Unlocking Green Growth: How Artificial Intelligence Policies Enhance Green Economic Efficiency—Evidence from China

Abstract

With growing environmental pressure and tightening resource constraints, artificial intelligence has become a key technical path for urban low-carbon transformation. This study aims to empirically examine whether and how AI-oriented pilot policies affect green economic efficiency (GEE) and identify its underlying mechanisms and boundary conditions. Taking China’s National New-Generation Artificial Intelligence Innovation Development Pilot Zone (NAIDPZ) as a quasi-natural experiment, we use a staggered difference-in-differences model to test the policy effect based on panel data of 267 Chinese prefecture-level cities from 2007 to 2023, with a series of robustness checks to ensure the reliability of the conclusion. We find that the NAIDPZ policy significantly improves urban GEE, with a stronger effect in inland, central, and non-resource-based cities. The composite NAIDPZ policy effect is associated with higher GEE, mainly through green technological innovation and industrial structure optimisation, while its impact is positively moderated by government attention and public environmental attention. These conclusions provide empirical reference for global governments to optimise artificial intelligence policies for low-carbon development.

1. Introduction

In the global context of addressing climate change, enhancing the green economy has become a core issue for achieving sustainable development [1]. According to the International Energy Agency, carbon dioxide emissions from energy sources worldwide reached 3.74 billion tons in 2023, highlighting the tension between economic growth and carbon reduction [2]. Against this backdrop, whether it is possible to effectively decrease the economic output per unit of environmental cost has emerged as a key indicator for assessing green transformation progress. Unlike traditional total factor productivity, green economy efficiency (GEE) not only measures economic output and conventional factor inputs but also integrates environmental contamination into the accounting system as an undesirable product. Its core lies in encouraging superior long-term economic growth through technological development and the optimal distribution of elements [2]. However, at present, China continues to struggle with structural path dependence due to its high emissions and energy consumption [3], as shown in Figure 1. From 2010 to 2023, China’s total carbon dioxide emissions have been increasing, significantly constraining the advancement of green economic efficiency while impeding the real green transformation process. Against this background, identifying effective policy tools that can simultaneously promote technological upgrading and green transformation has become an important academic and practical issue.

Existing studies have explored the key drivers of GEE from multiple aspects, generally classifying them into three aspects: technology [4,5], factors [6,7], and institutions [8,9,10]. However, important gaps remain in the existing literature. Most studies focus on single technological changes or conventional environmental regulatory tools, whereas the green effects of emerging digital technologies and their related policy interventions, especially the causal impact of AI-oriented policies on urban GEE, have not yet been sufficiently examined. To effectively transform AI’s technological potential into an actual catalyst for higher-quality economic growth, China has successively approved NAIDPZ policies in batches since 2019. These policies attempt to build a conducive ecosystem by conducting demonstrations of AI technologies, policy trials, and social experiments. Firstly, the NAIDPZ policies strongly encourage the adoption of green scenarios by increasing investment in public computing power, leveraging their advantages in big data processing, risk prediction, and intelligent decision-making [11]. In addition, through pioneering institutional exploration, NAIDPZ policies have increasingly demonstrated technological advantages in improving the accuracy of environmental governance and allocating resources optimally [12]. More importantly, NAIDPZ is not merely a technology-support policy but also a place-based policy experiment that combines infrastructure investment, subsidies, and institutional reforms. It represents an effective policy tool with dual attributes of institutional innovation and technological empowerment for GEE. Therefore, studying NAIDPZ not only helps evaluate the green effects of AI-oriented policy interventions but also helps explain how emerging technology policies can be translated into concrete improvements in urban green development.

The comprehensive advantages of AI have become a driving force for urban green development. Specifically, the NAIDPZ policies mainly influence GEE through two core channels: green technological innovation and industrial structure optimisation. Firstly, according to targeted technological progress theory, NAIDPZ policies, through incentive mechanisms such as research subsidies and patent protection, can break the high-carbon lock of traditional high-energy-consuming industries and guide innovation resources to converge in low-carbon fields, thereby enhancing energy utilisation efficiency [8]. Secondly, based on structural change theory, the digital transformation guided by NAIDPZ policies can expedite the movement of production inputs from inefficient to productive industries, promote the development of high-tech low-emission intelligent industries, and reduce the environmental cost per unit output through the industrial structure’s modernisation and simplification [13]. Accordingly, this study develops a unified analytical framework that links AI-oriented pilot policies, green technological innovation, industrial structure optimisation, and urban GEE to better explain the potential transmission mechanisms of the policy effect.

Therefore, this study aims to respond to the following questions: (1) Can NAIDPZ policies package significantly improve cities’ GEE? (2) Through what paths do NAIDPZ policies enhance the GEE? (3) Will the effect of the NAIDPZ policies on GEE produce heterogeneous effects due to the different characteristics of the city? To respond to the questions above, the research sample for this study involves data from 267 Chinese cities at the prefecture level between 2007 and 2023. The NAIDPZ policies are regarded as a quasi-natural experiment, and a staggered difference-in-differences (DID) model is constructed for empirical testing. The results show that the NAIDPZ policies significantly improve cities’ GEE. The heterogeneity analysis reveals that the policies’ effectiveness is stronger in central cities, non-resource-based cities, and inland cities. The mechanism test shows that the policies mainly function through two paths: encouraging green and technology innovation and optimising the industrial structure. In addition, government and public attention play a key regulatory role in enhancing the effect of the NAIDPZ policies. These findings indicate that AI-oriented policy interventions can generate not only technological effects but also broader structural and governance-related effects in the process of urban green transformation.

The possible contributions lie in three aspects: First, this study provides city-level causal evidence on the green effects of AI-oriented pilot policies, thereby extending the literature on AI policy and urban green transformation. The majority of studies have focused on the economic and environmental effects of AI [13], while few have systematically identified its green impact from this perspective. Second, by incorporating green technological innovation and industrial structure optimisation into a unified analytical framework, this paper clarifies the potential transmission mechanisms through which NAIDPZ policies may improve GEE. Although the existing literature has examined paths such as technological progress [4] and environmental regulations [14], few studies have combined shifts in industrial organisation and the development of technological bias for methodical analysis. Third, this paper further extends the analysis to the governance context and shows that the effectiveness of AI policies depends on formal and informal governance conditions. In this sense, the paper not only identifies whether AI-oriented pilot policies work but also explains why they work and under what conditions they are more effective.

2. Literature Review

The extant literature on GEE primarily examines its underlying drivers. As a combined measure that captures both the standard of economic expansion and the extent of alignment between economic activity and environmental sustainability, GEE’s evolution is shaped by the interplay of multiple mutually reinforcing factors—not by any single determinant. The scholarly consensus identifies three interdependent dimensions as central to GEE advancement: technology-driven innovation, resource reallocation and structural optimisation, and institutional governance [15].

First, technological progress constitutes a pivotal driver of sustained GEE improvement. Grounded in endogenous growth theory and directed technical change frameworks, empirical studies demonstrate that the diffusion of clean and energy-efficient technologies significantly reduces pollution intensity while preserving output growth [4,5]. Second, according to the structural dividends hypothesis, GEE gains arise not merely from intra-sectoral efficiency improvements but crucially from the allocative efficiency of production factors across sectors. Furthermore, industrial structure rationalisation and upgrading serve both as foundational prerequisites for industrial chain modernisation and as strategic levers for advancing green productivity [16]. Third, while much of the literature centres on environmental regulation, the findings reveal nonlinear effects: command-and-control instruments and market-based incentives exhibit heterogeneous impacts on GEE, contingent upon local implementation rigor and administrative capacity [17]. Complementing formal rules and informal institutions elevates firms’ noncompliance costs, thereby fostering region-level improvements in green economic performance [18].

