Digital Economy, Regional AI Orientation, and Industrial Structure Upgrading Under Economic Policy Uncertainty: Evidence from China

Abstract

This study examines whether the digital economy helps provincial economies sustain industrial structure upgrading under economic policy uncertainty (EPU), and whether regional AI orientation strengthens this role. Using a balanced panel of 30 Chinese provinces from 2015 to 2023, the study uses the standardised logarithm of a provincial digital economy index as its core measure of digital development. Province and year fixed-effects models show that the triple interaction among digital economy development, regional AI orientation, and high EPU is positive and statistically significant. Marginal effect analysis indicates that the digital economy effect under high EPU only becomes positive when regional AI orientation exceeds a threshold, suggesting a conditional rather than universal effect. Robustness checks, alternative dependent variables, province-grouped machine learning validation, and supplementary policy exposure evidence based on Broadband China pilots are consistent with this state-dependent complementarity, although the estimates are interpreted as conditional associations rather than definitive causal effects.

1. Introduction

Industrial structure upgrading—the reallocation of productive resources from labour-intensive secondary industries toward knowledge-intensive, service-oriented, and technology-enabled activities—is central to long-run regional development (Kuznets, 1966; Herrendorf et al., 2014; Chenery & Taylor, 1968; Baumol, 1967; Ngai & Pissarides, 2007). In China, the transition from investment-led manufacturing toward service-oriented and innovation-driven growth has become an explicit policy priority. Yet such structural transformation remains sensitive to macroeconomic disruptions. Among the most consequential sources of disruption in recent years has been economic policy uncertainty (EPU), which refers to ambiguity surrounding future taxation, regulation, trade policy, fiscal stance, and broader macroeconomic policy.

EPU constrains industrial upgrading through several economic channels. Via the real-options mechanism, uncertainty raises the option value of waiting before committing to irreversible investment and structural adjustment (Bloom, 2009). Via the financial frictions channel, lenders tighten credit under uncertainty, raising the cost of capital for longer-horizon service and technology activities (Gulen & Ion, 2016; Phan et al., 2019). Via the risk-premia channel, elevated uncertainty increases the required return on investment projects (Pástor & Veronesi, 2013; Julio & Yook, 2012). These mechanisms are especially relevant for industrial upgrading because upgrading involves cross-sector resource reallocation, the development of producer-service ecosystems, technology-market transactions, and the formation of new institutional arrangements. Unlike short-run output fluctuations, which can often be absorbed through changes in capacity utilisation, industrial upgrading requires long-term commitments that are difficult to reverse. Recent evidence from China confirms that EPU restrains industrial upgrading at the provincial level (Ye et al., 2024; Zhu et al., 2022).

Two potential buffering mechanisms motivate this paper. The first is the digital economy. Digital infrastructure and digital financial inclusion can reduce search costs, information frictions, verification costs, and financing barriers, thereby sustaining credit flows, market matching, and resource reallocation under uncertainty (Goldfarb & Tucker, 2019; Beck et al., 2007; Brynjolfsson & Hitt, 2000). A large body of Chinese literature documents that digital finance and the digital economy are positively associated with industrial upgrading, innovation, and high-quality development (Wu & Shao, 2022; Ren et al., 2023; Y. Shen & Ren, 2023; Zheng et al., 2023; Jiao et al., 2024; Li et al., 2024; H. Shen et al., 2024; Chen et al., 2025). However, much of this literature focuses on average effects. Whether digital infrastructure functions as uncertainty-buffering infrastructure—providing stronger support for industrial upgrading when EPU is high—has not been systematically examined.

The second mechanism is regional AI orientation. Artificial intelligence is a general-purpose technology that reduces the cost of prediction, supports adaptive decision-making, and reshapes the organisation of tasks and production (Brynjolfsson & McElheran, 2016; Agrawal et al., 2018, 2019; Acemoglu & Restrepo, 2019; Autor, 2015). Provinces where listed firms strategically engage with AI-related technologies may be better positioned to use digital infrastructure for industrial upgrading, particularly under high-EPU conditions when prediction, adaptation, and rapid reallocation become more valuable. Nevertheless, AI orientation is unlikely to operate as an independent regional buffer by itself. Without deep digital financial services, functioning technology markets, and knowledge-intensive producer services, strategic AI engagement may not translate into aggregate structural upgrading. AI orientation is therefore more plausibly understood as a complementary capability that amplifies the role of digital infrastructure.

China’s provincial panel provides a suitable setting for examining this complementarity. EPU in China is largely driven by national policy cycles that are common to all provinces in a given year. At the same time, provinces differ substantially in their digital economy development, the AI orientation of their listed firms, and their industrial structures. This combination of common national uncertainty shocks and cross-provincial variation in digital and AI capabilities provides the basis for testing whether provinces that are both digitally developed and AI-oriented respond differently to high-EPU periods.

