The Impact Mechanism of Artificial Intelligence Development on Water–Energy–Food System Technical Efficiency—An Empirical Study in China

Abstract

To investigate the interaction between artificial intelligence development (AID) and water–energy–food system technical efficiency (WEF-TE), panel data from 264 cities in China from 2013 to 2023 were utilized, and WEF-TE in the study areas was estimated using Stochastic Frontier Analysis (SFA). Subsequently, the Error Correction Model (ECM) and a random forest model were adopted for empirically examining the adjustment and driving mechanisms of AID on WEF-TE from three dimensions, namely enterprise scale, application level, and workforce literacy. The results indicate the following: (1) China’s WEF-TE generally shows an increasing trend; however, clear differences remain between high-value and low-value regions, and the deviation in lagging areas can reach 0.507. Meanwhile, the Yellow River Basin, which is the core region of China’s WEF system, remains below the national average in the process of technical efficiency optimization. (2) AID has a long-term equilibrium relationship with WEF-TE across the research dimensions and can effectively adjust technological inefficiencies in the short term, with adjustment coefficients ranging from 0.004 to 0.021 under different test rules. (3) In terms of enterprise scale and application level, the driving effect of AID on WEF-TE is relatively strong, with feature weights of 0.16 and 0.155, which are close to those of human capital input (0.172) and industrial structure rationalization (0.15). This study provides important reference value for constructing an interdisciplinary research framework that integrates WEF Nexus with AID.

1. Introduction

Water, energy, and food constitute the three important resources supporting human social development. However, due to rapid global population expansion, limited resource reserves, and environmental degradation, the supply of these three resources continues to encounter serious challenges, thereby restricting sustainable human development. As a result, related research has consistently remained a key focus of academic attention [1]. Since the Bonn Conference held in 2011 in Germany [2], interdisciplinary investigations on water, energy, and food have gradually attracted broader attention. The academic community generally agrees that, due to the numerous and complex interdependencies among these three resources in their production, consumption, and management processes, formulating strategies based on a single resource or only two resources involves considerable risks [3,4]. Accordingly, a unified research framework for water, energy, and food—the WEF Nexus theory—has been developed, emphasizing the importance of placing these three resources on a shared platform for integrated analysis.

At the WEF Nexus level, studies on WEF system technical efficiency (WEF-TE) have long been a central topic in academia, and their significance is mainly reflected in two aspects. First, improving efficiency can reduce the consumption and loss of these three resources during system operation, and it represents a critical approach to mitigating the conflict between local development and limited resources in the presence of diverse constraints [5,6]. Second, within the framework of a “green economy”, optimizing efficiency can enhance the expected output ratio of WEF systems and decrease the generation of common byproducts such as carbon dioxide, sulfides, and effluent, thereby alleviating continuous pressure on the system caused by environmental deterioration [7,8,9,10]. This discussion also reflects the sustainable development objectives embedded in WEF theory.

Considerable progress has been achieved in academic research concerning WEF-TE. Some researchers have directly assessed the efficiency of WEF systems within specific study regions. For instance, Maia et al., Ali et al., and Sun et al. all applied Data Envelopment Analysis (DEA) to evaluate the ecological WEF-TE in Brazil, Pakistan, and China, respectively. The primary differences were that Maia et al. adopted an industrial perspective of WEF systems, Ali et al. concentrated on analyzing the influence of external factors on local WEF systems, and Sun et al. examined the spatiotemporal evolution of regional WEF system efficiency [10,11,12]. Hu et al., using a super slack-based measure model, estimated the total factor productivity of WEF systems at the provincial scale in China and emphasized the role of recycling in system coupling and efficiency [13].

Some studies also integrate the WEF system with additional factors to perform a comprehensive technical efficiency evaluation. For example, Ibrahim et al. utilized DEA to assess the technical efficiency of the water–energy–land–food (WELF) system in OECD countries and reported that drought conditions reduce its operational efficiency [14]. Similarly, Yao et al. employed multidimensional performance indicators and spatial analysis approaches to evaluate WELF system technical efficiency and analyze heterogeneity at the provincial level in China [15]. After developing the Slack-Based Measure Data Envelopment Analysis (SBM-DEA) model, Ren et al. incorporated climate change factors to provide a more comprehensive assessment of the eco-efficiency of WEF systems in Chinese provinces [16]. Zhang et al. included food security within the WEF framework and used an extended dynamic series-loop data envelopment analysis model to assess the water–energy–food–food security (WEF-FS) efficiency across China’s provinces [17]. In addition, some researchers have applied the WEF framework to evaluate individual resource efficiency. For instance, Zhang et al. used an SBM model based on undesirable outputs to measure green energy efficiency across 30 provinces in China [18], whereas Wang et al. applied a similar approach to evaluate the grain total factor productivity within the Yellow River Basin of China [19].

Artificial intelligence development (AID) is increasingly emerging as a structural factor shaping future technology [20,21]. As a relatively broad concept, AID has been defined and measured in existing studies from multiple dimensions. Ma et al. and Dong et al. focused on enterprise scale and used the number of AI enterprises to represent the level of AI development in a region [22,23]. In contrast, Qian et al. and Lin et al. paid greater attention to AI application and R&D dimensions, measuring AID through industrial robot density and the number of AI-related patents derived from macroeconomic data [24,25]. Meanwhile, Zhang et al. and Guan et al. adopted a policy-oriented perspective, regarded China’s “National New Generation Artificial Intelligence Innovation and Development Pilot Zones” policy as an exogenous variable, and used the DID model to examine policy-driven changes in AID levels and their effects [26,27]. In addition, some scholars have defined AID from the perspectives of labor force literacy, infrastructure, and market advantages and have explored its direct or indirect influences on social and economic development [25,28]. Overall, because of limitations in data availability and research methods, existing studies on the dimensional definition of AID have mainly concentrated on enterprise scale and application levels, while different dimensions and measurement approaches have also been developed according to specific research subjects.

