Wu et al. (2024) find that AI deployment can benefit lower-level employees, noting that wage increases among regular staff have occurred through internal restructuring, while executive pay has remained stable – ultimately reducing wage inequality between managerial and non-managerial staff. To reconcile these divergent findings, Afonso, Sequeira and Almeida (2023) adopt a task-based theoretical perspective. They propose that the impact of AI depends on how technological capabilities interact with the task composition of different skill groups. High-skilled workers benefit from complementarities with AI technologies, enhancing their productivity and wages. By contrast, middle-skilled workers, whose tasks are more routine, face higher substitution risk, leading to stagnant or declining wages. Meanwhile, some low-skilled, non-routine workers may retain wage resilience due to task-specific protection. Overall, this framework explains why AI may widen skill-based wage gaps while allowing for heterogeneity in firm-level wage outcomes, depending on internal labour structures.

The nature and outcome of such labour restructuring may also depend on national contexts and institutional settings. For instance, Austria and Germany, which have robust vocational training systems, experience labour market adjustments that help contain wage gaps. In contrast, countries such as Australia and the United States, where labour markets are more service-oriented and reskilling systems are weaker, often face widening disparities as a result of technological adoption (Cords and Prettner 2022; Georgieff 2024). Accordingly, we propose two further hypotheses:

H3-1: AI transformation restructures labour demand, thereby widening the within-firm wage gap.

H3-2: AI transformation restructures labour demand, thereby narrowing the within-firm wage gap.

3. Research design

3.1 Data source

To achieve the objectives of this study, we use data from all A-share manufacturing firms¹ listed on the main boards of the Shanghai and Shenzhen Stock Exchanges from 2012 to 2022. The firm-level basic information and financial data were obtained from the Wind Database, China Stock Market and Accounting Research (CSMAR) Database and Cninfo, the official information disclosure platform for China’s listed companies. Additionally, regional-level economic data were taken from the China Statistical Yearbook and the China City Statistical Yearbook.

In line with standard practices in the literature, data cleaning involved removing firms with substantial data gaps and those marked as ST or *ST due to financial distress.² We also excluded companies with negative total operating revenue, a debt-to-asset ratio outside the range of 0–1, negative employee compensation, a labour income ratio greater than total revenue, and firms with fewer than 100 employees. To further ensure robustness, continuous variables were winsorized at the top and bottom 1 per cent. After data cleaning and matching, the final sample comprised 2,720 listed manufacturing firms in China, yielding a total of 18,414 observations.

3.2 Model specification

To examine the impact of industrial intelligent transformation on internal wage gaps within firms, we construct the following two-way fixed effects model:

$W s p i t = β 0 + β 1 A I i t + γ X i t + λ i + δ t + ε i t$ (1)

Where:

Wsp_it is the dependent variable, representing the wage gap within firm i in year t.
AI_it is the key explanatory variable, measuring the degree of intelligent transformation within the firm.
X_it is a vector of control variables.
i indexes firms, and t indexes years.
λ_i and δ_t denote firm-level and year-level fixed effects, respectively, controlling for time-invariant firm-specific characteristics and common shocks or trends across time.
ε_it is the error term.

This two-way fixed effects model accounts for unobserved heterogeneity across firms and over time by controlling for firm-level fixed characteristics (λ_i) and year-specific macroeconomic or industry-wide effects (δ_t), thus ensuring the robustness of the empirical estimates.

3.3 Variable definitions

3.3.1 Dependent variable: Within-firm wage gap

Following the methodology of Faleye, Reis and Venkateswaran (2013) and Banker, Bu and Mehta (2016), we define the within-firm wage gap as the logarithm of the ratio between the average compensation of the management team and that of regular employees. This metric captures hierarchical wage differentials and serves as a meaningful proxy for within-firm income inequality in the context of technological transformation.

Executives and senior managers – typically engaged in non-routine, strategic and decision-making tasks – are more likely to benefit from AI-enabled productivity enhancements. By contrast, lower-tier employees, often assigned to routine or automatable tasks, may face stagnant wages or even displacement as firms adopt intelligent technologies (Faleye, Reis and Venkateswaran 2013; Banker, Bu and Mehta 2016).

In operational terms, the management team includes all executives, directors (excluding independent directors) and supervisors. Regular employees refer to all non-managerial personnel. The average compensation of the management team is calculated as the total annual compensation paid to directors, supervisors and executives, divided by the size of the management team. The size of the management team is measured by summing the number of directors, executives and supervisors, and subtracting the number of independent directors and any members who did not receive remuneration.

Likewise, the average compensation of regular employees is calculated as the sum of: (i) the change in “payable employee compensation”; and (ii) “cash paid to or on behalf of employees”, minus the total compensation of directors, supervisors and executives. This value is then divided by the number of regular employees.

3.3.2 Core explanatory variable: AI transformation index

Building on the methods of Yu, Wang and Li (2020), we construct the AI transformation index through text mining, using specific keywords to reflect the degree of AI adoption. Following standard keyword association analysis for smart industrial transformation, we identify a set of 27 keywords related to AI adoption in firm reports.³ Using Python’s Jieba segmentation module, we analyse the management discussion and analysis sections of annual reports to extract AI-related language. We then normalize the frequency of these keywords by dividing the total number of occurrences by the length of the management discussion and analysis segment.