This paper uses a balanced province–year panel of 30 mainland Chinese provinces from 2015 to 2023. The dependent variable is the logarithm of the ratio of tertiary industry value added to secondary industry value added. Digital economy development is measured by the standardised logarithm of a provincial digital economy index. The measurement follows the entropy-weighting approach widely used in digital economy studies and draws on a multidimensional framework covering digital infrastructure, digital industrialisation, and industrial digitisation (Zhao et al., 2020; Peng & Dan, 2023). The index is strictly positive in the sample, so log(Score) is used without adding a shift term. Regional AI orientation is constructed by aggregating annual report-based AI text-mining scores of A-share listed firms to the province–year level.

This study makes four contributions. First, it introduces a state-dependent perspective into the literature on the digital economy and industrial upgrading. Existing studies mainly estimate average effects, whereas this paper asks whether the digital economy plays a stronger role during high-EPU periods. Second, it constructs a regional AI orientation measure by aggregating firm-level annual report AI signals to the province–year level, thereby linking firm-level technology orientation to regional structural transformation. Third, it examines the three-way relationship among digital economy development, regional AI orientation, and economic policy uncertainty. This allows the paper to distinguish between AI as an independent regional driver and AI as a complement to digital infrastructure. Fourth, it asks whether the digital–AI combination carries additional predictive information for upgrading using cross-validated machine learning models; this evidence is predictive only and is not used for causal identification.

The empirical results support a conditional complementarity argument. The triple interaction among digital economy development, regional AI orientation, and high EPU is positive and statistically significant. The marginal effect analysis (Figure 1) shows that the digital economy is not an unconditional buffer: its positive association with upgrading under high EPU appears mainly when regional AI orientation is sufficiently high. Robustness checks show that the finding is not driven by the exclusion of Tibet, by the digital finance component of the index, or by overlapping controls. Supplementary policy exposure and machine learning exercises provide additional, non-causal evidence consistent with the main mechanism.

2. Literature Review and Hypothesis Development

2.1. Economic Policy Uncertainty and Industrial Upgrading

Economic policy uncertainty (EPU) has become an important framework for understanding how ambiguity in future policy conditions affects investment, innovation, and structural adjustment. Baker et al. (2016) formalise EPU as a distinct macroeconomic risk and construct widely used text-based EPU indices (see (Al-Thaqeb & Algharabali, 2019) for a review). Unlike general macroeconomic volatility, EPU captures uncertainty about the direction, timing, and intensity of government policy, including taxation, fiscal expenditure, industrial regulation, trade policy, financial supervision, and other forms of macroeconomic intervention. Because policy rules shape firms’ expectations about future costs, demand, financing conditions, and regulatory constraints, elevated EPU can affect both firm-level decision-making and regional economic transformation.

The economic consequences of EPU are usually explained through three channels. First, the real-options channel suggests that uncertainty raises the value of waiting before committing to irreversible investment (Bloom, 2009). When firms face uncertainty about future policy conditions, they may postpone investment in new technologies, new service activities, and new organisational capabilities. This delay effect is especially important for industrial upgrading because upgrading often requires long-term and partially irreversible commitments. Second, the financial frictions channel suggests that lenders become more cautious under uncertainty, reducing credit supply and increasing financing costs for firms undertaking long-horizon projects (Gulen & Ion, 2016; Phan et al., 2019). Since industrial upgrading often depends on technology adoption, producer-service expansion, and intangible investment, tighter credit conditions can directly slow structural transformation. Third, the risk-premia channel suggests that investors require higher returns when policy uncertainty increases, thereby raising the threshold for new investment and discouraging projects with uncertain payoffs (Pástor & Veronesi, 2013; Julio & Yook, 2012).

Industrial structure upgrading is likely to be more sensitive to EPU than ordinary output growth. Short-run output fluctuations can often be absorbed through changes in capacity utilisation, inventories, or temporary labour adjustments. By contrast, industrial upgrading requires the reallocation of capital, labour, technology, and managerial attention from traditional secondary industries toward services, knowledge-intensive activities, and technology-enabled sectors. This process involves coordination among firms, financial institutions, labour markets, technology suppliers, and local governments. It also requires the formation of complementary institutions such as technology markets, producer-service platforms, and digital financial channels. These forms of structural adjustment are difficult to reverse once implemented, and their returns depend heavily on expectations about future policy stability.