On the other hand, through data-driven methods [29] and intelligent algorithms [30], AI has demonstrated distinct capabilities in optimizing resource allocation [31], mitigating resource security risks [32], promoting the transformation and upgrading of resource-related industries [33], and reducing the costs of resource development [34]. It is reshaping patterns of resource utilization across multiple dimensions, thereby promoting the efficiency management of the WEF system. In the energy sector, on hand, AI can help achieve demand-driven allocation of energy in agricultural and water conservancy scenarios, and enhance the supply reliability of renewable energy sources such as solar and wind power [31,32,33]. On the other hand, the advancement of AI can optimize the efficiency of unexpected outputs through technology, scale effects, and technically skilled workforce [35,36,37], thereby alleviating pressure on WEF systems at the environmental level. And Yuan et al. and Liu et al. both further emphasize that this adjustment effect typically follows a U-shaped pattern [38,39]. Within the agricultural sector, Ding et al. And Amin et al. found that the application of AI can significantly enhance agricultural productivity, and the impact varies across different regions [40,41,42]. Nath et al. indicated that AI plays an important role in transforming food systems by improving efficiency of water, electricity and other basic elements [43]. In the field of water resources, adaptive AI techniques can improve water use efficiency through process optimization [44] while AI models can enhance the efficiency of water resource management in areas such as industrial wastewater recycling, water allocation, and agricultural irrigation [45,46].

In summary, on one hand, the academic community has conducted extensive investigations on measuring the WEF-TE and has achieved substantial progress in areas such as factor integration, model refinement, and heterogeneity assessment. On the other hand, scholars have clearly demonstrated that the application of AI can effectively enhance the technical efficiencies of energy, water, and food systems and contribute to their synergistic management. However, several limitations remain:

(1) Because of limitations in data availability, related studies have mainly focused on national and provincial scales, while city-level research remains relatively limited. This issue is especially evident in studies on China, which restricts the accuracy of related policy formulation. Meanwhile, when provinces are used as the unit of analysis, the sample size is relatively small; therefore, methods for measuring WEF system efficiency are mainly limited to DEA and its derivative models. (2) Scholars tend to emphasize the mechanisms through which AI affects the efficiency of individual resource systems and concentrating on the micro-technical perspective rather than adopting comprehensive analyses within the WEF Nexus framework based on larger samples, thereby restricting the empirical support at the macro level and development of interdisciplinary research. (3) In terms of research methods for the impact mechanisms of WEF system efficiency, existing studies have mainly used static and dynamic panel models to examine causal relationships and spatial econometric models to analyze spatial effects, while relatively limited attention has been given to the long-term and short-term effects of these mechanisms. Meanwhile, analyses of effect strength based on regression coefficients from econometric models remain insufficiently persuasive.

Based on these limitations, this study makes the following contributions: (1) Using the SFA model and panel data from 264 prefecture-level cities in China, WEF-TE was evaluated, thereby complementing existing studies in terms of both research scale and efficiency measurement method and providing a more refined and diversified basis for local policy-making. (2) By integrating econometric approaches with machine learning methods, the adjustment and driving mechanisms of AID on WEF-TE were empirically examined from three dimensions: enterprise scale, application level, and workforce literacy. While confirming the short- and long-term interactive relationship between AI and the WEF system, a macro interdisciplinary research framework was also established, thereby addressing a gap in the existing literature. The research framework of this article is illustrated in Figure 1.

2. Sample Selection and Data Processing

This study selected prefecture-level cities in China from 2013 to 2023 as the research sample. The specific reasons for this are as follows: (1) China possesses relatively favorable resource endowments, and its total consumption of water, energy, and food ranks first globally, making it easier to observe the causal relationships among relevant factors; (2) China’s vast territory and uneven regional development have led to differentiated resource utilization patterns and varying levels of AID [47,48,49], which facilitates heterogeneity analysis; (3) China’s WEF system has long encountered efficiency challenges due to a complex environment, imbalanced factor allocation, ineffective management, and technological lag. At the same time, local governments regard sustainable essential resource development, like water, energy, and food as the primary strategic objective. An increasing number of scholars consider AID as a key pathway to addressing China’s resource utilization challenges [50], thereby providing a theoretical basis for this study.

This study processed the data as follows: (1) samples with more than 30% missing data were removed; (2) samples lacking statistics for five or more consecutive years for key variables were excluded; (3) individual outliers were reasonably adjusted according to their historical trends; (4) individual missing values were supplemented using interpolation methods. Ultimately, a total of 2904 valid observations from 264 prefecture-level cities were obtained.

3. Method and Data Sources

3.1. Measurement of WEF-TE

3.1.1. Model Description and Variable Settings

For quantitative analysis, this study measures the WEF-TE in the study region to establish a data basis for subsequent analysis. Currently, the primary econometric methods used in academia for measuring technical efficiencies of production units include Data Envelopment Analysis (DEA) alongside Stochastic Frontier Analysis (SFA). In comparison, SFA is based on a clearly specified production function, which allows more effective identification of the influence of external factors—such as technology, management, and resource utilization—on technological deviations. It is more suitable for evaluating the technical efficiency of production systems in the public sector, enterprises, and agriculture relative to the optimal frontier [51], and corresponds well with the main sectors of the WEF system (water resources, energy industry and manufacturing, agriculture, etc.) [52]. Therefore, this study adopts SFA to measure the WEF-TE of the study area.

Table 1. WEF system technical efficiency input–output indicators.

Variable Type	Element Type	Measurement Indicators	Unit
Input variables	Water system	Total water supply (W)	108 m3
	Energy system	Total energy consumption (E)	104 tons of standard coal
	Food system	Area sown with grain crops (F1)	km2
		Total power of agricultural machinery (F2)	kWh
Output variable	Economic benefits	Regional Gross Domestic Product (GDP) (Y)	108 yuan

presents the specific input–output indicator system. Input variables are represented by the quantities of basic inputs in each subsystem; total energy consumption is calculated by converting regional coal, oil, and natural gas consumption into tons of standard coal and summing the results; and input variables for the food system are represented by the quantities of agricultural machinery and cultivated land [19,53]. Output variables are measured using regional gross domestic product (GDP), which reflects the economic benefits of the WEF system.