3.3.3 Control variables

To avoid endogeneity problems caused by unobserved variables and to obtain more accurate estimation results, we select seven control variables. These are:

Firm age (Age): defined as the number of years since the firm’s establishment, this variable accounts for the potential impact of firm maturity on wage structures. Older firms may have more established wage-setting practices compared to younger firms (Santhosh 2023).
Asset structure (KS): calculated as the ratio of net fixed assets and inventory to total assets, this variable reflects the firm’s capital intensity, which can influence wage disparities by determining the firm’s reliance on capital versus labour (Eliasy and Przychodzen 2020; Wang 2022).
Capital-labour ratio (KL): measured as the logarithm of capital stock divided by the number of employees, this variable represents the firm’s capital-labour mix, a critical factor influencing wage differentials, particularly in the context of AI transformation (ElFayoumi 2024).
Leverage (Lev): defined as the ratio of total liabilities to total assets, this variable captures the firm’s financial risk. Highly leveraged firms may have less flexibility in wage setting due to financial constraints (Giroud and Mueller 2021).
Profitability (ROA): defined as net income divided by total assets, this variable controls for the effect of firm performance on wages. More profitable firms may offer higher wages due to better financial health (Babina et al. 2024).
Ownership concentration (TOP): Measured as the sum of shares held by the top three shareholders, this variable reflects the degree of ownership concentration, which can influence managerial power and wage-setting policies. High ownership concentration may lead to more centralized decision-making in wage distribution (Yasser and Mamun 2017).
Regional GDP per capita (AGDP): defined as the average GDP per capita in the region where the firm operates, controlling for regional economic development’s effect on wage levels (Edeh and Prévot 2024).

Table 1 presents descriptive statistics for the variables used in the empirical analysis. In addition to the dependent variable (Wsp) and the key explanatory variable (AI), the table reports summary statistics for the firm-level and regional control variables included in the baseline regressions.

Table 1. Descriptive statistics

Variable symbol	Variable meaning	Mean	Median	Std.	Min	Max
Wsp	Within-firm wage gap	8.475	8.373	1.135	1.486	14.012
AI	AI transformation	0.916	0.6931	1.134	0	4.454
Age	Firm age	2.884	2.944	0.331	1.792	3.526
KS	Asset structure	0.364	0.349	0.159	0	0.954
KL	Capital-labour ratio	14.108	14.064	1.027	10.612	17.219
Lev	Leverage	0.392	0.379	0.200	0.053	0.959
ROA	Profitability	0.055	0.044	0.077	0.000	0.866
TOP	Ownership concentration	49.697	49.381	15.252	0.565	98.29
AGDP	Regional GDP per capita	11.102	11.138	0.454	9.482	12.013

Sources: Our own calculations, based on data taken from the Wind Economic Database, the China Stock Market and Accounting Research (CSMAR) Database, Cninfo, the China Statistical Yearbook and the China City Statistical Yearbook.

4. Results

4.1 Baseline regression results

Table 2 presents the baseline regression results examining the impact of industrial intelligent transformation on within-firm wage disparities. Columns (1) through (3) display estimates of how AI adoption influences wage gaps within manufacturing firms.

Table 2. Baseline regression results

	Wsp
	(1)	(2)	(3)
AI	0.017** (0.007)	0.029*** (0.007)	0.018*** (0.007)
Age		0.663*** (0.065)	–0.355*** (0.096)
KS		–0.535*** (0.321)	–0.420*** (0.062)
KL		0.097*** (0.011)	0.113*** (0.011)
Lev		–0.422*** (0.053)	–0.354*** (0.053)
ROA		0.806*** (0.068)	0.809*** (0.067)
TOP		0.009*** (0.001)	0.008*** (0.001)
AGDP		0.404*** (0.045)	–0.031 (0.057)
Constant	8.458*** (0.008)	3.308*** (0.374)	11.290*** (0.712)
Firm	YES	YES	YES
Year	YES	NO	YES
Adjusted R²	0.771	0.777	0.781
Observations	18 314	18 314	18 314

Notes: *, ** and *** indicate statistical significance at the 10, 5 and 1 per cent levels, respectively. Standard errors are reported in parentheses.

Sources: Our own calculations, based on data taken from the Wind Economic Database, the China Stock Market and Accounting Research (CSMAR) Database, Cninfo, the China Statistical Yearbook and the China City Statistical Yearbook.

In column (1), the regression includes only the core explanatory variable while controlling for firm and year fixed effects. The estimated coefficient on AI is positive and statistically significant at the 5 per cent level, indicating that intelligent transformation significantly increases wage disparities within firms.

Column (2) adds a set of firm-level control variables – such as firm age, size, capital intensity, leverage, profitability and productivity – while maintaining firm fixed effects. The coefficient on AI remains positive and becomes even more significant at the 1 per cent level, with a magnitude of 0.029. This suggests that the inclusion of control variables does not attenuate the core relationship.

To enhance the robustness of the baseline results, column (3) further includes year fixed effects alongside firm fixed effects and control variables. The AI coefficient remains statistically significant at the 1 per cent level, with a similar magnitude, reinforcing the finding that AI adoption exacerbates within-firm wage disparities.

These results suggest a systemic effect of AI on wage inequality across the manufacturing sector, independent of firm-specific characteristics or macroeconomic fluctuations. The findings align with previous studies highlighting the inequality-enhancing effects of AI and automation (Autor, Katz and Kearney 2006; Autor and Dorn 2013; Goos, Manning and Salomons 2014; Barth et al. 2020; Humlum 2022). We therefore accept H1-1 and reject H1-2.

4.2 Addressing endogeneity in AI transformation and wage gaps

A potential concern in the baseline analysis is the issue of reverse causality: firms exhibiting wider within-firm wage gaps may be more inclined to adopt intelligent manufacturing technologies in pursuit of improved labour productivity. This introduces the possibility of endogeneity, which may bias the estimated impact of AI on wage disparities.

To address this concern, we employ an instrumental variable (IV) approach using two-stage least squares (2SLS) estimation. Specifically, we use the number of AI-related invention patents as the instrument. This variable is constructed by matching the main classification codes of invention patents with AI-related keywords, identifying and aggregating the number of AI-related patents filed by listed firms during the sample period, and then taking the logarithm.

This choice of instrument is justified on two grounds. First, innovation in AI technologies is a fundamental component of intelligent transformation and reflects a firm’s technical capability and innovation intensity. Second, while AI patent activity strongly correlates with a firm’s engagement in intelligent transformation, it is unlikely to directly affect within-firm wage disparities – thus satisfying the exogeneity condition required for a valid instrument.