Another category of literature related to this is the discussion of the effectiveness of artificial intelligence policies. The existing studies mainly focus on two aspects: macro-regional development and micro-enterprise behaviour. At the macro level of regional development, AI policies have significantly enhanced regional innovation efficiency through institutional innovation and factor agglomeration [19,20]. Additionally, AI policies have demonstrated a significant green transformation effect [21], encouraging the growth of green finance and strengthening the resilience of industrial chains [22]. At the enterprise behaviour micro level, extensive literature confirms the empowering effects of AI policies on various aspects of enterprises, including innovation, productivity, and ESG performance [23]. Specifically, AI policies have encouraged companies to invest more in R&D and promoted the emergence of high-quality, collaborative, and disruptive innovations by opening up public computing power, providing application scenarios, and establishing special funds [24]. Although the above-mentioned literature has extensively explored the direct economic output and environmental impact of AI policies, systematic studies that integrate the specialised field of GEE with the quasi-natural experiment of NAIDPZ policies to investigate causal links are limited.

In addition, recent research from a global perspective has also begun to highlight the broader relationship between AI and sustainable development. For example, Gohr et al. [25] show that AI has significant potential to advance sustainable development goal-related research internationally, while also emphasising that its full promise depends on combining advanced AI applications with deep sustainability expertise. Costa et al. [26] further characterise AI as a general-purpose technology capable of driving cross-sectoral and transnational sustainable transformation, and Mancuso et al. [27] stress that AI innovation may generate both sustainable value creation and value destruction, depending on how it is governed and implemented in different regions. Relatedly, Zhou et al. [28] demonstrate that AI can reinforce cleaner production when coupled with green finance and ESG-oriented strategies, suggesting that AI’s green effects are often conditional on complementary institutional and policy arrangements. These studies help place the present study within a broader discussion on whether AI contributes to sustainability primarily through technological efficiency, institutional coordination, or their interaction.

Nevertheless, three critical research gaps still exist. First, current studies on GEE predominantly examine isolated technological innovations or conventional environmental regulatory instruments. While prior work has extensively analysed how technological progress, factor reallocation, and environmental regulation influence green development, it largely overlooks the distinct causal impact of NAIDPZ policies. Second, the literature on transmission pathways underlying green development remains fragmented, lacking an integrated framework that jointly accounts for technological bias evolution and industrial structural transformation. Although existing studies emphasise efficiency gains via technological upgrading, they seldom provide rigorous theoretical articulation or empirical validation of the dual-channel mechanism. Third, although the importance of institutional context is widely acknowledged, few studies jointly incorporate formal and informal institutions into a cohesive analytical framework to evaluate how divergent governance settings shape the green-enabling efficacy of NAIDPZ policies.

3. Theoretical Analysis and Research Hypothesis

3.1. Policy Background

As the continuous technology revolution and industrial transformation intensify, AI has gradually become the key general technology driving the rapid growth of the economy. The Chinese government highly values AI advancement, and it has incorporated it into the national innovation-driven development strategy system, continuously improving the institutional setup and top-level design. In 2015, the Chinese government clearly proposed supporting the development of AI, establishing the governmental framework for relevant industry applications and technological research. In 2017, the State Council released the “New Generation Artificial Intelligence Development Plan”. Under the guidance of the national strategy, NAIDPZ policies and artificial intelligence clusters have been successively established, gradually building an implementation mechanism centred on pilot demonstrations and experience promotion.

In 2019, the “Construction Work Guidelines for National New Generation Artificial Intelligence Innovation Pilot Zones” were published by the Ministry of Science and Technology, clearly stating that through policy pilot and technology demonstration coordination, it could explore replicable and promotable development models to encourage the widespread use of artificial intelligence in key fields. The period from 2020 to 2023 marked a key stage in China’s AI policies, shifting from top-level design to practical implementation, concentrating on encouraging the close connection between technology and the real economy through pilot zones and simultaneously establishing a standardisation system. In 2025, the State Council issued the “Opinions on Deeply Implementing the AI + Action, systematically outlining the application paths for integrating AI across different fields. The specific development of the AI policy is reflected in Figure 2.

It should be noted that AI policy differs conceptually from both traditional environmental regulation and digital infrastructure policy. Traditional environmental regulation mainly operates through compliance pressure, pollution control standards, and cost internalisation, thereby inducing firms to reduce emissions or adopt cleaner production methods. By contrast, AI policy is not primarily designed to impose environmental constraints but to enhance technological capability, application efficiency, and innovation incentives. It is also distinct from digital infrastructure policy. While digital infrastructure policy mainly focuses on improving connectivity, data transmission, and broadband access, AI policy places greater emphasis on algorithmic application, intelligent scenario demonstration, cross-sector integration, and institutional experimentation. In this sense, NAIDPZ is not merely an infrastructure programme or a regulatory arrangement but a composite policy intervention that combines technological empowerment and institutional coordination [29]. This dual attribute helps explain why AI-oriented pilot programmes may influence urban green transformation through both innovation-related and structural upgrading channels.

3.2. Research Hypotheses

3.2.1. The Direct Effect of NAIDPZ Policies

NAIDPZ policies are not merely a technical support measure; instead, their effect identified in this paper should be understood as the outcome of a composite intervention that combines technological support, application incentives, and institutional coordination. They reshape the institutional environment for technological application, exerting a guiding influence on resource allocation and the direction of technological progress.

Based on institutional change theory, the coordinated evolution of institutional supply and technological progress is the inherent mechanism driving the improvement of the economic development mode. On this basis, NAIDPZ policies enhance GEE through the collaborative effects of the supply and demand sides. Figure 3 presents the framework illustrating the impact of NAIDPZ policies on GEE in urban areas.

On the supply side, NAIDPZ policies improve the supply of production factors, lower the entry threshold for green transformation, and, through optimised factor allocation, enhance a city’s green productivity [30]. Specifically, the government alleviates the factor constraints of enterprises in terms of computing power acquisition, technological experimentation and R&D investment through research subsidies, public computing power platform construction in intelligent infrastructure, and facilitating AI adoption in the green industry, thereby reducing the capital costs and technical uncertainties faced by enterprises in implementing intelligent transformation and green technology innovation [31,32]. This micro-level cost reduction can generate significant economies of scale effects. The extensive penetration of intelligent factors enhances the city’s overall energy utilisation efficiency and total factor productivity, enabling it to achieve higher economic output with lower resource and environmental costs [33,34].

On the demand side, NAIDPZ policies have expanded the effective market demand for green technologies, stabilising the expected returns from their application, thereby accelerating their diffusion and penetration at the urban level. Relying on government green procurement, open application scenarios and demonstration project construction, NAIDPZ policies have provided clear application spaces for artificial intelligence in areas such as pollution control, energy conservation, etc., reducing the market uncertainty during the initial stage of new technology promotion and encouraging the transition from potential to actual demand for green technologies [35]. This demand-driven mechanism not only prompts individual enterprises to update equipment and restructure processes but also guides urban industrial resources to shift from high-energy-consuming sectors to high-efficiency, intelligent, and green sectors, promoting the synchronisation of the urban consumption structure and production mode [7]. Therefore, we propose the following:

H1: 

NAIDPZ policies can improve the urban GEE.