Under high EPU, firms may therefore delay or scale down upgrading-related activities. Manufacturing firms may postpone service-oriented transformation, technology adoption, or participation in technology-market transactions. Financial institutions may become more conservative in extending credit to emerging service sectors or technology-intensive businesses. Local governments may also face greater difficulty in coordinating industrial policy, attracting investment, and supporting new market institutions when macroeconomic and regulatory signals are uncertain. As a result, high-EPU periods may slow the transition from manufacturing-oriented growth toward service-oriented and innovation-driven development.

Recent studies provide evidence that EPU affects structural transformation in China. Ye et al. (2024) and Zhu et al. (2022) show that EPU can restrain provincial industrial upgrading and that its effects may vary across regions and resource endowments. Related research also documents that policy uncertainty influences firm investment, cash holdings, innovation behaviour, and environmental adjustment (Yu et al., 2021). These studies indicate that EPU is not only a macroeconomic background condition, but also a constraint on regional resource allocation and industrial transformation.

However, the existing EPU-and-upgrading literature pays less attention to the conditions under which regions can mitigate the adverse effects of uncertainty. If EPU works by increasing waiting incentives, financing constraints, information frictions, and coordination costs, then regional capabilities that reduce these frictions should help sustain upgrading during uncertain periods. Digital infrastructure is one such capability. Digital finance, digital platforms, and data-based matching mechanisms may help maintain credit access, reduce search and verification costs, and support cross-sector resource reallocation when policy uncertainty is high. Therefore, the role of the digital economy may be especially important in high-EPU periods rather than uniform across all macroeconomic conditions.

This reasoning suggests a state-dependent relationship between the digital economy and industrial upgrading. During normal periods, digital finance may promote upgrading by improving efficiency and supporting market expansion. During high-EPU periods, its role may become more defensive and resilience-oriented: it may help firms and regions continue structural adjustment despite weaker expectations and tighter financing conditions. This leads to the first hypothesis.

H1: 

Digital economy development is positively associated with industrial structure upgrading during high-EPU periods.

2.2. Digital Economy as Uncertainty-Buffering Infrastructure

The digital economy affects regional development by changing the cost structure of economic exchange. Goldfarb and Tucker (2019) argue that digital technologies reduce several fundamental economic costs, including search, replication, transportation, tracking, and verification costs. These reductions are directly relevant to industrial upgrading because structural transformation depends on the efficient movement of capital, labour, technology, and services across sectors. When search and verification costs are lower, firms can more easily identify new suppliers, customers, technology partners, and financing channels. When digital transactions and digital platforms become more widespread, market matching improves, and the cost of entering new service-oriented activities declines.

In the financial domain, digital finance can broaden access to financial services and reduce information asymmetry between borrowers and lenders. Traditional financial institutions often rely on collateral, long-term relationships, and balance-sheet indicators, which may disadvantage smaller firms, service firms, and firms engaged in intangible or technology-oriented activities. Digital finance uses transaction records, platform data, payment information, and other digital traces to improve credit assessment and risk monitoring. This can ease financing constraints and support more inclusive resource allocation (Beck et al., 2007; Guo et al., 2020). For industrial upgrading, this matters because new service activities, producer services, and technology-intensive businesses often require flexible financing and have fewer tangible assets than traditional manufacturing.

In China, the Peking University Digital Financial Inclusion Index has become a widely used measure of provincial digital finance development. The index captures three dimensions: coverage breadth, usage depth, and the digitisation level of financial services (Guo et al., 2020). Coverage breadth reflects the extent to which digital financial services reach users. Usage depth captures how intensively these services are used for payments, credit, insurance, investment, and other financial activities. The digitisation level reflects the convenience, affordability, and digital intensity of financial transactions. The growing literature links digital financial inclusion and the broader digital economy to industrial upgrading, innovation, green development, and high-quality growth in China (Wu & Shao, 2022; Ren et al., 2023; Y. Shen & Ren, 2023; Zheng et al., 2023; Jiao et al., 2024; Li et al., 2024; H. Shen et al., 2024; Chen et al., 2025). Digital finance is treated here as one important resource-allocation channel within the broader digital economy, rather than as a substitute for the broader digital economy concept; the empirical section uses a multidimensional provincial digital economy index that encompasses digital infrastructure, digital industrialisation, and industrial digitisation alongside digital financial inclusion.

However, much of the existing literature focuses on the average effect of the digital economy. It usually asks whether provinces or cities with higher levels of digital finance experience faster industrial upgrading on average. This approach is useful, but it does not fully capture the state-dependent role of digital infrastructure. From a friction-reduction perspective, the economic value of digital infrastructure should be especially visible when frictions are severe. High EPU increases information asymmetry, weakens confidence, tightens financing conditions, and makes firms more cautious about entering new sectors or investing in new technologies. Under these conditions, digital finance and digital platforms may not merely promote growth; they may help preserve the functioning of resource-allocation channels that would otherwise weaken.