The first-stage results in column (1) of table 3 show that the IV is positively and significantly correlated with the endogenous variable (AI adoption) at the 1 per cent level, supporting the instrument’s relevance. Further diagnostic tests confirm the strength and validity of the instrument:

First, the Kleibergen–Paap rk Wald F statistic is 1,698.36, which is well above the Stock-Yogo weak identification critical value of 16.38 for a 10 per cent maximal IV size distortion. This result rejects the null hypothesis of weak identification.
Second, the Kleibergen–Paap rk LM statistic is significant at the 1 per cent level, rejecting the null hypothesis of under-identification.

Table 3. Addressing endogeneity

	(1)	(2)
	AI	Wsps
AI		0.049** (0.022)
IV	0.082*** (0.002)
Controls	YES	YES
Firm	YES	YES
Year	YES	YES
K-P rk Wald F	1 698.36
K-P rk LM	1 799.95
Observations	18 314	18 314

Notes: *, ** and *** indicate statistical significance at the 10, 5 and 1 per cent levels, respectively. Standard errors are reported in parentheses.

Sources: Our own calculations, based on data taken from the Wind Economic Database, the China Stock Market and Accounting Research (CSMAR) Database, Cninfo, the China Statistical Yearbook and the China City Statistical Yearbook.

Taken together, these results affirm the validity of the constructed instrument.

The second-stage regression results, reported in column (2), indicate that AI adoption continues to exert a significantly positive effect on within-firm wage disparities, even after accounting for potential endogeneity. The estimated coefficient is statistically significant at the 5 per cent level, reinforcing the robustness of the baseline findings. This suggests that intelligent transformation remains a key driver of increased intra-firm wage gaps, particularly along skill and task dimensions.

4.3 Robustness checks

To ensure the robustness of the baseline results, we conduct a series of robustness checks by modifying both the dependent and key explanatory variables. First, we conduct additional tests by employing alternative measures of within-firm wage inequality based on skill and task premiums, following the measurement strategies of Egger and Kreickemeier (2009) and Chen, Yu and Yu (2017). Two alternative indicators of wage disparity are constructed. For the skill premium, employees with an undergraduate and/or graduate degree are classified as high-skilled workers, while those with lower educational attainment are categorized as low-skilled workers. For the task premium, employees engaged in research and development, management, marketing and finance are classified as non-routine-task workers, whereas workers involved in general production and administrative activities are categorized as routine-task workers.

Because firm-level wage data by worker type are not publicly available, we adopt a proxy approach inspired by the fair wage model. Specifically, the wage of low-skilled or routine workers is approximated using the lowest firm-level average wage within the same industry-region cell, which serves as a benchmark wage for low-tier labour.

In addition to the established approaches discussed above, this proxy is also consistent with structural characteristics of China’s labour market. Wage levels vary substantially across regions and industries due to differences in economic development, cost of living and labour mobility. In this context, the lowest firm-level wage within an industry-region cell often approximates the prevailing wage for low-skilled or routine labour, particularly in labour-intensive manufacturing sectors. Moreover, these wages frequently align with local minimum wage standards, which tend to bind compensation for entry-level or unskilled workers. Therefore, under current data constraints, this proxy provides a reasonable approximation of baseline labour compensation.

Based on this benchmark, the firm-level wage premium is calculated as follows:

$W s p i t j = W i t H j − W i t L j = W ¯ i t j − 1 − α i t j W i t L j α i t j − W i t L j = W ¯ i t j − W i t L j α i t j, j ∈ s, t$

Where:

$W s p i t j = W i t H j − W i t L j$ , the skill or task premium.
$α i t s$ and $α i t t$ represent the proportion of high-skilled or non-routine task labour, respectively.
$W ¯ i t$ is the average wage in year t.
$W i t H j$ and $W i t L j$ refer to the high-wage and low-wage groups, respectively, based on skill level or task type.

Based on this procedure, two alternative dependent variables are constructed: within-firm skill premium (Wsp_s, where s denotes skill) and within-firm task premium (Wsp_t, where t denotes task). The regression results reported in columns (1) and (2) of table 4 show that the coefficient on AI adoption remains positive and statistically significant, indicating that the main findings are robust to alternative measures of wage inequality.

Table 4. Robustness checks

	(1)	(2)	(3)	(4)	(5)	(6)
	Wsps	Wspt	Wsp	Wsp	Wsp	Wsp
AI	0.124*** (0.043)	0.107** (0.042)			0.021** (0.011)	0.017** (0.007)
AI1			0.025** (0.009)
AI2				0.022*** (0.008)
Controls	YES	YES	YES	YES	YES	YES
Firm	YES	YES	YES	YES	YES	YES
Year	YES	YES	YES	YES	YES	YES
Year×province	NO	NO	NO	NO	NO	YES
Industry×province	NO	NO	NO	NO	NO	YES
Adjusted R²	0.676	0.678	0.808	0.782	0.609	0.804
Observations	18 314	18 314	11 941	16 123	10 412	18 292

Notes: *, ** and *** indicate statistical significance at the 10, 5 and 1 per cent levels, respectively. Standard errors are reported in parentheses.

Sources: Our own calculations, based on data taken from the Wind Economic Database, the China Stock Market and Accounting Research (CSMAR) Database, Cninfo, the China Statistical Yearbook and the China City Statistical Yearbook.

Next, we replace the core explanatory variable (AI) with alternative proxies:

AI1: the frequency of AI-related keywords disclosed in firm annual reports. While such references may include standardized language, this variable offers a broader perspective on a firm’s AI-related activities across operations.
AI2: the penetration rate of industrial robots at the firm level, constructed using data from the International Federation of Robotics, reflecting the intensity of intelligent transformation driven by robotics.

Results in columns (3) and (4) show that both alternative proxies yield significant and positive coefficients, reaffirming the baseline conclusion.