3.1.2. Model Construction

Construction of the SFA model first requires specifying the functional form of the production function. Considering that this study mainly focuses on regional analysis and that production technologies in the WEF system are largely general-purpose, the Cobb–Douglas function is adopted as the production function for the SFA model. In addition, to avoid heteroscedasticity caused by large variations in variable values, all variables are logarithmically transformed. Referring to relevant studies [51,54], the model is specified as follows:

$$\ln Y_{it}={\beta }_1\ln W_{it}+{\beta }_2\ln E_{it}+{\beta }_3\ln F_{1it}+{\beta }_4\ln F_{2it}+v_{it}-u_{it}$$

(1)

where Y_it represents the economic output of city i in year t; W_it, E_it, F_1it, and F_2it denote the input levels of the water, energy, and food subsystems of city i in year t, respectively; β_n represents the coefficients to be estimated; v_it denotes the random error term of city i in year t and follows a normal distribution; u_it denotes the technical inefficiency term of city i in year t, reflecting efficiency losses caused by external factors, and it follows a truncated normal distribution, while u_it and v_it are assumed to be independent. Since the model includes two unobservable variables, u_it and v_it, the maximal likelihood estimation approach is applied for predicting β_n.

Following the standard procedure for constructing an SFA model, the validity of the model is tested using the likelihood ratio (LR) test based on the sample data. The results (Chi = 4747.446, p = 0.000 < 0.05) indicate that the null hypothesis that “the quality of the model is the same whether input variables are included” is rejected, demonstrating the validity of input variables and significance of model construction.

Since the ultimate objective of model construction is to measure WEF-TE, u_it is separated from the composite error term (v_it − u_it). The formula for calculating WEF-TE is expressed as follows:

$$T{E}_{it}={\displaystyle \frac{Y_{it}}{\exp \left[f(X_{it},\beta )+v_{it}\right]}}=\exp (-u_{it})$$

(2)

where TE_it denotes the WEF-TE value of city i in year t; exp [f(X_it, β) + v_it] represents the frontier output level of technology; X_it denotes the set of input variables. Those meanings for the rest variables are consistent with those in Equation (1). Once the estimated value of u_it is obtained, TE_it can be calculated. The value of TE_it generally lies within the range [0, 1]. TE_it = 1 suggests the location of the production system on the frontier, which operates at full technical efficiency, whereas TE_it < 1 indicates the presence of efficiency losses. The closer TE_it is to 1, the higher the level of technical efficiency [55].

3.2. The Impact Mechanism of AID on WEF-TE

3.2.1. Variables Settings

(1)

Core Explanatory Variables.

The core explanatory variable in this study reflects the level of artificial intelligence development (AID). AID is a comprehensive concept that covers multiple dimensions, including technological foundations, industrial applications, policy governance, enterprise scale, international cooperation, and public literacy [56,57]. To explore the impact mechanism of AID on WEF-TE, considering data availability, the multidimensional characteristics of AID, and its compatibility with resource-based technical efficiency, this study classifies AID into three dimensions: enterprise scale (ai.e), application level (ai.a), and workforce literacy (ai.w). The detailed variable settings are described as follows:

AI enterprise scale (ai.e): The scale of AI enterprises within a region reflects its industrial activity and technological strength in the AI sector, serving as an important indicator of local AI development. Since direct data such as total output value or fixed assets of local AI enterprises are not available, following Wang et al. [58], the logarithm of AI enterprise number across an area is used to be a proxy variable for ai.e. The identification of AI enterprises follows Wang et al. [58], using Python-based web scraping to extract information from the publicly available business scopes of sample enterprises from 2013 to 2023. Enterprises whose business scopes include keywords like chips, image recognition, speech recognition, computer vision, or sensors are classified as AI enterprises.

AI application level (ai.a): The penetration rate of industrial robots reflects the extent and intensity of AI technology application within a region and is an important indicator of the progress of AID [24]. Therefore, this study adopts the industrial robot penetration rate to represent the AI application level in study regions.

Workforce AI literacy (ai.w): This variable represents the workforce’s level of understanding and interaction with AI technologies. Higher AI literacy enables the workforce to more effectively apply relevant technologies [59]. Since these characteristics are closely associated with educational attainment and attention to AI, the product of average years of education per capita and AI attention level can be used as a proxy variable and then logarithmically transformed. The AI attention level is measured using the Baidu Index by entering the keyword “artificial intelligence” to obtain the annual aggregate search index from both PC and mobile platforms in China during the study period. It should be emphasized that Baidu Index data are obtained from the internet, and their statistical results are influenced by current events and media coverage. Therefore, using the Baidu Index to measure the level of AI attention may inevitably introduce bias into the results. However, because better indicators and data for supporting municipal-level research are lacking, this indicator was retained in this study. To reduce the possible influence of this bias, ai.w was excluded from the ECM in the robustness test, and regression analysis was then performed to isolate the impact of ai.w on the overall model. The regression results including ai.w are provided for reference only.

(2)

Control Variables.

The inclusion of control variables in the present work serves two purposes: first, to enhance our constructed model robustness; and second, to provide a comparative benchmark between control variables and core explanatory variables when analyzing contribution mechanisms using the random forest model, thereby clarifying the relative importance of AID in influencing WEF-TE among various factors. The specific settings are as follows:

Technical input (tec) directly contributes to accelerating the commercialization of technologies, strengthening organizational technological capabilities, and improving resource allocation, and it is one of the key determinants affecting technical efficiency. The current work uses the logarithm of scientific and technological government expenditure as its measure.