3.2.2. The Indirect Effect of NAIDPZ Policies

According to targeted technological progress theory, the direction of technological innovation is deeply influenced by the institutional environment and incentive mechanisms. First, the NAIDPZ policies change the incentive structure, break the high-carbon lock-in, and guide technological progress towards the green side [36]. In a market environment without external intervention, constrained by sunk costs and scale effects, enterprises tend to pursue incremental innovation along the current technological pathways that have significant levels of energy use and pollution [37]. Through differentiated incentive measures such as fiscal subsidies, green procurement, and special funds, artificial intelligence policies have changed the relative returns between clean technologies and polluting technologies. This institutional arrangement effectively reversed the excessive concentration of R&D resources in traditional high-carbon sectors, promoting technological progress toward resource conservation and environmentally friendly technologies [38]. Through this biased technological progress, urban economies can reduce their excessive reliance on energy sources, thereby lowering the marginal generation rate of undesirable outputs.

Another essential way that NAIDPZ policies affect the GEE is by optimising the industrial structure. Based on structural change theory, the leap in economic efficiency fundamentally stems from the continuous flow of factors of production from inefficient to efficient sectors. First, NAIDPZ policies, through mechanisms of factor price incentives and market guidance, alleviate the misallocation of resources in traditional industries, promoting the aggregation of labour, capital, and data factors to intelligent industries with low energy intensity, high added value, and high technological density [39]. In addition, traditional industries like manufacturing, energy, and transportation are deeply linked to NAIDPZ policies. By effectively cooperating both upstream and downstream of the production chain, they complete the intelligent transformation of traditional high-carbon industries, reducing production process losses and improving overall operational efficiency [40]. Moreover, NAIDPZ policies have also given rise to new business forms with inherently low-carbon attributes, such as virtual power plants and intelligent recycling. Their expansion directly increases the proportion of the green economy in cities, thereby reducing environmental pressure at the source [41]. Therefore, we propose the following:

H2: 

By encouraging green technology innovation and optimising the industrial structure, NAIDPZ policies improve urban GEE.

3.2.3. The Moderation Effect of NAIDPZ Policies

The government’s attention and the public’s environmental concern constitute the formal institutional execution environment and the informal institutional supervision environment, which jointly regulate the extent to which NAIDPZ policies promote urban GEE.

The government’s attention reflects the priority of policy goals within the public governance system. According to the attention allocation theory, government’s attention directly determines the level of resource mobilisation. On the one hand, in a vertical governance system, a high level of government attention can be translated into strong administrative capacity and resource guarantees [42]. When local governments attach high importance to artificial intelligence and green development, they often introduce clear administrative instructions, special funds support, and cross-departmental coordination mechanisms. This effectively reduces information friction and implementation resistance in the policy transmission process, ensuring that the land, funds, and human resources required for the construction of AI infrastructure and the green technology transformation are prioritised [43]. On the other hand, government attention shapes market expectations in urban areas through signalling effects [44]. Including artificial intelligence and green development in the government’s core agenda sends a strong policy commitment signal and institutional endorsement to the market. This not only guides social capital and innovative elements to cluster in the green intelligent industry within the city and alleviates the external financing constraints of regional green transformation but also optimises the factor allocation structure of the city and encourages a general increase in the city’s green economic efficiency [11].

Compared with the enhancement of formal institutional efficacy through government attention, public environmental attention mainly exerts external constraints and amplification effects on informal institutions. According to the stakeholder theory and the legitimacy mechanism, first, public attention strengthens the rigidity of policy implementation through the social supervision mechanism [45]. High-intensity public attention to environmental issues implies more sensitive public oversight and more frequent media exposure, which puts significant legitimacy pressure on local governments and enterprises. This bottom-up social accountability mechanism forces polluting entities to actively use artificial intelligence to achieve precise pollution control and energy efficiency management, thereby transforming external policy pressure into an endogenous driver of improving green economy efficiency [4]. Secondly, public attention guides the direction of technological application through the market demand mechanism [46]. Green consumption demands force enterprises to carry out intelligent and green transformation at the production end. Public preference for green products and services has increased the economic return on investment in the green field, prompting enterprises to pursue economic benefits while also considering environmental performance [47]. Therefore, we propose the following:

H3: 

The impact of NAIDPZ policies on urban GEE is positively moderated by government attention and public concern.

4. Research Design

4.1. Model Setting

4.1.1. Baseline Regression Model

To investigate how artificial intelligence policies affect the effectiveness of the urban green economy, a two-way fixed effects model is set up as follows:

$${\mathrm{G}\mathrm{E}\mathrm{E}}_{\mathrm{i}\mathrm{t}}={\mathsf{\beta }}_0+{\mathsf{\beta }}_1{\mathrm{N}\mathrm{A}\mathrm{I}\mathrm{D}\mathrm{P}\mathrm{Z}}_{\mathrm{i}\mathrm{t}}+\mathsf{\phi }{\mathrm{Z}}_{\mathrm{i}\mathrm{t}}+u_i+{\lambda }_t+{\epsilon }_{it}$$

(1)

Among them, ${\mathrm{G}\mathrm{E}\mathrm{E}}_{\mathrm{i}\mathrm{t}}$ in the model represents green economic efficiency, ${\mathrm{N}\mathrm{A}\mathrm{I}\mathrm{D}\mathrm{P}\mathrm{Z}}_{\mathrm{i}\mathrm{t}}$ represents the dummy variable for the artificial intelligence policy. The coefficient ${\mathsf{\beta }}_1$ is the estimation quantity this paper focuses on, reflecting the net driving effect of implementing AI policies on GEE. ${\mathrm{Z}}_{\mathrm{i}\mathrm{t}}$ represents a series of control variables that could affect the GEE; the year fixed effect is represented by ${\mathsf{\lambda }}_{\mathrm{t}}$, the city fixed effect is represented by ${\mathrm{u}}_{\mathrm{i}}$, and the random error term is represented by ${\mathsf{\epsilon }}_{\mathrm{i}\mathrm{t}}$.

4.1.2. Mediating Effect Model

In the section on theoretical analysis, we propose that NAIDPZ policies would enhance urban GEE by encouraging the development of green technologies and optimisation of industrial structures. To test the above mechanism, this study constructs the following econometric model:

$${\mathrm{M}\mathrm{e}\mathrm{d}}_{\mathrm{i}\mathrm{t}}={\mathsf{\alpha }}_0+{\mathsf{\alpha }}_1{\mathrm{N}\mathrm{A}\mathrm{I}\mathrm{D}\mathrm{P}\mathrm{Z}}_{\mathrm{i}\mathrm{t}}+{\mathsf{\phi }\mathrm{Z}}_{\mathrm{i}\mathrm{t}}+u_i+{\lambda }_t+{\epsilon }_{it}$$

(2)

In the formula, ${\mathrm{M}\mathrm{e}\mathrm{d}}_{\mathrm{i}\mathrm{t}}$ represents the mechanism variable, explicitly referring to green technological innovation and industrial structure optimisation. The definitions of the other variables are identical to those in the baseline model.