To address potential sample selection bias, which arises because AI adoption is not random, but influenced by firm characteristics (e.g. human capital, management practices and technology capabilities), we apply Propensity Score Matching (PSM). Firms are divided into a treatment group (those mentioning AI keywords) and a control group, and matched 1:1 (with replacement) based on observable covariates.

Balance tests confirm that post-matching, the standard differences in covariates between the two groups reduce by 54.6–97.9 per cent, with most standardized biases falling below 5 per cent and t-tests showing no significant differences – indicating strong matching quality. The PSM regression result in column (5) remains significantly positive, confirming that the findings are robust to self-selection concerns.

Lastly, to account for potential omitted variable bias from unobserved macroeconomic or regional shocks, following Bai (2009), we include multi-dimensional fixed effects in the baseline model, adding both year-province and industry-province interactions. As shown in column (6), the coefficient remains significantly positive and consistent with previous results, reinforcing the finding that AI adoption significantly increases within-firm wage gaps, even after controlling for more complex contextual factors.

Overall, across different definitions of the dependent and independent variables, as well as multiple estimation methods, the coefficient for AI adoption remains consistently positive and significant. This confirms the robustness and credibility of the baseline findings: intelligent transformation exacerbates within-firm wage disparities along both skill and task dimensions.

5. The impact of AI transformation on wage disparities

5.1 Mechanisms of AI transformation and their influence on wage disparities

5.1.1 Technology bias and productivity effect

The results presented in table 5 suggest that AI transformation influences within-firm wage disparities primarily through technology bias and productivity effects. This study uses Total Factor Productivity (TFP) – calculated via the DEA-Malmquist index method – as a proxy for technical progress. To test for biased technical change, an interaction term between TFP and AI adoption is introduced into the fixed effects model. Specifically, the coefficient of AI on TFP is positive and significant (0.152***), indicating that the adoption of AI technologies enhances firms’ overall productivity and efficiency. This finding is consistent with previous studies demonstrating that intelligent transformation fosters productivity growth and operational optimization (Domini et al. 2022; Barth et al. 2020).

Table 5. Mechanisms of AI transformation and their influence on wage disparities

Variables	Technological bias		Productivity effects	Skill structure		Task structure
Variables	(1) Wsp	(2) TFP	(3) Wsp	(4) Skill	(5) Wsp	(6) Task	(7) Wsp
AI		0.152*** (0.039)		0.261*** (0.075)		0.375*** (0.102)
TFP	0.399*** (0.015)		0.822** (0.264)
AI×TFP	0.002*** (0.000)
Skill					2.875** (1.186)
Task							2.873** (1.347)
Control variables	YES	YES	YES	YES	YES	YES	YES
Fixed effects	YES	YES	YES	YES	YES	YES	YES
Sample size	16 184	16 108	16 108	18 314	18 314	18 314	18 314

Notes: *, ** and *** indicate statistical significance at the 10, 5 and 1 per cent levels, respectively. Standard errors are reported in parentheses.

Sources: Our own calculations, based on data taken from the Wind Economic Database, the China Stock Market and Accounting Research (CSMAR) Database, Cninfo, the China Statistical Yearbook and the China City Statistical Yearbook.

However, the productivity gains from AI adoption are not evenly distributed among different groups of employees. When the interaction term between AI and TFP is included in the model, the coefficient remains positive and significant (0.002***), suggesting that AI-driven productivity improvements are skill- and task-biased. In other words, technological progress tends to complement workers engaged in non-routine and cognitively demanding tasks – such as executives, managers and technical professionals – while replacing routine and lower-skilled roles. As a result, AI-driven productivity improvements increase the relative wage of high-skilled and non-routine workers, thereby widening intra-firm wage disparities.

This pattern is consistent with the predictions of SBTC, TBTC and RBTC frameworks (Autor, Levy and Murnane 2003; Acemoglu and Autor 2011; Violante 2008), as well as empirical evidence from Humlum (2022), who found that technology workers benefited from robot adoption while production-line employees suffered wage declines. Thus, even though AI enhances firm-level productivity, the associated biased technological progress amplifies hierarchical wage disparities within firms – a pattern also observed by Barth et al. (2020). We therefore accept H2-1 and reject H2-2.

5.1.2 Changes in skill and task structure

AI transformation also alters firms’ internal labour demand structures by reshaping the relative importance of skill and task composition. Columns (4) through (7) of table 5 show that AI adoption significantly increases the proportion of high-skilled and non-routine-task labour within firms. This indicates that intelligent transformation induces firms to upgrade their human capital structure, favouring employees capable of complementing intelligent systems and engaging in analytical, managerial and creative functions.

The coefficient for AI adoption (0.261***) demonstrates that AI adoption enhances the proportion of high-skilled workers, consistent with evidence that automation and digitalization increase the demand for skilled labour and contribute to a skills premium (Acemoglu 2002; Moll, Rachel and Restrepo 2022; Hémous and Olsen 2022). Furthermore, when the skill structure variable is incorporated into the regression, its coefficient remains positive and significant (2.875**), suggesting that firms with a higher proportion of skilled employees also experience a larger wage gap between managerial and regular workers. This implies that skill upgrading, while enhancing firm efficiency, contributes to wage stratification.

Similarly, the results for the task structure channel (columns (6) through (7)) show that AI adoption is positively associated with the share of non-routine-task workers (0.375***), and that task structure itself significantly increases the wage gap (2.873**). This supports the technology-task complementarity hypothesis (Autor, Katz and Kearney 2006; Violante 2008), which posits that AI technologies complement non-routine, cognitive and interactive tasks more effectively than routine ones. Consequently, AI transformation amplifies intra-firm income inequality by increasing the relative demand and bargaining power of workers engaged in non-routine, high-responsibility roles.

Overall, these results confirm that AI-driven intelligent transformation affects within-firm wage disparities through multiple, interrelated mechanisms. Productivity gains from AI adoption are biased toward high-skilled and non-routine labour, while the transformation of firms’ skill and task structures further deepens the internal wage divide between management and regular employees.