A highly skilled workforce is better able to acquire, assimilate, and apply emerging technologies. Although the depreciation rate of human capital is accelerating in the AI era, a highly skilled workforce can adapt more rapidly to technological changes and improve technical efficiency [60]. Therefore, human capital input (hum) is selected as a control variable, measured as the logarithm of full-time equivalent R&D personnel from industrial enterprises above designated size.

Optimizing industrial structure can guide resources toward high-technology industries, eliminate outdated production capacity, and constrain energy-intensive sectors, thereby encouraging traditional industries to adopt digital and intelligent technologies, ultimately enhancing their technical efficiency [61]. Accordingly, rationalization of the industrial structure (ind) is selected as a control variable and measured by the ratio of value added by the tertiary industry to regional GDP.

Based on existing studies, the influence of government policies on resource and technical efficiency is dual in nature: on the one hand, they significantly enhance efficiency through institutional guidance, macroeconomic regulation, and market incentives [62]; on the other hand, they may also produce diminishing returns or even negative effects in certain stages or sectors [63]. Specifically, an increase in the number of policies may easily lead to conflicts among multiple objectives, resulting in duplicated or inefficient resource use and reducing marginal benefits. Furthermore, government intervention may direct resources toward politically prioritized areas rather than areas with the highest market efficiency, thereby causing resource misallocation. Therefore, government intervention (gov) is selected as a control variable and measured by the ratio of government expenditure to regional GDP.

Descriptive statistics for all the above variables are presented in

Table 2. Descriptive statistics of variables.

Variables	Observed Values	Maximum	Minimum	Mean	Standard Deviation
TE	2904	0.919	0.439	0.683	0.091
ai.e	2904	11.031	1.099	5.004	1.704
ai.a	2904	10.584	1.562	5.329	1.295
ai.w	2904	7.100	0.003	4.011	0.896
tec	2904	15.529	3.784	9.630	1.884
hum	2904	6.811	0.601	3.575	1.054
ind	2904	0.839	0.133	0.482	0.115
gov	2904	0.872	0.035	0.199	0.102

.

3.2.2. Construction of Adjustment Mechanism Analysis Model

From the perspective of economic development, a long-term equilibrium relationship may exist between technological progress and factor productivity, while short-term fluctuations in relevant variables reflect adjustments toward this long-term equilibrium [64]. In this context, error correction models have clear advantages in analyzing such relationships, including the ability to avoid spurious regression caused by non-stationary data and to simultaneously capture both long-term and short-term effects among variables [65]. Therefore, to examine whether a long-term equilibrium relationship and short-term adjustment mechanism exist between AID and WEF-TE in the study regions, this study employs an autoregressive distributed lag error correction model (ADL-ECM). Referring to relevant literature [65,66,67], the model construction process is described as follows:

Typically, before constructing an ECM, it is necessary to conduct a cointegration test on the regression data to determine whether a long-term equilibrium relationship exists. If the test results reject the null hypothesis, this indicates that the original series maintain a long-term equilibrium relationship, and the ECM can then be used to analyze short-term relationships among variables. Accordingly, this study first applies the F-test to examine the cointegration of the original series. The results (F = 212.836, p = 0.000 < 0.1) indicate that the model satisfies the cointegration requirements.

Under the cointegration condition, exploratory regression was performed on variables lagged by 1 to 4 periods according to the AIC and BIC criteria, and the optimal lag order was identified as 1. Meanwhile, the ADF test showed that all variables remained stationary after first-order difference transformation, thereby satisfying the requirements for constructing a panel model. Finally, an ECM of WEF-TE with respect to AID-related variables was constructed:

$$\begin{array}{rr}\Delta T{E}_{it}& =\left({\alpha }_0+{\alpha }_{1}T{E}_{it-1}+{\alpha }_{2}AI{D}_{it-1}+{\alpha }_3{X}_{it-1}\right)+{\beta }_1\Delta T{E}_{it-1}\hfill \\ & +{\beta }_2\Delta AI{D}_{it-1}+{\beta }_3\Delta X_{it-1}+{\mu }_{it}\hfill \end{array}$$

(3)

where Δ denotes the first-order difference of variables; TE_it represents the measured value of WEF-TE of city i in year t; AID_it−1 denotes the first-order lagged terms of variables representing different dimensions of AID (ai.e, ai.a, ai.w); X_it−1 denotes the first-order lagged terms of control variables (tec, hum, ind, gov); μ_it represents the random error term; α_n and β_n are coefficients to be estimated. Furthermore, the long-term equilibrium relationship of the model can be derived from the equation “α₀ + α₁TE_it−1 + α₂AID_it−1 + α₃X_it−1 = 0”.

3.2.3. Construction of Driving Mechanism Analysis Model

To explore the AID driving mechanism in WEF-TE in our study regions and to identify the relative contributions of core explanatory variables compared with other control variables, following relevant studies [68,69], this study adopts the mean squared error increment (IncMSE) and node purity increment (IncNode Purity) from the random forest model to represent the contribution rates of all explanatory variables to the dependent variable. Random forest serves as the machine learning method based on the bagging algorithm and feature randomization, with the primary objective of constructing multiple decision trees and averaging their prediction results. It can reduce sensitivity to outliers while avoiding multicollinearity across variables relative to conventional linear regression models; therefore, it is suitable for exploring driving mechanisms and relative contributions when multiple explanatory variables are involved [70]. Following the method of Huang et al. [68], SPSSAU software (https://spssau.net/?100000001) was used to construct a random forest model, and the specific parameters were set as shown in

Table 3. Random forest parameters [68].

Model	Parameter	Parameter Settings
Random forest	Dependent variable	TE
	Explanatory variables	ai.e; ai.a; ai.w; control variables
	Training set share	0.8
	Decision tree number	100
	Node split standard	Gini index
	Minimal sample size for node splitting	2
	Minimum value for leaf node splitting	1
	Maximal depth of a tree	No limitation
	Maximal feature count limit	Automatically setting
	If replacement sampling be permitted	YES
	If out-of-bag data be tested	YES

.