4.1.3. Moderation Effect Model

To evaluate the moderating impacts of public and governmental environmental attention on the GEE and the adoption of artificial intelligence policies, the following econometric model is constructed:

$${\mathrm{G}\mathrm{E}\mathrm{E}}_{\mathrm{i}\mathrm{t}}={\mathsf{\gamma }}_0+{\mathsf{\gamma }}_1{\mathrm{N}\mathrm{A}\mathrm{I}\mathrm{D}\mathrm{P}\mathrm{Z}}_{\mathrm{i}\mathrm{t}}+{\mathsf{\gamma }}_2{\mathrm{M}\mathrm{o}\mathrm{d}}_{\mathrm{i}\mathrm{t}}+{\mathsf{\gamma }}_3{\mathrm{N}\mathrm{A}\mathrm{I}\mathrm{D}\mathrm{P}\mathrm{Z}}_{\mathrm{i}\mathrm{t}}\times {\mathrm{M}\mathrm{o}\mathrm{d}}_{\mathrm{i}\mathrm{t}}{+\mathsf{\phi }\mathrm{Z}}_{\mathrm{i}\mathrm{t}}+{\mathrm{u}}_{\mathrm{i}}+{\mathsf{\lambda }}_{\mathrm{t}}+{\mathsf{\epsilon }}_{\mathrm{i}\mathrm{t}}$$

(3)

Among them, ${\mathrm{M}\mathrm{o}\mathrm{d}}_{\mathrm{i}\mathrm{t}}$ serves as the moderating variable, representing government attention (Gover) and public environmental attention (Publi), respectively. The coefficient ${\mathsf{\gamma }}_3$ of the interaction term ${\mathrm{N}\mathrm{A}\mathrm{I}\mathrm{D}\mathrm{P}\mathrm{Z}}_{\mathrm{i}\mathrm{t}}\times {\mathrm{M}\mathrm{o}\mathrm{d}}_{\mathrm{i}\mathrm{t}}$ is the focus of this study. A significantly positive value of ${\mathsf{\gamma }}_3$ suggests that this moderating variable can strengthen the beneficial influence; conversely, it implies a weakening effect of AI policies on the GEE. The remaining variables’ definitions coordinate with those of the baseline model.

4.2. Variable Definition

4.2.1. Dependent Variable

This paper’s dependent variable is green economic efficiency. Its measurement necessitates extensive and well-coordinated improvement in the economy, environment, and society, concentrating on aggressively promoting energy conservation and emission reduction as well as divorcing economic growth from excessive consumption and pollution. Following [31], this paper comprehensively measures the above input–output variables using the super-efficiency SBM model for non-expected outputs, and the final result is expressed as green economic efficiency.

Table 1. System for measuring green economic efficiency.

Category	Variable	Index
Input variable	Capital	Fixed capital stock
	Labour	Number of employees
	Energy	Total energy consumption
Expected output variables	Total economic output	Real GDP
Non-expected output variable	Environmental pollution emissions	Sulphur dioxide in industry
		Emissions of industrial smoke and dust
		Discharge of industrial wastewater

depicts the indicator system.

4.2.2. Explanatory Variable

NAIDPZ is the explanatory variable, measured by whether the city has been approved as a pilot zone. Accordingly, the explanatory variable captures the implementation of an AI-oriented pilot policy package rather than the pure development level of AI technology itself, based on the list of national new-generation artificial intelligence innovation development pilot zones announced by the Ministry of Science and Technology. If the city i is approved as a pilot zone in a given year, its value would be 1 in that year and the following years, or it would be 0.

4.2.3. Control Variable

To guarantee that the empirical findings are accurate, this paper follows the approaches of Lin and Zhou [48] and Zhou et al. [49], and selects a series of control variables closely related to urban GEE to exclude additional influences from the benchmark regression results: environmental regulation (Er), urbanisation level (Urb), marketisation level (Mar), economic development level (Pgdp), government intervention (Gov), population size (Popd), and financial development level (Fin). See

Table 2. Variable definitions and measurement method.

Category	Variable	Symbol	Measurement
Dependent variable	Green economic efficiency	GEE	Super-efficiency SBM model.
Explanatory variable	National new-generation artificial intelligence innovation development pilot zone	NAIDPZ	In 2019 and subsequent years and in the pilot areas, the value is 1; otherwise, it is 0.
Mediating variable	Green technological innovation	Techn	The total number of city-level patent authorisations.
		Green	Number of green patent grants.
	Industrial structure optimisation	Advan	Proportion between the tertiary and secondary industries’ added values.
		Justi	Measured by the Theil index.
Moderation variable	Government attention	Gover	The ratio of total keyword frequency to the total word count of the report.
	Public environmental attention	Publi	The natural logarithm of the Baidu annual average haze search index.
Control variable	Environmental regulation	Er	The rate at which industrial solid waste is used overall.
	Urbanisation level	Urb	The proportion of the population living in permanent urban areas to the overall population.
	Marketisation level	Mar	The private economic development index.
	Economic development level	Pgdp	The natural logarithm of per capita GDP.
	Government intervention	Gov	The ratio of local general public budget expenditure to regional GDP.
	Population size	Popd	Year-end permanent population divided by the administrative area.
	Financial development level	Fin	The proportion between the year-end loan balance of financial institutions and regional GDP.

for details.

4.2.4. Mediating Variable

(1) Green technological innovation (Paten): Following Liu and Dong [50], this paper measures green technological innovation from two aspects: green innovation (Green) and technological innovation (Techn). Among them, the technological innovation level reflects the activity level and output scale of regional technological research and development, which is measured by counting all of the patent authorisations at the city level; green innovation level reflects the structural bias of innovation activities towards the domains of the environment, which is measured by the quantity of green patent authorisations.

(2) Industrial structure optimisation (Indus): Following Tian and Zhang [51], this study measures industrial structure optimisation from two dimensions: advancedization (Advan) and rationalisation (Justi) of the industrial structure. Specifically, advancedization of the industrial structure represents the process of a structure evolving to a higher level, which is measured by the proportion between the tertiary and secondary industries’ added values; rationalisation of the industrial structure mainly represents the degree of coordinated development across different industries, as determined by the Theil index.

4.2.5. Moderation Variable

(1) Government Attention (Gover): Following Du et al. [52], we use the city government work reports as the data source, after preprocessing the text, such as ecological protection, pollution control, and green development, and the frequency of keyword occurrences is counted. The logarithm of the ratio of the total frequency of keywords to the total number of words in the report is used to obtain this indicator. The higher the value, the more environmental issues are discussed in the government’s policy agenda, and the closer the government’s attention.

(2) Public Environmental Attention (Publi): Following the approach of Yu and Jin [53], the logarithm of the annual average search index for smog on Baidu is used to evaluate public environmental attention. Specifically, by searching for smog as the keyword in the Baidu Index, the average daily search volume for smog across cities is obtained, and the annual average smog search index for the region is calculated.

4.3. Data Source

In this study, the prefecture level is selected as the research sample for the years 2007 through 2023. The list of newly approved AI policies is mainly based on the “Construction Work Guidelines for National New-Generation Artificial Intelligence Innovation Development Pilot Zones” issued by the Ministry of Science and Technology in 2019. Authoritative databases such as the China Urban Statistical Yearbook are the source of the city-level characteristic variables. For the small amounts of missing data in some cities, a combination of linear interpolation and the average value of adjacent years is used to supplement. Every continuous control variable is truncated using the 1% quantile to prevent outliers’ impact on regression findings.

Table 3. Descriptive statistics.

	(1)	(2)	(3)	(4)	(5)
Variables	N	Mean	Sd	Min	Max
NAIDPZ	4539	0.0123	0.110	0.000	1.000
GEE	4539	0.348	0.155	0.081	1.213
Er	4539	0.975	0.498	0.168	14.410
Urb	4539	0.534	0.166	0.115	1.000
Mar	4539	15.400	1.156	5.472	19.010
Pgdp	4539	10.550	0.705	8.138	12.460
Gov	4539	0.181	0.094	0.043	1.485
Popd	4539	5.910	0.677	2.868	8.136
Fin	4539	2.367	1.145	0.504	7.621

shows the descriptive statistics.

5. Empirical Results

5.1. Baseline Regression Results

Table 4. Baseline regression.