We therefore accept H3-1 and reject H3-2.

5.2 Firm-level heterogeneity in the impact of AI transformation on the wage gap

The impact of AI-driven transformation on intra-firm wage inequality may vary significantly across different types of firms and industry contexts. To explore this heterogeneity, we analyse the differential effects across two dimensions: factor intensity and digital infrastructure.

5.2.1 Factor intensity: Labour-intensive versus capital-intensive industries

The degree of capital intensity in firms can shape how they benefit from AI adoption. Labour-intensive firms, which rely heavily on manual labour, may experience fewer productivity gains from AI due to their lower technological base. Capital-intensive firms, on the other hand, are better positioned to integrate AI into production processes, magnifying returns to high-skilled labour.

To measure factor intensity, we use the log of the ratio of net fixed assets to total employment. Firms above the median are classified as capital-intensive, and those below as labour-intensive.

Table 6, column (1) shows that AI adoption has an insignificant effect on wage inequality in labour-intensive firms (0.016, SE = 0.016), while column (2) shows a significant positive effect for capital-intensive firms (0.025***, SE = 0.049). This finding reinforces the notion that AI benefits accrue more rapidly in technologically advanced settings, which demand a higher proportion of skilled and non-routine workers – thus exacerbating internal wage disparities.

Table 6. Heterogeneous effects of AI transformation on the wage gap: Factor intensity and digital infrastructure

Variables	Factor intensity			Digital infrastructure
	(1)	(2)	(3)	(4)
	Labour-intensive	Capital-intensive	High	Low
AI	0.016 (0.016)	0.025*** (0.049)	0.025** (0.010)	0.016 (0.012)
Controls	YES	YES	YES	YES
Firm	YES	YES	YES	YES
Year	YES	YES	YES	YES
Adjusted R²	0.804	0.778	0.816	0.776
Observations	8 810	10 164	8 862	9 200

Notes: *, ** and *** indicate statistical significance at the 10, 5 and 1 per cent levels, respectively. Standard errors are reported in parentheses.

Sources: Our own calculations, based on data taken from the Wind Economic Database, the China Stock Market and Accounting Research (CSMAR) Database, Cninfo, the China Statistical Yearbook and the China City Statistical Yearbook.

5.2.2 Digital infrastructure: High- versus low-development regions

AI transformation depends heavily on the availability of supporting digital infrastructure. Regions with more robust digital ecosystems – such as high-speed internet and data connectivity – are likely to facilitate deeper integration of intelligent technologies, amplifying their labour market effects.

We classify regions based on the median ratio of internet access ports to population. Firms located in regions above the median are considered to have high digital infrastructure, while those below fall into the low infrastructure category.

Results in columns (3) and (4) show a clear divergence. In regions with advanced digital infrastructure, the impact of AI on wage inequality is positive and significant (0.025**, SE = 0.010), whereas in low-infrastructure regions, the effect is positive but statistically insignificant (0.016, SE = 0.012). This suggests that firms in digitally advanced regions can more effectively leverage AI technologies, leading to stronger differentiation in labour value and compensation.

5.3 Technology bias and skill displacement effects

While AI-driven transformation can enhance productivity and drive skill upgrading, it may also generate unintended labour market consequences – particularly for low-skilled workers. One such consequence is the phenomenon of “skill crowding-out”, whereby increased demand for high-skilled labour displaces low-skilled workers within firms.

To explore this effect, we begin by examining the impact of industrial AI transformation on the employment of low-skilled labour (Lski) and routine-task labour (Rtask). As shown in columns (1) and (2) of table 7, the coefficients for AI are significantly negative for both Lski (–0.036***, SE = 0.005) and Rtask (–0.010***, SE = 0.002), indicating that AI adoption leads to a substantial decline in the employment of low-skilled and routine-task workers. These results suggest that AI technologies, while boosting overall efficiency, may replace workers engaged in routine, easily codifiable tasks.

Table 7. Impact of AI capital investments on labour market dynamics

Variables	Skill crowding-out		Capital-skill complementarity
	(1)	(2)	(3)	(4)
	Lski	Rtask	Wsps	Task
AI	–0.036*** (0.005)	–0.010*** (0.002)
Hski×AIK			0.004*** (0.001)
Lski×AIK			–0.001 (0.001)
Utask×AIK				0.002*** (0.000)
Rtask×AIK				–0.001 (0.001)
Controls	YES	YES	YES	YES
Firm	YES	YES	YES	YES
Year	YES	YES	YES	YES
Adjusted R²	0.930	0.972	0.806	0.806
Observations	17 047	16 911	13 074	13 074

Notes: *, ** and *** indicate statistical significance at the 10, 5 and 1 per cent levels, respectively. Standard errors are reported in parentheses.

Sources: Our own calculations, based on data taken from the Wind Economic Database, the China Stock Market and Accounting Research (CSMAR) Database, Cninfo, the China Statistical Yearbook and the China City Statistical Yearbook.

To further probe the capital-skill complementarity mechanism, we analyse whether AI-related capital complements or substitutes for different types of labour. We use the logarithm of intangible assets related to intelligent technology (AIK) as a proxy for a firm’s AI-specific capital. Interaction terms between AIK and the logarithm of high-skilled (Hski) and low-skilled (Lski) labour, as well as between AIK and unconventional (Utask) and routine-task (Rtask) labour, are introduced to assess heterogeneity in complementarity effects.

In column (3), the interaction term Hski × AIK shows a significantly positive coefficient (0.004***, SE = 0.001), while Lski × AIK is statistically insignificant (–0.001, SE = 0.001). This indicates that AI capital significantly enhances the wage premium of high-skilled workers, supporting the theory of capital-skill complementarity. In other words, AI-related intangible capital and high-skilled labour are complementary, reinforcing the wage advantages of skilled workers.