3.3. Data Sources

Business scope information of sample enterprises is obtained from China’s Tianyancha system. Data on AI attention levels are derived from China’s Baidu Index. All other data are obtained based on the China City Statistical Yearbook, China Energy Statistical Yearbook, provincial statistical yearbooks, as well as the China DataSeed database.

4. Results and Analysis

4.1. Measurements and Analysis of the WEF-TE

According to estimation from the SFA, a box plot showing the WEF-TE in the study area is presented in Figure 2. The box length is analyzed through upper and lower 25th percentiles, with a longer box indicating higher data dispersion in that year. The ends of the upper and lower whiskers represent the maximum and minimum values for that year, respectively. The line within each box corresponds to the median value of the data for each year and reflects the overall temporal trend.

As shown in Figure 2, the overall WEF-TE in the study regions exhibited a gradual upward trend from 2013 to 2023, with an average growth rate of 18% across the sample cities. This suggests that China’s policies on technological development and resource management were effectively implemented during the study period. On the one hand, guided by the principle of “innovation-driven development”, China has introduced supportive financing and tax policies for high-tech industries, reduced barriers to technological R&D, and promoted the effective transformation of emerging technologies into practical production. On the other hand, since 2015, China has steadily implemented its “capacity reduction” policy. By shutting down highly polluting, high-emission, and low-efficiency enterprises, the policy has not only reduced environmental pressure on the WEF system but also encouraged firms to actively optimize their input–output efficiency, thereby decreasing resource waste. These measures have formed a mechanism for the continuous improvement of WEF-TE. Observing the box length for each year, fluctuations remain relatively stable, indicating that improvements in China’s WEF-TE have been relatively balanced across regions. An examination of extreme values across different years reveals considerable disparities in China’s WEF-TE, with lower-end values deviating markedly from the box range. To determine whether the deviations were associated with specific regions or regional anomalies, this study counted the number of times each region fell within the lower 25th percentile across all years and mapped their spatial distribution, as shown in Figure 3. According to the statistical results, 51 prefecture-level cities exhibited low outlier values six or more times, and 18 of them remained outliers throughout the study period. In terms of spatial clustering, low-value areas formed significant contiguous clusters in the northeast, the southern coastal regions, and the upper and middle reaches of the Yellow River Basin. These findings suggest that some regions still experience weak links in basic resource conversion technologies, which require urgent attention from the relevant authorities. However, from a long-term perspective, efficiency levels in these regions have still improved notably compared with the initial year of the study period.

To further compare regional differences in the temporal evolution of WEF-TE, the average efficiency values for each year in western, central, and eastern cities, and in the Yellow River Basin, were calculated and presented in a grid-based heat map, as shown in Figure 4.

Based on Figure 3, during the study period, China’s WEF-TE exhibits a gradient pattern: eastern cities > central cities > western cities. This pattern is consistent with the regional economic and social development landscape, indicating that the level of technological optimization within the WEF system is closely associated with the overall development level of a region. Specifically, the WEF-TE in eastern and central cities was relatively similar in the initial stage of the study period, but from 2014 onward, the improvement rate in eastern cities became significantly faster than that in central cities. One possible reason is that 2014 marked the transition of China into a new economic normal. Benefiting from the long-term accumulation of high-quality production factors, eastern China was better positioned to shift development drivers from basic inputs to technological innovation, thereby forming a virtuous cycle that enhances both resource utilization and technical efficiency. In contrast, the central region serves as the main contributor to China’s WEF system supply side and inevitably bears part of the negative externalities associated with resource development. As a result, it faces greater marginal difficulty in improving technical efficiency. In addition, the growth trends in both eastern and central regions showed fluctuations around 2020. Possible explanations include the continued implementation of supply-side structural reform policies and the impact of the COVID-19 pandemic, both of which disrupted the normal operation of the WEF system to some extent. Given the larger industrial base and higher population density in the eastern and central regions, these areas were more significantly affected, leading to more pronounced fluctuations.

In the Yellow River Basin, its annual efficiency values have consistently remained below the national average, highlighting long-term challenges to sustainable development caused by inefficiencies within the basin. However, this gap has gradually narrowed since China introduced policies for high-quality development and ecological conservation in the Yellow River Basin in 2019, and by 2023, the efficiency levels had approached the national average.

4.2. Test Results and Analysis of the Adjustment Mechanism in WEF-TE

Based on the research design presented in Section 3.3, the cointegration test results and ECM regression results of WEF-TE with respect to AID-related variables are reported in

Table 4. Results of Cointegration Test and ECM regression.

Cointegration Test Results		Error Correction Model Results
Variables	Coefficient	Variables	Coefficient
ai.e	0.012 *** (7.236)	Δai.e	0.019 *** (14.929)
ai.a	0.005 *** (3.589)	Δai.a	0.016 *** (13.156)
ai.w	0.021 *** (6.901)	Δai.w	0.004 * (1.836)
Constant	0.576 *** (62.094)	Constant	−0.012 (−0.016)
Control variables	NO	Control variables	YES
Year FE	YES	Year FE	YES
City FE	YES	City FE	YES
R2	0.187	R2	0.500
Observations	2904	Observations	2904
		ECM(-1)	−0.124 *** (−13.928)

Notes: *** and * denote significance at the 1%, and 10% levels, respectively, with t-values reported in parentheses. ECM(-1) represents the lagged error term from the previous period.

.

As indicated by the cointegration test results in Table 4, AI enterprise scale (ai.e), AI application level (ai.a), and workforce AI literacy (ai.w) all exert significant positive long-term effects on WEF-TE. According to the ECM regression results in Table 4, current changes in ai.e, ai.a, and ai.w all show a positive adjustment effect on contemporaneous fluctuations in WEF-TE.

