The researchers reached that conclusion by treating the staggered rollout of the pilot zones as what econometricians call a quasi-natural experiment. Because different cities were approved in different years—the first batch of six in 2019, with subsequent batches approved between 2020 and 2021—the policy arrived at distinct times in distinct places, allowing the team to compare cities before and after designation against cities that had not yet been designated. They assembled a balanced panel covering 271 prefecture-level and higher cities from 2010 to 2023, drawing on national and provincial statistical yearbooks, and applied an overlapping difference-in-differences model, a technique suited to policies that switch on at different moments for different units.

The outcome variable was deliberately concrete. Intensification of industrial land use, measured as the logarithm of secondary-sector value added per unit of industrial land area, captures how much economic value each hectare of factory and industrial ground produces. In the baseline regression, the policy indicator carried a coefficient of 0.203, significant at the 1 percent level, which the authors translate into an approximately 22.5 percent increase in land-use intensity in pilot cities. Because the dependent variable is a ratio, however, the team had to rule out an unflattering alternative explanation: that the number rose not because output grew but because industrial land shrank as factories were displaced. Separate regressions on the numerator and denominator showed that AI policies significantly increased secondary-sector output while having a negative but statistically insignificant effect on land area, pointing to genuine efficiency gains rather than passive spatial contraction.

The study then subjected that headline finding to an unusually thorough battery of robustness checks. A parallel trends test confirmed that pilot and non-pilot cities were on statistically indistinguishable trajectories before designation, with pre-policy effect sizes hovering around zero. A placebo test randomly reassigned the pilot designation across cities 1,000 times and found the true coefficient sitting in the far right tail of the resulting distribution, far from the zero-centered noise. A propensity score-matched difference-in-differences design addressed self-selection bias—the worry that only already-dynamic cities were chosen as pilots—by matching treated cities to statistically comparable controls on economic development, science spending, government intervention, financial depth, wages, and industrial structure. After matching, standardized deviations between the two groups collapsed, with some variables showing bias reductions above 99 percent, and the policy effect survived.

The authors also confronted a subtler methodological hazard. In staggered adoption settings, the conventional two-way fixed-effects estimator can be distorted when treatment effects vary over time, because it implicitly uses already-treated units as comparison groups. Using the Goodman-Bacon decomposition, they showed that 92.81 percent of the identifying weight came from comparisons between treated and never-treated cities, the most reliable type, while the potentially problematic timing comparisons accounted for just 1.02 percent. Subsample regressions by approval batch told a consistent story: the effect was strongest in the first-batch cities, insignificant in the tiny second batch of four cities where statistical power was limited, and highly significant in the third batch. Excluding the first batch entirely still left a positive, significant coefficient of 0.110, indicating the result was not driven solely by flagship cities such as Beijing, Shanghai, and Shenzhen.

Having established that the policy worked, the researchers turned to how. Their answer centers on digital industry agglomeration, measured through a location entropy index of employment in telecommunications, computer services, and software. The logic is a causal chain running from technology policy to industrial space to land use: AI pilot zones attract digital enterprises, talent, and infrastructure; those enterprises cluster; and clustering raises the economic density of industrial land. Empirically, the policy raised digital industry agglomeration with a coefficient of 0.405, agglomeration in turn predicted land-use intensity with a coefficient of 0.078, and mediation analysis attributed about 15.53 percent of the total effect to this pathway. Sobel tests and bootstrap confidence intervals confirmed the indirect effect was statistically significant, and the finding held when the mediating variable was lagged by one period or replaced with the density of registered ICT firms.

The mechanism makes economic sense for reasons rooted in agglomeration theory. Digital industries are characterized by high value-added output and relatively low land consumption, so their clustering mechanically raises economic output per unit of land. Knowledge spillovers and technology diffusion within clusters then push traditional manufacturers toward intelligent transformation, lifting productivity on existing industrial parcels without new construction. The framework also draws on the economics of misallocation: AI, as a general-purpose technology, helps concentrate land and other resources in high-productivity firms rather than low-productivity ones, extracting more value from a fixed land supply. Notably, the authors caution that agglomeration is not automatically beneficial at every scale—past a threshold, crowding and congestion can generate negative externalities—which keeps the policy prescription nuanced rather than simply maximalist.

Perhaps the most striking result concerns policy combinations. The team examined what happened when AI pilot zones overlapped with municipal public data openness policies, identified using the Fudan University Digital and Mobile Governance Laboratory’s reporting on city data platforms. In isolation, open data policies showed a significantly negative direct effect on industrial land intensity, consistent with the theoretical argument that early-stage data platforms impose compliance costs, inconsistent standards, and data-quality headaches on firms—data available but not yet usable. But the interaction between the two policies was strongly positive. When implemented together, the total effect significantly exceeded the sum of the individual effects, a superadditive result the authors describe as a one-plus-one-greater-than-two amplification. The interpretation is complementary: AI technologies are data-hungry, so opening public data lowers the cost of the key input, while AI raises the value extractable from that data.

The synergy was not uniform across geography or time. Ordinary prefecture-level cities, which start from lower baselines of land-use intensity, showed slightly larger synergy effects than higher-tier cities, where marginal room for improvement is smaller. More tellingly, cities whose data platforms launched before 2019 displayed a highly significant synergy coefficient of 0.407, while late-launching cities showed an insignificant 0.053. Platform maturity, accumulated datasets, and established data standards appear to be prerequisites for the technology-data feedback loop to ignite. A moderated mediation test sharpened the point: the indirect pathway from AI policy through digital industry clustering to land intensification was significant only where open data policies were in place, with an indirect effect of 0.0542 and a confidence interval excluding zero; without data openness, that same pathway was statistically indistinguishable from zero.

The implications stretch beyond land economics. China’s industrial land has long expanded rapidly and been used extensively, contributing to urban sprawl that consumes suburban farmland and heightens tension between development and food security. If AI and data policies can raise the economic density of existing industrial land, they reduce demand-side pressure for horizontal expansion, creating policy space for arable land conservation even if converted parcels do not automatically return to agriculture. The authors recommend expanding the pilot zone system, incorporating land intensity into local policy evaluation, investing in digital infrastructure, and pairing AI initiatives with mature open data platforms. They also flag open questions: the analysis captures average temporal effects but not spatial spillovers between neighboring cities, and future work with firm-level data could dissect the numerator-driven versus denominator-driven pathways more finely. What the study establishes now is that artificial intelligence policy, at least as deployed across 18 Chinese pilot zones, is not merely a technology program—it is a land-use program, and it works best when the data flows freely.

Subject of Research: Effects of China's AI pilot zone policy on the intensification of industrial land use

Article Title: Artificial intelligence policy, digital industry agglomeration, and industrial land intensive use: evidence from a quasi-natural experiment in China

Article References: Wang, N., Fu, Y., & Liu, Z. (2026). Artificial intelligence policy, digital industry agglomeration, and industrial land intensive use: evidence from a quasi-natural experiment in China. Frontiers in Sustainable Food Systems, 10, Article 1858439. https://doi.org/10.3389/fsufs.2026.1858439

Image Credits: AI Generated

DOI: 10.3389/fsufs.2026.1858439

Keywords: artificial intelligence policy, industrial land use, digital industry agglomeration, difference-in-differences, quasi-natural experiment, pilot zones, public data openness, policy synergy, urban sustainability, land economics, China, digital economy