Similarly, column (4) reports a significantly positive interaction effect between Utask × AIK (0.002***, SE = 0.000) and an insignificant effect for Rtask × AIK (–0.001, SE = 0.001). This suggests that AI capital increases the relative demand and remuneration for non-routine labour, while routine-task workers do not experience significant gains – further reinforcing the task-biased nature of technological change.

Together, these findings indicate that AI transformation exacerbates intra-firm wage disparities by disproportionately benefiting high-skilled and non-routine workers. Low-skilled and routine-task workers not only fail to benefit from productivity gains but also face crowding-out effects as their roles are increasingly automated or devalued.

6. Conclusion

This study offers an in-depth analysis of how AI transformation affects within-firm wage gaps in Chinese manufacturing firms. Leveraging firm-level textual disclosures, we capture not only the extent of AI adoption, but also the strategic intent behind AI transformation, offering a more comprehensive and granular measure of technological change (Yu, Wang and Li 2020; Jiang, Wang and Liu 2024).

Our findings reveal that AI transformation significantly widens the wage gap between executives and regular employees, reinforcing the view that AI-driven technological change redistributes income unevenly within firms. This effect is particularly pronounced in China’s manufacturing sector, which has shifted from being traditionally labour-intensive to being increasingly capital-intensive and digitally integrated. These structural shifts heighten the concentration of AI benefits among a small group of high-skilled, high-responsibility employees while limiting the wage gains – or even employment prospects – of routine-task and low-skilled workers.

The widening of the within-firm wage gap is driven by two interconnected mechanisms. The first is the technology bias effect, wherein AI-induced productivity improvements disproportionately benefit high-skilled, non-routine task workers. Firms with stronger AI capabilities – measured through higher TFP – experience more pronounced wage divergence between executives and employees. These productivity gains exhibit clear skill and task bias, as AI technologies complement analytical, managerial and creative roles while reducing reliance on routine- and repetitive-task workers (Domini et al. 2022; Acemoglu and Restrepo 2018; Graetz and Michaels 2015). The second is the skill displacement effect, whereby AI transformation restructures firms’ labour demand. As intelligent systems automate routine processes, firms reduce the share of low-skilled and routine-task workers while expanding roles that require higher levels of judgment, technical expertise and decision-making. This reallocation of labour resources contributes directly to the divergence between executive and employee compensation, as routine workers become increasingly redundant or undervalued relative to strategic and technical staff.

Heterogeneity analyses further support these findings. Capital-intensive firms show stronger AI-inequality links, which is likely to be due to their greater capacity to absorb and implement automation technologies. Importantly, regional digital infrastructure also plays a role: firms operating in digitally advanced areas experience a greater widening of the wage gap, as infrastructure enables deeper AI integration.

Although this study primarily focuses on within-firm wage inequality, robustness checks based on skill and task premium measures confirm that similar patterns of unequal benefit distribution persist across alternative definitions of wage disparity. These findings underscore the systemic nature of AI’s redistributive effects in the organizational context.

From a policy perspective, mitigating the inequality risks posed by AI adoption requires multi-layered interventions. Reskilling programmes tailored to low-skilled workers, investment in digital infrastructure in underdeveloped regions and wage equity initiatives in high-tech firms are all crucial to ensuring that the gains from intelligent transformation are more evenly distributed. Moreover, ownership-specific guidelines and industry-level wage regulations may help balance firm efficiency with social equity.

This study acknowledges some limitations. First, the data span from 2012 to 2020 and may not capture long-term AI impacts or the latest technological developments. Extending the time frame to include post-2020 data would yield more forward-looking insights. Second, while focused on the manufacturing sector, future research could apply similar methods across services, agriculture or finance to offer comparative cross-sectoral perspectives.

In summary, this study finds that AI transformation intensifies within-firm wage gaps in Chinese manufacturing, largely through technology-biased productivity gains and labour restructuring that displaces low-skilled workers. These results highlight the importance of aligning industrial upgrading with inclusive labour market policies, ensuring that AI-driven growth does not come at the cost of rising inequality.

Acknowledgments

This research was supported by the National Social Science Fund of China (Grant No. 25CJY119). The authors would like to thank their colleagues, the anonymous reviewers and the editor for their helpful comments on earlier versions of this manuscript. Any remaining errors or omissions are the authors’ own.

Competing interests

The authors declare that they have no competing interests.

Notes

A-shares are RMB-denominated ordinary shares issued by mainland Chinese firms and traded on domestic stock exchanges. In this study, we focus on firms listed on the main boards of the Shanghai and Shenzhen Stock Exchanges to ensure consistency and comparability in financial reporting and listing requirements. ⮭
In China’s A-share market, ST denotes “special treatment”, a regulatory designation for listed firms subject to special risk monitoring because of abnormal financial conditions or other major risks. *ST indicates a more severe status, usually associated with continued financial distress and a higher risk of delisting. ⮭
The AI transformation keyword set includes intelligent manufacturing (智能制造), digitalization (数字化), intelligence (智能化), informatization (信息化), automation (自动化), cloud computing (云计算), cloud platform (云平台), Internet of Things (物联网), cyber-physical systems (信息物理系统), networked physical systems (网络物理系统), big data (大数据), sensing technology (感知技术), cloud manufacturing (云制造), proactive manufacturing (主动制造), smart manufacturing (智慧制造), intelligent enterprise (智能企业), intelligent terminal (智能终端), intelligent recognition (智能识别), robots (机器人), Industry 4.0 (工业4.0), industrial internet (工业互联网), Internet+ (互联网+), human-machine interaction (人机交互), sensors (传感器), controllers (控制器), data mining (数据挖掘) and data visualization (数据可视化). ⮭

References

Acemoglu, Daron. 2002. “Technical Change, Inequality, and the Labor Market”. Journal of Economic Literature 40 (1): 7–72. http://doi.org/10.1257/0022051026976.