Specifically, on the one hand, AID can effectively regulate inefficiencies of WEF-TE in the short term across multiple dimensions: the sufficient AI enterprise scale can provide a more diverse range of technologies, ensuring adaptability with WEF system technologies; the broad applications of AI facilitates the integration of related technologies with the WEF system; and improvements in workforce skills further ensure that AI technologies realize their functional value in practice. On the other hand, the widespread application of AI inevitably increases data center computational demand [29]. This significantly raises regional energy consumption pressures and affects the stability and sustainability of WEF systems, thereby creating a threshold effect in technical efficiency optimization. Consequently, these opposing forces ultimately promote AID maintains a long-term equilibrium relationship with WEF-TE.

In addition, the error correction coefficient ECM(-1) reflects the degree of deviation of variables from their long-term equilibrium in the previous period. The policy interpretation of this coefficient requires comprehensive consideration of its sign, absolute value, and significance. As shown in

Table 5. Results of robustness test for ECM.

Variables	Re-Estimate the Dependent Variable Using the DEA Model	Exclude Samples (Eastern Cities)	Exclude Explanatory Variable (ai.w)	Using Terrain Slope as the Instrumental Variable
Δai.e	0.013 ** (9.362)	0.008 *** (7.330)	0.019 *** (14.678)	0.016 *** (12.154)
Δai.a	0.014 ** (15.003)	0.011 ** (9.672)	0.018 *** (14.241)	0.007 *** (7.512)
Δai.w	0.021 * (6.026)	0.019 (4.669)		0.010 ** (4.347)
ECM(-1)	−0.267 *** (−30.775)	−0.078 *** (−6.840)	−0.127 *** (−14.293)	−0.119 *** (−10.063)
Control variables	YES	YES	YES	YES
Year FE	YES	YES	YES	YES
City FE	YES	YES	YES	YES
Observations	2904	1815	2904	2904

Notes: ***, **, and * denote significance at the 1%, 5%, and 10% levels, respectively, with t-values reported in parentheses. ECM(-1) represents the lagged error term from the previous period.

, ECM(-1) is negative, indicating that WEF-TE and AID exhibit a negative feedback mechanism. When these variables deviate from equilibrium, they automatically adjust in the opposite direction to restore equilibrium. The absolute value of ECM(-1) is 0.124 and is significant at the 1% level, indicating that deviations of WEF-TE from marginal changes in AID can be corrected with an adjustment strength of 12.4% over time. In summary, although the equilibrium relationship between WEF-TE and AID shows a tendency toward self-correction, the extent of this correction is limited. Therefore, targeted policies should be implemented to ensure the quality of AID and the effective transformation of research outcomes and to guide the integration of AI technologies into the WEF system.

4.3. Roustness Test for ECM

To verify the reliability of the ECM results, robustness tests were conducted by replacing the dependent variable, excluding samples from eastern cities, and incorporating instrumental variables. The specific rationale is as follows:

First, following Sun et al. [12], we employed the DEA model to measure WEF-TE in the study regions and re-estimated the ECM.

Second, as the primary driver of China’s economic development, the eastern region has a leading effect on the practical implementation of AI technologies, which may amplify the adjustment effect of AI on WEF-TE. Therefore, eastern cities were excluded from the sample in this study, and the model was estimated again.

Third, because the indicator of ai.w is unstable, as explained in Section 3.2.1, this variable was excluded, and the ECM was re-estimated to avoid its influence on the overall results.

Fourth, based on findings of this study, WEF-TE in China is closely associated with regional development levels. More developed regions tend to attract clusters of high-tech enterprises, which introduces a reverse causal relationship and leads to potential endogeneity issues. To address this, instrumental variables and a two-stage least squares approach are employed within the ECM framework. Following Yao et al. [71], terrain slope is selected as an instrumental variable. Terrain slope reflects the flatness of regional topography; lower slopes facilitate the agglomeration of fundamental factors such as population, capital, and technology, reduce factor mobility costs, and provide favorable conditions for the development of high-technology industries such as AI, thereby satisfying the relevance requirement. At the same time, terrain slope is an objective geographical characteristic and is not influenced by technical efficiency, satisfying the exogeneity condition. The calculation method follows Yao et al. [66]. Specifically, ArcGIS 10.2 software is used to extract panels from China’s Digital Elevation Model (DEM) data using a 500-m resolution, after which slope values per unit area are calculated for each sample city. In addition, because terrain slope does not vary over time and represents cross-sectional data, to align with panel data requirements, the interaction term of each city’s terrain slope with time is used to be an instrumental variable. Terrain slope data are acquired based on the GSCloud platform developed by the Chinese Academy of Sciences. Based on the above approach, Table 5 presents results of robustness tests.

As shown in Table 5, after replacing dependent variable, excluding eastern city samples and incorporating instrumental variables, the regression results of the core variables in the ECM remain broadly consistent with those in Table 4, indicating that the empirical results of this study are relatively robust. However, after excluding eastern cities, the adjustment effect of ai.w becomes insignificant. This may be explained by two factors. First, the relatively weak human capital base or technological infrastructure in central and western regions may limit the observable effect of ai.w on WEF-TE. Second, because the measurement indicators for ai.w rely on network data, and the eastern region already has a higher level of network development, the exclusion of these samples may substantially reduce the total information contained in the indicators, thereby resulting in insignificant regression coefficients. In contrast, after ai.w was excluded, the regression results of Δai.e, Δai.a, and ECM(-1) were very close to those reported in Table 4, confirming the adjustment effect of AID on WEF-TE through enterprise scale and application level.

4.4. Assessment Results and Analyses of the Driving Mechanism in WEF-TE

In line with the research framework described in Section 3.3, a random forest model was adopted for examining the driving mechanism of AID on WEF-TE. Following standard validation procedures, the model performance was evaluated by MAE, MSE, RMSE, R², and EVS [60].

Table 6. Results of indicator tests for random forest model.

