The pilot zone policy has its roots in the State Council’s 2017 New Generation Artificial Intelligence Development Plan, which laid out a three-stage roadmap through 2030 for embedding AI into China’s economy and society. Two years later, the Ministry of Science and Technology issued guidelines for constructing national AI innovation and development pilot zones, approving successive batches of cities with established industrial capacity and suitable conditions for AI application. Unlike emissions trading schemes or environmental taxes, which directly raise the cost of pollution, the pilot zones work indirectly. They combine technological research and development, application demonstrations, policy experimentation, and infrastructure building, while supporting data openness, technology commercialization, intellectual property protection, talent development, and financial services designed to fuse AI with the real economy.

What makes this study distinctive is its outcome variable. Rather than measuring productivity, green patents, or emissions in isolation, the researchers constructed a composite index of corporate sustainable development performance that captures whether economic viability and environmental improvement genuinely move together. They standardized each firm’s return on assets and its environmental score from the Huazheng ESG rating system onto a zero-to-one scale, then combined them using a geometric mean and a coordination term. The geometric mean ensures that strong performance in one dimension cannot fully compensate for weakness in the other, while the coordination term rewards firms whose economic and environmental scores are closely balanced. This design directly addresses a gap in prior research, which had examined the pilot zones’ effects on labor allocation, green innovation, and energy use separately without establishing whether economic gains coincided with environmental progress.

The empirical setting is formidable. The researchers assembled 17,483 firm-year observations of Chinese A-share listed companies from 2009 to 2023, excluding the four centrally administered municipalities of Beijing, Shanghai, Tianjin, and Chongqing because their administrative status differs from prefecture-level cities. Roughly 16.3 percent of the sample firms sit in cities affected by the policy. Because exposure to the pilot zones is not random, cities with stronger pre-existing AI foundations were more likely to be designated, the team turned to double machine learning, a method that uses machine-learning algorithms to estimate the conditional expectations of both the outcome and the treatment, then extracts the policy effect from the residualized variables. Orthogonalization and five-fold cross-fitting mitigate regularization bias and overfitting, allowing the estimate to remain robust even when the relationships between firms, cities, and sustainability outcomes are nonlinear.

The headline result is clear: the establishment of AI pilot zones raises corporate sustainable development performance by an estimated 0.026 index points, significant at the 1 percent level even after firm and year fixed effects are included. Relative to the sample mean of 0.384, that represents an improvement of roughly 6.8 percent, or about 0.15 standard deviations of the overall variation in the measure. The finding survived an unusually thorough battery of robustness checks. An instrumental variable built from 1984 fixed-line telephone penetration per 100 residents, interacted with year dummies, produced a second-stage estimate of 0.068, with a Kleibergen-Paap statistic of 45.89 far exceeding weak-instrument thresholds. Alternative winsorization thresholds, z-score normalization, a three-dimensional entropy-weighted indicator spanning economic, environmental, and social performance, and the overall Huazheng ESG score all yielded positive and significant coefficients. So did swapping the Lasso algorithm for ElasticNet, support vector machines, and gradient boosting, and adjusting the cross-fitting split ratio.

Perhaps most striking is what the decomposition reveals. When the researchers broke the composite index into its components, return on assets actually declined by 0.008 points, significant at the 5 percent level, while environmental performance rose by 0.035, the balance between the two dimensions improved by 0.028, and the joint development level climbed by 0.022, all significant at the 1 percent level. In other words, the policy’s positive effect does not stem from firms simultaneously becoming more profitable and cleaner in the short run. Instead, it reflects greener operations and a better-coordinated trajectory, with the dip in profitability plausibly reflecting the upfront costs of AI adoption, equipment upgrades, data infrastructure, and employee training. The authors interpret this through the lens of the Porter hypothesis’s innovation offset: cleaner technologies may generate economic returns over a longer horizon, even as they impose near-term adjustment costs.