Acemoglu, Daron, and David Autor. 2011. “Skills, Tasks and Technologies: Implications for Employment and Earnings”. In Handbook of Labor Economics, Vol. 4, Part B, edited by David Card and Orley Ashenfelter, 1043–1171. Amsterdam: Elsevier.

Acemoglu, Daron, and Pascual Restrepo. 2018. “Artificial Intelligence, Automation and Work”. NBER Working Paper No. 24196. Cambridge, MA: National Bureau of Economic Research.

Acemoglu, Daron, and Pascual Restrepo. 2020. “The Wrong Kind of AI? Artificial Intelligence and the Future of Labour Demand”. Cambridge Journal of Regions, Economy and Society 13 (1): 25–35. http://doi.org/10.1093/cjres/rsz022.

Afonso, Oscar, Tiago Sequeira, and Derick Almeida. 2023. “Technological Knowledge and Wages: From Skill Premium to Wage Polarization”. Journal of Economics 140 (2): 93–119. http://doi.org/10.1007/s00712-023-00833-y.

Autor, David H., and David Dorn. 2013. “The Growth of Low-Skill Service Jobs and the Polarization of the US Labor Market”. American Economic Review 103 (5): 1553–1597. http://doi.org/10.1257/aer.103.5.1553.

Autor, David H., Lawrence F. Katz, and Melissa S. Kearney. 2006. “The Polarization of the US Labor Market”. American Economic Review 96 (2): 189–194. http://doi.org/10.1257/000282806777212620.

Autor, David H., Frank Levy, and Richard J. Murnane. 2003. “The Skill Content of Recent Technological Change: An Empirical Exploration”. The Quarterly Journal of Economics 118 (4): 1279–1333. http://doi.org/10.1162/003355303322552801.

Babina, Tania, Anastassia Fedyk, Alex He, and James Hodson. 2024. “Artificial Intelligence, Firm Growth, and Product Innovation”. Journal of Financial Economics 151 (January): Article No. 103745. http://doi.org/10.1016/j.jfineco.2023.103745.

Bai, Jushan. 2009. “Panel Data Models with Interactive Fixed Effects”. Econometrica 77 (4): 1229–1279. http://doi.org/10.3982/ECTA6135.

Banker, Rajiv D., Danlu Bu, and Mihir N. Mehta. 2016. “Pay Gap and Performance in China”. Abacus 52 (3): 501–531. http://doi.org/10.1111/abac.12082.

Barth, Erling, Marianne Roed, Pål Schøne, and Janis Umblijs. 2020. “How Robots Change Within-Firm Wage Inequality”. IZA Discussion Paper No. 13605. Bonn: Institute for Labor Economics.

Chen, Bo, Miaojie Yu, and Zhihao Yu. 2017. “Measured Skill Premia and Input Trade Liberalization: Evidence from Chinese Firms”. Journal of International Economics 109 (November): 31–42. http://doi.org/10.1016/j.jinteco.2017.08.005.

Cords, Dario, and Klaus Prettner. 2022. “Technological Unemployment Revisited: Automation in a Search and Matching Framework”. Oxford Economic Papers 74 (1): 115–135. http://doi.org/10.1093/oep/gpab022.

Damioli, Giacomo, Vincent van Roy, and Daniel Vertesy. 2021. “The Impact of Artificial Intelligence on Labor Productivity”. Eurasian Business Review 11 (1): 1–25. http://doi.org/10.1007/s40821-020-00172-8.

Deloitte China. 2020. AI Enablement on the Way to Smart Manufacturing: Deloitte Survey on AI Adoption in Manufacturing.

Domini, Giacomo, Marco Grazzi, Daniele Moschella, and Tania Treibich. 2022. “For Whom the Bell Tolls: The Firm-Level Effects of Automation on Wage and Gender Inequality”. Research Policy 51 (7): Article No. 104533. http://doi.org/10.1016/j.respol.2022.104533.

Edeh, Jude, and Frédéric Prévot. 2024. “Beyond Funding: The Moderating Role of Firms’ R&D Human Capital on Government Support and Venture Capital for Regional Innovation in China”. Technological Forecasting & Social Change 203 (June): Article No. 123351. http://doi.org/10.1016/j.techfore.2024.123351.

Egger, Hartmut, and Udo Kreickemeier. 2009. “Firm Heterogeneity and the Labor Market Effects of Trade Liberalization”. International Economic Review 50 (1): 187–216. http://doi.org/10.1111/j.1468-2354.2008.00527.x.

ElFayoumi, Khalid. 2024. “Firm Financing and the Relative Demand for Labor and Capital”. Journal of Economic Dynamics & Control 168 (November): Article No. 104946. http://doi.org/10.1016/j.jedc.2024.104946.

Eliasy, Ashkan, and Justyna Przychodzen. 2020. “The Role of AI in Capital Structure to Enhance Corporate Funding Strategies”. Array 6 (July): Article No. 100017. http://doi.org/10.1016/j.array.2020.100017.

Faleye, Olubunmi, Ebru Reis, and Anand Venkateswaran. 2013. “The Determinants and Effects of CEO-Employee Pay Ratios”. Journal of Banking & Finance 37 (8): 3258–3272. http://doi.org/10.1016/j.jbankfin.2013.03.003.

Freeman, Richard B., Ina Ganguli, and Michael J. Handel. 2020. “Within-Occupation Changes Dominate Changes in What Workers Do: A Shift-Share Decomposition, 2005–2015”. AEA Papers and Proceedings 110 (May): 394–399. http://doi.org/10.1257/pandp.20201005.

Georgieff, Alexandre. 2024. “Artificial Intelligence and Wage Inequality”. OECD Artificial Intelligence Papers No. 13. Paris: Organisation for Economic Co-operation and Development.

Giroud, Xavier, and Holger M. Mueller. 2021. “Firm Leverage and Employment Dynamics”. Journal of Financial Economics 142 (3): 1381–1394. http://doi.org/10.1016/j.jfineco.2021.05.006.