Indicators	Test Standards	Training Set	Testing Set
R2	Fit level—should be as close to 1 as possible.	0.846	0.787
MAE	The difference between the means of actual and fitted values—must approach 0 as much as possible.	0.016	0.043
MSE	Mean of squared errors—must approach 0 as much as possible.	0.000	0.003
RMSE	MSE square root—average gap value.	0.021	0.048
EVS	Measures explanatory capacity of the model for data fluctuations—must approach 0 as much as possible.	0.846	0.787

shows the evaluation criteria and corresponding results for diverse indicators. From Table 6, R² and EVS values of training and testing sets are 0.78–0.85, suggesting potent explanatory capability. A small difference in R² values was detected between two sets, indicating the absence of overfitting. In addition, MAE, MSE, and RMSE are <0.05, confirming that this model meets the required evaluation standards.

On the basis of the random forest model outcomes, Figure 5 presents the feature importance diagram for all explanatory variables of WEF-TE. Higher feature importance values indicate stronger driving effects on WEF-TE.

As illustrated in Figure 5, the driving mechanisms of different explanatory variables on WEF-TE exhibit a hierarchical structure. Specifically, the feature weight of technical input (tec) reaches 0.27, indicating a strong contribution to WEF-TE and forming the first tier independently. In addition, the feature weights of AI enterprise scale (ai.e) and AI application level (ai.a) are 0.166 and 0.15, respectively, placing them in the second tier together with human capital input (hum) and industrial rationalization (ind). The driving effects of government intervention (gov) and workforce AI literacy (ai.w) are comparatively weaker and belong to the third tier.

More specifically, investment in technology and human capital remains the dominant driver for improving the technical efficiency of China’s basic resources. At the same time, ai.e and ai.a also exert a notable driving influence and show synergies with ind, suggesting that AID in China is closely linked with the optimization of regional industrial structures. For example, the research, development, and application of AI technologies can not only directly enhance resource conversion technologies but also, by supporting industrial rationalization, guide the allocation of production factors, reduce homogeneous competition, coordinate upstream and downstream industrial chains, and minimize resource losses throughout the lifecycle, thereby improving overall efficiency. On the other hand, considering that the coefficient of hum (0.172) is much higher than that of ai.w (0.036), it can be inferred that although improvements in WEF-TE at the current stage still rely on a highly skilled workforce, the demand for AI literacy has not yet become prominent. A possible explanation is that AI-related workforce capabilities are mainly concentrated in high-level sectors such as macro decision-making and technological R&D within the WEF system, while their penetration into mid- and lower-level sectors remains limited, and thus is not fully captured in the model. This also provides an indirect explanation for why the coefficient of Δai.w becomes insignificant after excluding eastern city samples in Table 5. However, in the long term, with the increasing diffusion of AI technologies within the WEF system, the driving effect of ai.w on WEF-TE is expected to expand considerably.

5. Discussion

5.1. Comparison and Innovation

The differences between this study and previous research can be summarized in three aspects.

First, existing studies on WEF Nexus efficiency have mainly focused on the provincial level, particularly in research related to China. Provincial-level studies generally emphasize comprehensive and systematic evaluations of resource efficiency, with attention given to cross-regional coordination and strategic resource planning. In addition, relevant indicator data are more easily accessible at this scale. In contrast, prefecture-level cities were adopted as the unit of analysis in this study. Although the diversity of indicator data may be relatively limited, the findings are more detailed and precise. These findings include the spatial distribution of outliers, the identification of distinctive regions composed of different cities, such as the Yellow River Basin, and the formulation of targeted policies for specific cities. These aspects can help address the limitations of provincial-level studies.

Second, this study differs notably from existing research in terms of methodology. On the one hand, current measurements of WEF system efficiency mainly depend on DEA and its derivative models. This approach does not require a predefined production function and can support multiple output variables, allowing factors such as environmental performance and social welfare to be included in output indicators. Therefore, the information reflected in its measurement results is relatively comprehensive. However, its limitation is that random error terms cannot be separated from inefficiencies. Moreover, when the sample size is large, the results may be easily affected by outliers, leading to cases of “spurious efficiency.” In comparison, SFA can distinguish random error terms from inefficiencies, and the obtained results are more consistent with actual conditions. It should be noted that multi-indicator data are particularly difficult to obtain for city-level studies. Although SFA can only generate a single output variable and its measurement results contain less information than DEA, the use of SFA has other advantages under the inherent difficulty of data collection. Therefore, DEA is considered more suitable for WEF efficiency assessments at the provincial level, whereas SFA is more appropriate for studies at the city level or lower scales. On the other hand, existing analyses of the mechanisms affecting WEF system efficiency mainly rely on static and dynamic panel models for influencing factors and spatial econometric models for spatial effects. However, these methods generally overlook the long-term and short-term effects of these mechanisms, and the analysis of effect strength based only on regression coefficients from econometric models is not sufficiently convincing. To address these two issues, this article first used an ECM to distinguish long-term and short-term impact mechanisms and then applied machine learning techniques to estimate the contribution of each explanatory variable to WEF-TE, thereby cross-validating the results with those obtained from econometric models. The findings can provide mutual support and reference for existing research.

Third, existing econometric analyses of WEF system efficiency have mostly overlooked the influence of AID. The main reason for this is the difficulty of obtaining indicators for measuring AID. This article attempted to address this challenge by using Python 2.7 web scraping techniques and data sources such as the Baidu Index. Although the measurement results still need further optimization, the strong correlation between AID and WEF-TE was confirmed to a certain extent.

Based on the above discussion, the innovations of this study can be summarized in three main aspects:

First, considering data availability, study scale, the advantages and limitations of the model, and the suitability for this research, the SFA model was employed in this article instead of the commonly used DEA model to measure WEF-TE, thereby enriching the measurement framework of WEF efficiency research.

Second, by shifting the analytical focus of WEF system efficiency from the commonly adopted national or provincial level (especially in studies related to China) to the prefecture-level city scale, this study contributes to more precise and refined implementation of related policies.

Third, while addressing challenges in data collection, this study conducted a three-dimensional empirical analysis covering enterprise scale, application level, and workforce literacy to examine the adjustment and driving mechanisms of AID on WEF-TE. While confirming the interactive relationship between AI and the WEF system, this study also establishes an interdisciplinary research paradigm for both fields, thereby addressing a gap in existing interdisciplinary research.