Variables	(1)	(2)	(3)	(4)	(5)	(6)
	GEE	GEE	GEE	GEE	GEE	GEE
NAIDPZ	0.1513 **	0.2068 ***	0.1408 ***	0.1289 **	0.1493 ***	0.1322 ***
	(0.058)	(0.035)	(0.036)	(0.052)	(0.035)	(0.034)
Er				−0.0016	0.0004	−0.0008
				(0.005)	(0.004)	(0.003)
Urb				−0.0330	−0.0422	−0.0810
				(0.057)	(0.058)	(0.052)
Mar				0.0482 ***	−0.0008	−0.0148
				(0.017)	(0.007)	(0.009)
Pgdp				−0.0077	0.0438 ***	−0.1096 ***
				(0.023)	(0.013)	(0.025)
Gov				0.0607	−0.0294	−0.0832
				(0.093)	(0.071)	(0.063)
Popd				−0.0104	0.0317	−0.0545
				(0.021)	(0.073)	(0.081)
Fin				−0.0283 ***	0.0325 ***	−0.0085
				(0.006)	(0.007)	(0.011)
Constant	0.3466 ***	0.3459 ***	0.3467 ***	−0.1771	−0.3401	2.1317 ***
	(0.006)	(0.000)	(0.000)	(0.209)	(0.404)	(0.590)
Year FE	√	×	√	√	×	√
City FE	×	√	√	×	√	√
N	4539	4539	4539	4539	4539	4539
${\mathrm{R}}^2$	0.115	0.500	0.594	0.182	0.542	0.609

Notes: *** and **, denote statistical significance levels of 1% and 5%, respectively. Additionally, standard errors have been clustered at the city level and are presented in parentheses. √ indicates that the corresponding fixed effect is controlled for in the regression model, while × indicates that it is not.

presents the baseline regression results of NAIDPZ policies’ effect on the urban GEE. Among them, columns (1) to (3) do not include control variables, respectively controlling for the fixed-year effect, the fixed-city effect, and simultaneously controlling both; columns (4) to (6) add control variables on this basis, also successively controlling the fixed-year effect, the fixed-city effect, and the dual fixed effect. The regression results show that, regardless of whether control variables are included, the coefficient on NAIDPZ is significantly positive. From an economic perspective, assuming that all other variables remain constant, cities approved as pilot zones have, on average, green economic efficiency approximately 13.22% higher than non-pilot cities. This empirical result preliminarily verifies H1, indicating that NAIDPZ policies serve as the institutional guarantee for achieving high-quality urban green transformation.

A more precise interpretation is that the NAIDPZ policy package improves green economic efficiency through a combination of technological support, scenario expansion, and institutional coordination on both the supply and demand sides. From the supply side, NAIDPZ policies reduce the cost of technology adoption through research and development subsidies and infrastructure investment and encourage the transfer of production factors to intelligent manufacturing that uses less energy and is more efficient. From the demand side, green government procurement and intelligent application scenario pilots create stable market expectations and incentivise enterprises to pursue innovations in energy conservation and emission reduction, while accelerating the digital intelligence of their production processes [54].

5.2. Robust Tests

5.2.1. Parallel Trend Test

The validity of the DID approach depends on the parallel trend assumption, which states that before the national authorities officially approved the establishment of the NAIDPZ policies, the trends in changes in green economic efficiency for the treatment and control groups should not differ significantly. Therefore, this study conducts a parallel trend test using the model as follows:

$${\mathrm{G}\mathrm{E}\mathrm{E}}_{\mathrm{i}\mathrm{t}}={\mathsf{\delta }}_0+\sum _{\mathrm{k}=-7,\mathrm{k}\ne -1}^4{\mathsf{\delta }}_{\mathrm{k}}{\mathrm{A}\mathrm{I}}_{\mathrm{i}\mathrm{t}}+\mathsf{\phi }{\mathrm{Z}}_{\mathrm{i}\mathrm{t}}+u_i+{\lambda }_t+{\epsilon }_{it}$$

(4)

The dynamic changes in the policy effects of the NAIDPZ policies can be investigated by contrasting the statistical and economic importance of the parameters in equation (4). Figure 4 shows the test results. Before the official approval of NAIDPZ policies, there was no apparent difference between the treatment and control groups’ green economic efficiency. As NAIDPZ policies were successively approved, their beneficial impact on enhancing green economic efficiency persisted.

5.2.2. Placebo Test

In order to confirm the robustness of the benchmark regression results, this study, based on the country’s approval status for NAIDPZ policies, randomly generated multiple virtual treatment groups and conducted the same regression analysis as in model (1). Specifically, by conducting 500 random samplings of the NAIDPZ policy variable, the GEE regression coefficient’s p-value distribution and kernel density plot are produced. As shown in Figure 5, when using randomly generated virtual policy variables, the estimated coefficients mostly concentrated around zero and were far from the benchmark regression coefficients. Therefore, it can be inferred that the enhancement in GEE is indeed attributable to the NAIDPZ policies’ implementation.

5.2.3. Exclude the Interference of Other Policies

To prevent the concurrent implementation of other policies from interfering with the urban green economic efficiency and thereby affecting the reliability of benchmark results, this study further controlled for “Broadband China” pilot city policies implemented during the sample period. To determine the independent impact of the NAIDPZ policies, the “Broadband China” pilot program was specifically added as a control variable into the benchmark regression model.

Table 5. Robustness test.

Variables	Excluding Other Policy Interference	PSM-DID Before Matching	PSM-DID After Matching
	(1)	(2)	(3)
NAIDPZ	0.1284 ***	0.1322 ***	0.0216 ***
	(0.034)	(0.034)	(0.006)
Constant	2.1223 ***	2.1317 ***	−0.3269 **
	(0.582)	(0.590)	(0.554)
Controls	√	√	√
Year FE	√	√	√
City FE	√	√	√
N	4539	4539	2615
${\mathrm{R}}^2$	0.611	0.609	0.761

Notes: *** and ** denote statistical significance levels of 1% and 5%, respectively. Additionally, standard errors have been clustered at the city level and are presented in parentheses. √ indicates that the corresponding control variable or fixed effect is controlled for in the regression model.

’s column (1) regression results demonstrate that after adding this policy variable, the NAIDPZ policies’ estimated coefficient remained significant. It suggests that the promotion impact of NAIDPZ policies on urban GEE still remains robust even after adjusting for the possible influence of the “Broadband China” pilot city policies.

5.2.4. PSM-DID Test

To effectively mitigate potential interference from sample selection bias on the estimation results, this paper further employs the PSM-DID method for regression analysis. This test is particularly necessary, because the designation of NAIDPZ pilot cities is not random. Cities selected as pilot zones are more likely to have stronger innovation capacity, higher levels of economic development, and better institutional conditions than non-pilot cities. Therefore, even if the baseline DID model controls for observable covariates and fixed effects, the estimation may still be affected by selection bias arising from pre-existing differences between pilot and non-pilot cities. Although PSM-DID cannot eliminate such bias, it can improve the sample comparability and substantially alleviate this concern.

Specifically, first, based on the Logit model, the propensity score of each city becoming an artificial intelligence experimental zone during the data period is estimated using all control variables as matching covariates. Secondly, based on the propensity score results, a 1:1 nearest neighbour matching method is adopted to match cities in the artificial intelligence experimental zones with cities in the non-experimental zones that have similar characteristics, thereby constructing comparable samples. Finally, after completing the propensity score matching, the matched samples are used to re-perform in equation (1), and the results are examined and contrasted with the baseline regression.