Mechanism tests point to three channels through which the policy works. First, green innovation: the share of green patent applications in total patents rose by 0.009 points, and citations to firms’ green patents increased by 0.046, both statistically significant. Computing infrastructure and AI-enabled tools appear to help firms analyze energy use and emissions, improving project selection and reducing uncertainty in green research and development, while collaboration among firms, universities, and research institutions spreads green technologies more widely. Second, governance: the management expense ratio fell by 0.007 and the ratio of other receivables to total assets declined by 0.001, suggesting that AI-enabled information systems make managerial decisions more traceable, weaken managers’ informational advantages, and constrain opportunistic or myopic behavior. Third, human capital: the proportion of technical employees in the workforce rose by 0.013, reflecting both rising demand for analytical skills and an expanded local supply of qualified workers concentrated in pilot cities.

The policy’s benefits are not evenly distributed, and the heterogeneity analysis is where the study’s implications become sharpest. The effect is statistically insignificant for state-owned enterprises but strong and significant at the 1 percent level for non-state firms, with a difference test yielding a p-value of 0.006. The authors reason that state firms already enjoy stable access to policy information, financing, and talent, so the pilot zones’ marginal contribution is smaller, whereas non-state firms facing tighter financing constraints and higher recruiting costs derive greater value from the zones’ technology platforms and talent programs. Similarly, firms in heavily polluting industries show no significant effect, while firms outside those industries gain significantly, consistent with the idea that carbon-intensive energy structures, specialized fixed assets, and technological path dependence make AI adoption alone insufficient to transform dirty production systems. High-tech firms, with their established R&D systems and digital infrastructure, absorb the policy’s resources far more effectively than non-high-tech peers.

For policymakers and executives, the findings suggest that place-based AI policy can serve as a lever for green corporate transformation, but only when the surrounding conditions are right. The authors recommend expanding demonstration scenarios for green technologies, improving access to environmental data, strengthening collaboration among firms, universities, and research institutions, and lowering the costs of obtaining relevant technologies. Firms, meanwhile, could deploy AI-enabled information systems to strengthen internal governance and reduce agency costs, and align workforce training with the digital and analytical skills that AI adoption demands. For heavily polluting firms, the study cautions that AI support must be paired with cleaner-energy substitution, equipment upgrades, and production-line transformation rather than expected to deliver green change on its own.

The research also carries honest caveats. The environmental component of the composite measure relies on a single rating provider, whose assessment frameworks and firm coverage may differ from those of other raters, and the sample is limited to listed firms, leaving small and medium-sized enterprises and unlisted companies outside the frame. The analysis estimates effects on firms located in designated pilot cities without explicitly modeling spatial spillovers, meaning that technology and talent flows across regional borders could contaminate the comparison group. Future work, the authors suggest, could cross-validate sustainability measures across multiple rating providers, extend the analysis with firm surveys and tax data, and apply spatial econometric models to separate direct policy effects from spillovers to non-pilot regions. Even with these limitations, the study offers some of the firm-level evidence yet assembled on how a technology-oriented, place-based AI policy can nudge companies toward a more sustainable balance between growth and the environment.

Subject of Research: Effects of China's artificial intelligence pilot zone policy on corporate sustainable development performance

Article Title: Corporate sustainable development effects of the artificial intelligence pilot zone policy: evidence from China based on double machine learning

Article References: Gao, Q., Hu, W., Wang, X., & Liu, Y. (2026). Corporate sustainable development effects of the artificial intelligence pilot zone policy: evidence from China based on double machine learning. Frontiers in Environmental Science, 14, Article 1909483. https://doi.org/10.3389/fenvs.2026.1909483

Image Credits: AI Generated

DOI: 10.3389/fenvs.2026.1909483

Keywords: artificial intelligence pilot zones, corporate sustainable development, double machine learning, green innovation, agency costs, human capital upgrading, China, ESG performance, environmental policy, quasi-natural experiment, listed firms, green transformation