5.2. Limitations and Future Research

In addition to the inherent limitations of the model discussed above, this study has two main limitations. First, because of restrictions related to research scale, data accessibility, and methodological gaps, only a relatively small group of indicators was adopted to evaluate the level of AID. Consequently, the effects of government policies, infrastructure, and other relevant factors on WEF-TE could not be examined in greater depth. Meanwhile, the selected indicators were obtained from general technological activities rather than from industry-specific indicators related to the water, energy, and food sectors, which reduces the explanatory strength of this study to a certain degree. In addition, the validity of the proxy variable for workforce AI literacy (ai.w) needs to be further confirmed in future research. In particular, the Baidu Index may be affected by the large volume of online information, substantial uncertainty, and residents’ spatiotemporal preferences. Therefore, the regression results of the ECM with and without ai.w are presented for reference and comparison. Nevertheless, the empirical findings clearly demonstrate the strong correlation and interactivity between AID and WEF-TE. Future research is expected to utilize updated data sources or more comprehensive indicator systems to further investigate the effects of AID on WEF-TE across additional dimensions.

Second, the present work merely focuses on the interaction of AID with WEF-TE. Although other key concepts in WEF Nexus theory—such as synergy, sustainability, and security—are mentioned, they are not examined in depth, indicating that substantial research gaps remain and warrant further exploration in future studies.

6. Conclusions and Implications

6.1. Conclusions

In line with panel data of 264 Chinese cities during 2013–2023, the present work first measured the WEF system technical efficiency (WEF-TE) in the study regions using Stochastic Frontier Analysis (SFA). It then examined the adjustment and driving mechanisms of artificial intelligence development (AID) on WEF-TE from three dimensions—enterprise scale, application level, and workforce literacy—using an Error Correction Model (ECM) and a random forest model. Our major conclusions can be drawn below:

(1)

Throughout our study process, China’s overall WEF-TE showed an upward trend, but significant technical disparities remained, with lagging regions exhibiting particularly severe deviations, indicating that vigilance is required to prevent potential decoupling. China’s WEF-TE demonstrates a gradient distribution pattern: eastern cities > central cities > western cities. Since China entered the “new normal” phase of economic development in 2014, the rate of improvement in eastern cities has been significantly faster than that in central cities. However, because of influence of supply-side structural reforms and the COVID-19 pandemic, growth trends in both eastern and central regions experienced fluctuations around 2020. As the core region of China’s WEF system, the Yellow River Basin has remained below the national average in terms of technological optimization, although this gap has narrowed to some extent since 2019.

(2)

AID can improve WEF-TE across multiple dimensions; however, due to constraints associated with AI-related energy consumption, a threshold effect exists in this optimization process. Specifically, this is reflected in the adjustment of WEF-TE losses, and over time, these two factors tend to reach a long-term equilibrium. This conclusion remains robust after applying various robustness tests.

(3)

Investments in technology and human capital continue to be the primary driving forces behind improvements in China’s WEF-TE. At the same time, AI enterprise scale and AI application level also play important roles, showing clear synergies with the rationalization of regional industrial structures. Although the driving effect of workforce AI literacy is not yet prominent, it has considerable potential for growth as AI adoption within the WEF system expands.

These findings highlight several challenges currently faced by China’s WEF-TE, including development planning for key regions, coordination of regional development, optimization of the net benefits of AI, deeper integration of AI with related industries, and the precision and effectiveness of policy implementation.

6.2. Policy Implications

Based on the literature and the conclusions of this study, the following policy implications are proposed:

First, to address the significant regional disparities in WEF system technical efficiency, the pronounced deviations in low-efficiency areas, and the relatively weak performance in the Yellow River Basin, several measures should be implemented. It is necessary to establish a diversified technology extension system in less developed regions. Local governments should create specialized technology extension agencies focused on the introduction and testing of key technologies, while research institutions and universities can directly support WEF system production through initiatives such as expert dispatch and field training programs. Furthermore, improving the accessibility of digital technologies is essential to achieve comprehensive network coverage, particularly in rural areas and urban fringe zones where WEF systems are concentrated, thereby overcoming technological diffusion barriers caused by geographic constraints. In addition, coordinated regional development should be emphasized: the government should use fiscal and financial tools to support technological innovation in less developed areas, while actively promoting the development of “enclave economies” to encourage advanced regions to transfer technology, capital, and management expertise, thus achieving complementarity in factor allocation. Moreover, more targeted and precise technical optimization policies should be implemented for the Yellow River Basin, including tailoring strategies to local conditions, scientifically assessing the characteristics of water, agriculture, and industry in each region, and promoting suitable technologies accordingly. This approach can leverage the strong interconnections among elements within the basin to enhance overall WEF-TE, expanding from pilot applications to broader implementation.

Second, there exists a threshold in the extent to which AID can enhance WEF-TE. In particular, large-scale model training and data center operations impose substantial demands on energy and other resources, thereby placing considerable pressure on WEF systems. Therefore, while promoting the integration of AI with WEF systems, attention should also be given to improving the energy efficiency of AI itself. Through systematic net-benefit assessments and the coordinated advancement of technological innovation and regulatory frameworks, AI can be guided to function as a sustainable driver of resource and technical efficiency improvements.

Finally, it is important to establish a rational driving framework for WEF-TE. Specifically, while ensuring sufficient investment in technology and human capital, greater attention should be paid to the interaction between AI and the industrial structure of regional WEF systems. On the one hand, AI can be utilized to integrate data across multiple stages—including exploration, development, production, transportation, and storage—to enhance intelligent and digital transformation of WEF systems, particularly within the energy sector, agriculture, and utility services (electricity, water, and gas). On the other hand, efforts should be made to cultivate a workforce with interdisciplinary expertise in both AI and related industries, thereby building a strong human capital base to support the deep integration of these fields.