Figure 6 shows the standardised deviation changes of each characteristic variable before and after matching for the experimental and the control group. The findings show that the two groups of cities differed significantly in the majority of variables before matching. After matching, the standardised deviations of each variable significantly decreased and approached 0, indicating that the PSM method effectively improved the comparability between the samples and well controlled the differences in individual characteristics. Table 5 columns (2) and (3) further report the DID estimation results before and after PSM-DID matching. It can be found that, in the matched samples, the NAIDPZ policy still had a significant positive effect on urban GEE, confirming that the main findings are not solely driven by observable differences in city characteristics before policy implementation. Therefore, although PSM-DID cannot fully remove all sources of selection bias, it substantially alleviates the concern that non-random pilot-city selection biases the baseline results.

5.2.5. Endogeneity Test

This study also uses the instrumental variable (IV) method for identification to address possible endogeneity problems, such as omitted-variable bias and reverse causation. It uses the two-stage least squares method to perform an endogeneity test. Following Tang et al. [55], this paper constructs the instrumental variable as the interaction between the number of fixed-line telephones per 100 residents in a city in 1984 and the contemporaneous internet penetration rate. This IV is relevant, because historical telecommunications infrastructure shaped the local foundation for subsequent digital development and thus affected cities’ capacity to adopt AI-related technologies and policies, a logic consistent with related studies such as Ren et al. [56], Xing et al. [57], and Xu et al. [58]. At the same time, the number of fixed-line telephones in 1984 captures historical communication conditions rather than current green development performance. After controlling for city and year fixed effects as well as observable covariates, it is less likely to affect current urban green economic efficiency directly, except through later digital development and AI-related policy adoption. Nevertheless, this strategy cannot completely rule out concerns related to long-term development trajectories. Therefore, the IV results are treated as supplementary evidence supporting the causal interpretation of the baseline findings.

The regression results are shown in

Table 6. Results of the endogeneity test.

Variables	(1)	(2)
	NAIDPZ	GEE
IV	0.0011 ***
	(0.000)
NAIDPZ		0.5039 **
		(0.215)
Controls	√	√
Year FE	√	√
City FE	√	√
$\mathrm{N}$	4505	4505
Kleibergen–Paap rk Wald F	21.195[16.38]
Kleibergen–Paap rk LM	22.202 ***

Notes: *** and **, denote statistical significance levels of 1% and 5%, respectively. Additionally, standard errors have been clustered at the city level and are presented in parentheses. √ indicates that the corresponding control variable or fixed effect is controlled for in the regression model.

. The Kleibergen–Paap rk Wald F statistics are significantly higher than the critical value of 16.38, and the Kleibergen–Paap rk LM statistics pass the test at the 1% significance level, demonstrating that the issues of unidentifiability and weak instrumental variables are not present in the chosen instrumental variables. The IV and the policy variables in the artificial intelligence experimental area have a significant positive correlation, according to the first-stage regression results; the second-stage regression also shows that the estimated coefficient for the NAIDPZ policies remains positive. These findings demonstrate the robustness of the NAIDPZ policies’ promotion effect on urban GEE, which supports the validity of the baseline regression even after effectively controlling for potential endogeneity.

5.2.6. CSDID Test

Since the NAIDPZ policies were implemented in a staggered manner across cities, the conventional two-way fixed effects estimator may suffer from treatment-timing bias when treatment effects are heterogeneous across cohorts or over time. In particular, recent studies have shown that two-way fixed effects estimates under staggered adoption may involve negative weighting and produce biased average treatment effects. To address this concern, we further employ the Callaway and Sant’Anna [59] estimator, which is robust to heterogeneous treatment timing, and re-estimate the policy effect under alternative aggregation schemes.

Table 7. Results of the CSDID test.

Variables	(1)	(2)	(3)	(4)	(5)
	Simple ATT	Calendar-Time ATT	Group ATT	Pre-Treatment ATT	Post-Treatment ATT
NAIDPZ	0.0472 ***	0.0351 ***	0.0405 ***	0.0061	0.0685 ***
	(0.012)	(0.009)	(0.006)	(0.0040)	(0.0130)
Control variables	√	√	√	√	√
Year FE	√	√	√	√	√
City FE	√	√	√	√	√
N	4539	4539	4539	4539	4539

Notes: This table reports the Callaway and Sant’Anna [59] DID estimates under alternative aggregation schemes. It presents the simple ATT, calendar-time ATT, group ATT and reports the average pre-treatment and post-treatment effects. Standard errors clustered at the city level are reported in parentheses. *** denotes significance at the 1% level. √ indicates that the corresponding control variable or fixed effect is controlled for in the regression model.

reports the CSDID estimates under alternative aggregation schemes. As shown in columns (1) to (3), the simple ATT, calendar-time ATT, and group ATT are all significantly positive, indicating that the positive effect of NAIDPZ policies on urban green economic efficiency remains robust after accounting for treatment effect heterogeneity under staggered policy adoption. Columns (4) and (5) further shows that the pre-treatment effect is statistically insignificant, whereas the post-treatment effect is significantly positive. This provides support for the absence of systematic differential pre-trends and suggests that the positive baseline results are unlikely to be driven by negative weighting or treatment-timing bias in the conventional two-way fixed effects specification, thereby further supporting the conclusion that AI-oriented policies contribute to the improvement of green economic efficiency.

5.3. Mechanism Tests

This section further examines the potential mechanisms through which NAIDPZ policies may improve urban green economic efficiency (GEE), focusing on green technological innovation and industrial structure optimisation. It provides mechanism-related evidence that is informative for understanding how the policy effect may arise. The results reported in this section should be interpreted as evidence consistent with the proposed mechanisms. At the same time, when read together with the baseline DID estimates and the supplementary identification results reported above, these findings offer stronger support for a mechanism-based interpretation of the main policy effect. It should be noted that the mechanisms discussed here should be understood as potential pathways associated with the composite NAIDPZ policies’ effect.

5.3.1. Green Technological Innovation

As indicated in

Table 8. Mechanism test.

Variables	Paten		Indus
	(1)	(2)	(3)	(4)
	Techn	Green	Advan	Justi
NAIDPZ	0.2927 ***	0.4640 ***	0.1614 ***	0.0456 **
	(0.105)	(0.107)	(0.034)	(0.019)
Constant	5.7417 ***	−22.6245 ***	4.3738 ***	1.0401 ***
	(1.383)	(0.566)	(0.448)	(0.222)
Controls	√	√	√	√
Year FE	√	√	√	√
City FE	√	√	√	√
$\mathrm{N}$	4539	4539	4505	4505
${\mathrm{R}}^2$	0.893	0.861	0.858	0.655

Notes: *** and **, denote statistical significance levels of 1% and 5%, respectively. Additionally, standard errors have been clustered at the city level and are presented in parentheses. √ indicates that the corresponding control variable or fixed effect is controlled for in the regression model.

, column (1), the total number of city-level patent authorisations is used in this study for measuring technical innovation. The findings show that the effect of the NAIDPZ policies on technological innovation is significantly positive at the 1% level. Column (2) further measures the green innovation based on the number of green patent authorisations. The findings indicate that the effect of NAIDPZ policies on green innovation is positive, which provides suggestive but relatively strong support for the view that green technological innovation may be an important mechanism through which NAIDPZ policies affect urban GEE.