Goos, Maarten, Alan Manning, and Anna Salomons. 2014. “Explaining Job Polarization: Routine-Biased Technological Change and Offshoring”. The American Economic Review 104 (8): 2509–2526. http://doi.org/10.1257/aer.104.8.2509.

Gordon, David M. 1996. Fat and Mean: The Corporate Squeeze of Working Americans and the Myth of Managerial “Downsizing”. New York, NY: Martin Kessler Books.

Graetz, Georg, and Guy Michaels. 2015. “Robots at Work”. CEP Discussion Paper No. 1335. London: Center for Economic Performance, London School of Economics and Political Science.

Hémous, David, and Morten Olsen. 2022. “The Rise of the Machines: Automation, Horizontal Innovation, and Income Inequality”. American Economic Journal: Macroeconomics 14 (1): 179–223. http://doi.org/10.1257/mac.20160164.

Howcroft, Debra, and Phil Taylor. 2023. “Automation and the Future of Work: A Social Shaping of Technology Approach”. New Technology, Work and Employment 38 (2): 351–370. http://doi.org/10.1111/ntwe.12240.

Huang, Hanwei, Jiandong Ju, and Vivian Yue. 2024a. “Accounting for the Evolution of China’s Production and Trade Patterns”. NBER Working Paper No. 32415. Cambridge, MA: National Bureau of Economic Research.

Huang, Hanwei, Jiandong Ju, and Vivian Yue. 2024b. “Understanding the Evolution of China’s Production and Trade Patterns”. VoxChina, 9 October 2024. https://voxchina.org/show-3-377.html.

Humlum, Anders. 2022. “Robot Adoption and Labor Market Dynamics”. Study Paper No. 175. Copenhagen: Rockwool Foundation Research Unit.

Hunt, Jennifer, and Ryan Nunn. 2019. “Is Employment Polarization Informative about Wage Inequality and Is Employment Really Polarizing?” NBER Working Paper No. 26064. Cambridge, MA: National Bureau of Economic Research.

IFR (International Federation of Robotics). 2023. World Robotics 2023: Industrial Robots. Frankfurt.

Jayachandran, Seema. 2006. “Selling Labor Low: Wage Responses to Productivity Shocks in Developing Countries”. Journal of Political Economy 114 (3): 538–575. http://doi.org/10.1086/503579.

Jiang, Hong, Xue Wang, and Chongguang Liu. 2024. “Automated Machines and the Labor Wage Gap”. Technological Forecasting & Social Change 206 (September): Article No. 123505. http://doi.org/10.1016/j.techfore.2024.123505.

Koch, Michael, Ilya Manuylov, and Marcel Smolka. 2021. “Robots and Firms”. The Economic Journal 131 (638): 2553–2584. http://doi.org/10.1093/ej/ueab009.

Mandelman, Federico S., and Andrei Zlate. 2022. “Offshoring, Automation, Low-Skilled Immigration, and Labor Market Polarization”. American Economic Journal: Macroeconomics 14 (1): 355–389. http://doi.org/10.1257/mac.20180205.

Moll, Benjamin, Lukasz Rachel, and Pascual Restrepo. 2022. “Uneven Growth: Automation’s Impact on Income and Wealth Inequality”. Econometrica 90 (6): 2645–2683. http://doi.org/10.3982/ECTA19417.

Ozgul, Pelin, Marie-Christine Fregin, Michael Stops, Simon Janssen, and Mark Levels. 2024. “High-Skilled Human Workers in Non-Routine Jobs Are Susceptible to AI Automation but Wage Benefits Differ between Occupations”. arXiv:2404.06472. http://doi.org/10.48550/arXiv.2404.06472.

Porter, Michael E., and James E. Heppelmann. 2014. “How Smart, Connected Products Are Transforming Competition”. Harvard Business Review 92 (11): 64–88.

Santhosh, Channappa. 2023. “Firm Age, Sales Growth and Survival of Internationalized SMEs: Does Human Capital Play a Role?” Journal of Economic and Administrative Sciences 39 (4): 1180–1194. http://doi.org/10.1108/JEAS-09-2020-0168.

Sharfaei, Shahab. 2024. “Robots and Employment: A Review”. International Labour Review 163 (2): 271–293. http://doi.org/10.1111/ilr.12417.

Violante, Giovanni L. 2008. “Skill-Biased Technical Change”. In The New Palgrave Dictionary of Economics, 2nd ed., edited by Steven N. Durlauf and Lawrence E. Blume, 1–6. London: Palgrave Macmillan.

Wang, Chenxi. 2022. “Firm Asset Structure and Risk Aversion”. Economics Letters 221 (December): Article No. 110913. http://doi.org/10.1016/j.econlet.2022.110913.

Wu, Yongqiu, Zhiwei Lin, Qingcui Zhang, and Wei Wang. 2024. “Artificial Intelligence, Wage Dynamics, and Inequality: Empirical Evidence from Chinese Listed Firms”. International Review of Economics & Finance 96 (Part C): Article No. 103739. http://doi.org/10.1016/j.iref.2024.103739.

Yasser, Qaiser Rafique, and Abdullah Al Mamun. 2017. “The Impact of Ownership Concentration on Firm Performance: Evidence from an Emerging Market”. Emerging Economy Studies 3 (1): 34–53. http://doi.org/10.1177/2394901517696647.

Yu, Feifei, Liting Wang, and Xiaotong Li. 2020. “The Effects of Government Subsidies on New Energy Vehicle Enterprises: The Moderating Role of Intelligent Transformation”. Energy Policy 141 (June): Article No. 111463. http://doi.org/10.1016/j.enpol.2020.111463.

Yuan, Yiming, Yongming Sun, and Hangyu Chen. 2025. “Does Artificial Intelligence Affect Firms’ Inner Wage Gap?” Applied Economics 57 (19): 2365–2371. http://doi.org/10.1080/00036846.2024.2324090.