The possible reason is that NAIDPZ policies have changed the income structure of technological innovation by strengthening R&D support, improving innovation infrastructure, and guiding the direction of technology application, thereby encouraging innovation activities to shift towards green and low-carbon fields. Firstly, as a typical general-purpose technology, breakthroughs in algorithms, computing power, and data in artificial intelligence can significantly reduce R&D costs, shorten the technology iteration cycle, and drive other innovation activities through the technology spillover effect [60]. Secondly, the use of AI in energy system optimisation, pollution control, and new material research has strongly increased the effectiveness of green technology research and development and may accelerate the transition of green technologies from laboratories to the market [61]. Furthermore, the integration of AI facilitates the formation of a collaborative innovation network, allowing for the real-time sharing of energy-saving data across different industrial chains. To the extent that these changes are induced by the policy shock, the expansion and diffusion of green technological innovation may reduce energy use and pollutant emissions while improving factor allocation efficiency. This process, in turn, can support higher economic output and product added value, thereby contributing to the improvement of GEE. Overall, the empirical evidence in this subsection is consistent with the argument that NAIDPZ policies may enhance GEE by fostering green technological innovation.

5.3.2. Industrial Structure Optimisation

The proportion of added value from tertiary to secondary industries is used to assess the extent of modernisation of the industrial structure. According to the results in Table 8 column (3), the coefficient of the NAIDPZ policies on the upgrading of the industrial structure is positive, suggesting that NAIDPZ policies may help promote the modernisation of the urban industrial structure towards higher value-added and lower energy consumption, which is consistent with a possible structural mechanism for improving green economic efficiency. Column (4) further measures the rationalisation of the industrial structure using the Theil index. The results show that the NAIDPZ policies significantly promoted the rationalisation of the industrial structure, offering supportive evidence that the policy shock may help create structural conditions conducive to higher green economic efficiency. This result indicates that NAIDPZ policies encouraged the optimisation of the industrial structure. This result supports research hypothesis H2.

The possible reason is that NAIDPZ policies not only realise structural upgrading across departments through industrial structure adjustment but also achieve structural rationalisation within departments through technological advancements and factor allocation optimisation, thereby producing a dual effect that enhances green economic efficiency. First, NAIDPZ policies lower the costs of transactions and information asymmetry, alleviating the long-term problem of factor misallocation in the traditional industrial system [62]. Furthermore, intelligent and digital transformation driven by NAIDPZ policies accelerates the effective movement and reorganisation of manufacturing elements in the city and across industries, encouraging the rational growth of the industrial structure [63]. This spatial and sectoral rebalancing facilitates the formation of green industrial clusters, where the proximity of digital infrastructure promotes the circular use of energy and materials, ultimately amplifying the aggregate green productivity of the urban economy.

6. Further Analysis

6.1. Moderation Effect Analysis

The regression results in

Table 9. Results of the moderation effect analysis.

Variables	(1)	(2)	(3)	(4)
	GEE	GEE	GEE	GEE
NAIDPZ	0.1379 ***	0.1291 ***	0.1217 ***	0.1123 ***
	(0.036)	(0.034)	(0.033)	(0.034)
Gover	0.0038	0.0050 *
	(0.003)	(0.003)
NAIDPZ × Gover	0.0156 ***	0.0146 ***
	(0.004)	(0.004)
Publi			0.0006	0.0010 **
			(0.000)	(0.000)
NAIDPZ × Publi			0.0013 ***	0.0012 **
			(0.000)	(0.000)
Constant	0.3496 ***	2.1045 ***	0.3707 ***	2.0820 ***
	(0.000)	(0.597)	(0.001)	(0.668)
Controls	×	√	×	√
Year FE	√	√	√	√
City FE	√	√	√	√
N	4386	4386	3096	3096
${\mathrm{R}}^2$	0.595	0.609	0.605	0.615

Notes: ***, **, and * denote statistical significance levels of 1%, 5%, and 10%, respectively. Addition-ally, standard errors have been clustered at the city level and are presented in parentheses. √ indicates that the corresponding control variable or fixed effect is controlled for in the regression model, while × indicates that it is not.

column (1) indicate that the interaction coefficient is significantly positive when the control variables are removed, demonstrating that government attention to the field of AI has strengthened the promoting impact of NAIDPZ policies. The positive regulatory effect of government attention has strong robustness, as seen by column (2), where the coefficient of the interaction term marginally reduces but remains highly significant after additional control variables are included. This finding implies that local governments’ level of attention and execution intensity significantly affects NAIDPZ policies. A possible reason is that greater government attention usually means stronger policy implementation constraints and clearer development orientations, which help reduce distortions and deviations in policy implementation and amplify the positive impact of artificial intelligence policies on the efficiency of the green economy [64,65].

Meanwhile, the interaction term is also considerably positive based on the regression findings in columns (3) and (4) of Table 9. This suggests that the high level of public awareness for environmental issues can strongly increase the promoting effect of NAIDPZ policies on GEE. This finding indicates that public environmental concern, as an important informal institutional arrangement, serves as an indispensable external constraint and incentive in the implementation of NAIDPZ policies. Specifically, when public environmental concern is high, local governments and enterprises face more social pressure for oversight and are more inclined to prioritise energy conservation and environmentally friendly solutions in the application of AI [66,67]. Moreover, the widespread public concern about environmental issues can also enhance market demand for green products and services, strengthen the economic returns from artificial intelligence-based green applications, and further amplify the policy’s positive incentive effect [68].

6.2. Heterogeneity Analysis

6.2.1. Resource Dependence Heterogeneity

Following Liu et al. [69], the study’s samples are divided into groups of resource-based cities and non-resource-based cities. The estimated coefficients of the NAIDPZ policies in both types of cities are significantly positive based on the results in

Table 10. Results of the heterogeneity analysis.

Variables	Development Resources		Geographical Location		Spatial Hierarchy
	Resou	Non-Resou	Coast	Non-Coast	Cente	Non-Cente
NAIDPZ	0.0609 ***	0.1280 ***	0.0796 ***	0.1574 ***	0.0670 ***	0.0332
	(0.018)	(0.037)	(0.021)	(0.025)	(0.016)	(0.022)
Constant	3.3546 ***	1.2116	0.5211	2.8379 ***	−0.6434	3.0123 ***
	(0.634)	(0.829)	(0.383)	(0.258)	(0.459)	(0.536)
Controls	√	√	√	√	√	√
Year FE	√	√	√	√	√	√
City FE	√	√	√	√	√	√
${\mathrm{R}}^2$	0.555	0.633	0.656	0.561	0.815	0.579
$\mathrm{N}$	1768	2737	1870	2635	595	3910

Notes: *** denotes statistical significance levels of 1%. Additionally, standard errors have been clustered at the city level and are presented in parentheses. √ indicates that the corresponding control variable or fixed effect is controlled for in the regression model.

’s columns (1) and (2). However, the policy effect is significantly higher in non-resource-based cities. This indicates that NAIDPZ policies are more effective at improving green economic efficiency in non-resource-based cities. The possible reason is that resource-based cities have long relied on primary processing and resource extraction, and their industrial systems exhibit typical characteristics of high energy consumption and low added value. Technological innovation and production models are easily locked into existing paths. Cities without a reliance on natural resources have a more diverse industrial makeup, with a larger proportion of high-end manufacturing and modern services. The industrial system is more adaptable to digital and intelligent technologies, and artificial intelligence is more likely to be applied a large scale to improve production efficiency, optimise energy utilisation, and strengthen pollution control, thereby fully releasing green economic dividends [70].

6.2.2. Urban Geographical Heterogeneity

Based on standards for dividing coastal areas, this paper conducts group regression using whether a city is located in a coastal area as the criterion. Table 10’s columns (3) and (4) reveal the findings. The findings show that NAIDPZ policies significantly improve the GEE in both inland and coastal cities. However, compared to coastal cities, the policy effect is significantly higher in inland cities. The following are the potential causes: Firstly, inland cities usually receive more targeted policy support for digital infrastructure development and green transformation. NAIDPZ policies often interact with other policies, such as digital economy support and green technology promotion, thereby significantly improving the efficiency of technology implementation and diffusion. In contrast, coastal cities began their green economic development earlier, and some have entered the marginal diminishing efficiency zone for efficiency improvement. Although NAIDPZ policies still have a positive effect, their incremental effect is relatively limited. Moreover, with rising factor costs and the acceleration of industrial transfer, some high-energy-consuming industries in coastal areas have spread to inland areas. NAIDPZ policies, by mitigating the environmental negative externalities arising from industrial transfer in inland cities, further amplify their impact on reducing GEE [71].

6.2.3. Spatial Hierarchical Heterogeneity

Based on the administrative hierarchy of cities and their regional functions, this paper separates the samples into non-central and central cities, which include municipalities directly under the central government, provincial capital cities, and sub-provincial cities. The promotion effect of NAIDPZ policies on the GEE of central cities is statistically significant based on the regression findings revealed in Table 10’s columns (5) and (6), while the significance test is not passed in peripheral cities, indicating that central cities are primarily affected by NAIDPZ policies. This difference reflects the crucial role of urban functional positioning and factor agglomeration capabilities in the implementation of artificial intelligence policies. Specifically, central cities, as the core of regional economic, technological, and innovation resources aggregation, usually have comprehensive digital infrastructure, high-density human capital, and a relatively mature green industrial system, which can provide a favourable institutional environment and application scenarios for the thorough integration of green economy with artificial intelligence technologies. Under such conditions, NAIDPZ policies can be more easily translated into tangible productivity gains and promote systematic improvements in green economic efficiency through technology spillovers and industrial linkages. In contrast, peripheral cities still have obvious shortcomings in digital infrastructure, innovation entities and the scale of green industries, and the introduction of artificial intelligence technology often faces insufficient application scenarios and low conversion efficiency, resulting in the inability of the green empowerment effect of policies to be fully released [72,73].

7. Conclusions

7.1. Main Results

Based on panel data from 267 prefecture-level Chinese cities from 2007 to 2023, this study uses the NAIDPZ’s establishment as a natural experiment. It identifies the composite policy effect of NAIDPZ on urban GEE using a DID model. The results show that, first, NAIDPZ policies significantly improve green economic efficiency. Secondly, green technology innovation and the upgrading of industrial structures constitute the primary transmission mechanism for AI policies to enhance green economic efficiency. In addition, the moderating impact study shows that public and government attention are crucial contextual regulatory factors in the implementation of NAIDPZ policies. In cities with higher government and public attention to the environment, NAIDPZ policies have a significantly larger promotion effect on green economic efficiency. Furthermore, the analysis of heterogeneity reveals that the green impact of NAIDPZ policies has obvious structural characteristics, and its effects are particularly noticeable in inland, central, and non-resource-based cities.

7.2. Theoretical Contribution and Practical Implications

This study contributes to the existing literature in several important ways. First, it extends the research on the economic and environmental consequences of AI by moving beyond firm-level or technological discussions to examine the policy effect of AI-oriented pilot programmes on urban green economic efficiency. In doing so, it provides empirical evidence on how AI policy serves as a form of institutional intervention that can shape green development outcomes at the city level. Second, this study enriches the analytical framework on the green effects of AI policies by identifying green technology innovation and industrial structure upgrading as two key transmission mechanisms, thereby clarifying how AI policy can be translated into green performance gains. Third, by incorporating government attention and public environmental concern into the analysis, this study further reveals the conditions under which AI policies generate stronger green effects, thus deepening the understanding of heterogeneity in policy implementation and effectiveness. More broadly, this paper also contributes to the emerging literature linking AI-oriented policy intervention, green transformation, and urban governance. At the same time, the present study still leaves room for further theoretical development, particularly in terms of constructing a more fully integrated conceptual framework to formalise the relationships among policy intervention, mechanism pathways, and green performance. This remains an important direction for future research.

These findings also have practical implications. They suggest that AI policy should not be understood merely as a tool for technological advancement but as an important governance instrument for promoting urban green transformation. Whether AI policy can effectively improve green economic efficiency depends not only on policy design itself but also on complementary conditions such as innovation capacity, industrial upgrading, government commitment, and public participation. Therefore, the green dividends of AI development can be better realised only when technological empowerment is coordinated with institutional support and differentiated local implementation. This provides governments with potential policy references for promoting green urban development through the implementation of sound AI policies.

7.3. Policy Recommendations

The following policy proposals are suggested in light of the previously mentioned research findings. First, improve the strategic coherence and implementation efficacy of NAIDPZ policies with a green development orientation. Policymakers should systematically refine AI-related regulatory frameworks, strengthen interdepartmental coordination mechanisms, and improve policy integration across levels of government. Building on experience from pilot initiatives, the scope of AI policy deployment should be expanded in a phased orderly manner. Policy replication and institutional diffusion should be leveraged to unlock the green innovation potential embedded in AI governance, thereby reducing the institutional transaction costs associated with green technology adoption. Concurrently, explicit green development targets—such as carbon intensity reduction, resource efficiency gains, and ecological restoration benchmarks—must be institutionalised within AI policy instruments to correct market actors’ tendency to prioritise technological advancement over environmental sustainability.

Second, accelerate green technological innovation and structural upgrading through targeted public investment and incentive alignment. Public R&D funding should be prioritised for cross-domain integration between AI and green technologies—including clean energy systems, intelligent environmental monitoring, and green manufacturing processes. Dedicated innovation funds should support collaborative R&D projects involving universities, research institutes, and enterprises, thereby expediting the commercialisation and diffusion of green technologies. Complementary fiscal instruments—including preferential tax treatment, enhanced R&D expense deductions, and green innovation subsidies—should be deployed to internalise environmental externalities, strengthen firms’ endogenous capacity, and facilitate the structural pivot of industrial chains toward high-value, low-carbon configurations that minimise carbon leakage.

Third, adopt context-sensitive differentiated policy implementation strategies to accommodate urban heterogeneity and mitigate regional innovation disparities. In core metropolitan areas, policy efforts should emphasise breakthroughs in foundational AI technologies and the clustering of green industries. In secondary and peripheral cities, priority should be placed on digital infrastructure modernisation and the systematic attraction of innovation talent and institutions. Non-resource-based cities should leverage AI policy tools to catalyse green industrial diversification and upgrading, whereas resource-dependent cities should embed AI applications into resource extraction optimisation, energy efficiency enhancement, emission control, and ecosystem rehabilitation. Furthermore, coastal cities should deepen international AI–green cooperation through open platforms and bilateral agreements, while inland cities should strategically deploy policy incentives—including land use concessions, talent subsidies, and streamlined administrative procedures—to attract external innovation resources, thereby fostering a more balanced and integrated national AI–green innovation ecosystem.

Fourth, strengthen institutional capacity for green-AI governance and foster inclusive multi-stakeholder engagement to ensure the long-term legitimacy of technological transitions. Governments at all levels must elevate AI-enabled green development as a strategic priority, integrating measurable green–AI performance indicators—such as green patent output, AI-driven emission reductions, and smart environmental governance coverage—into local official evaluation systems to ensure policy continuity and accountability. Concurrently, digital governance capabilities should be upgraded to harness AI for real-time environmental monitoring, predictive risk assessment, and adaptive public service delivery. Transparent information disclosure, participatory planning mechanisms, and co-design platforms should be institutionalised to mobilise social capital, civil society organisations, and citizen expertise, thereby cultivating a robust, adaptive, and accountable multi-level governance ecosystem that maximises the green productivity gains enabled by AI policies.