The uncertainty-buffering role of the digital economy can be understood through four mechanisms. First, digital infrastructure reduces information frictions. Digital platforms, payment systems, and data-based financial services generate information about firms, consumers, and transactions. This information helps reduce uncertainty about borrower quality, customer demand, supplier reliability, and market conditions. During high-EPU periods, when traditional signals become less reliable, these digital information channels can help firms and financial institutions continue making allocation decisions.

Second, digital finance can mitigate financial frictions. EPU often induces banks and investors to become more conservative. Firms engaged in service-oriented or technology-intensive upgrading may face greater difficulty obtaining credit because their assets are intangible and their returns are uncertain. Digital credit channels, online supply-chain finance, and platform-based risk assessment can provide alternative financing routes. These tools may be particularly valuable for firms attempting to maintain upgrading-related investment when traditional bank credit tightens.

Third, digital infrastructure reduces transaction costs. Digital payment systems, online platforms, logistics systems, and digital business services lower the cost of market entry and exchange. This is important for the emergence of producer services and new service activities because many such activities depend on frequent interactions, small transactions, and cross-regional service delivery. Lower transaction costs make it easier for firms to experiment with new business models and participate in service-sector expansion even under uncertain policy conditions.

Fourth, digital platforms improve spatial matching. Industrial upgrading often requires firms to connect with technology suppliers, financial service providers, professional service firms, and customers beyond their local market. Digital platforms expand the geographic scope of matching by allowing firms to access external knowledge, services, and financing. This can reduce the dependence of upgrading on local market conditions and help provinces sustain structural transformation even when local demand or policy expectations are weak.

These four mechanisms suggest that the digital economy should be interpreted not only as a source of average productivity improvement, but also as a form of uncertainty-buffering infrastructure. When EPU is low, digital infrastructure may support upgrading by improving efficiency and market expansion. When EPU is high, its role may become even more important because it helps reduce the very frictions that uncertainty amplifies. A province with stronger digital economy infrastructure is therefore expected to be better able to sustain industrial upgrading during high-EPU periods.

This state-dependent view extends the existing digital economy literature. Instead of only asking whether digital economy development promotes industrial upgrading on average, this paper asks whether the digital economy becomes more valuable when policy uncertainty is high. This distinction is important because a purely average effect approach may underestimate the resilience function of digital infrastructure. If digital economy development helps provinces maintain credit access, market matching, and resource reallocation under uncertainty, then its association with industrial upgrading should be particularly visible in high-EPU years.

2.3. Regional AI Orientation as a Complementary Capability

Artificial intelligence is widely regarded as a general-purpose technology that can reshape prediction, decision-making, task organisation, and resource allocation. Its economic significance does not only lie only in automation but also in its ability to reduce the cost of prediction and improve the quality of decisions under uncertainty (Brynjolfsson & McElheran, 2016; Agrawal et al., 2018, 2019; Acemoglu & Restrepo, 2019; Autor, 2015; Acemoglu et al., 2022). When firms face uncertain demand, changing regulation, and unstable market conditions, AI-related capabilities can help them process information, forecast risks, identify new opportunities, and adjust production or service strategies more quickly. These functions are particularly relevant under high economic policy uncertainty, when traditional decision-making based on historical experience becomes less reliable.

However, AI does not automatically translate into regional industrial upgrading. AI-related technologies require complementary digital infrastructure, data resources, organisational routines, skilled labour, and market institutions. Without digital financial services, technology markets, and producer-service platforms, AI-related strategic attention may remain symbolic or confined to individual firms. This is why AI orientation should be understood as a complementary capability rather than a stand-alone driver. It may strengthen the effect of digital infrastructure, but it is unlikely to generate broad regional upgrading outcomes in isolation.

The growing literature has examined AI adoption, automation, and digital transformation at the firm, sectoral, and macroeconomic levels. These studies show that AI and related digital technologies can affect productivity, innovation, labour demand, task structure, and firm performance. In the Chinese context, recent studies also link AI adoption or digital transformation to industrial upgrading and high-quality development. Yet much of this literature treats AI either as a firm-level technology adoption decision or as a macro-level technological trend. Less attention has been paid to the regional aggregation of firm-level AI orientation and how this regional AI orientation interacts with digital infrastructure under policy uncertainty.

The measurement of AI orientation increasingly relies on textual information from firms’ public disclosures. Corporate annual reports contain information about firms’ strategic priorities, technology adoption, and managerial attention. Loughran and McDonald (2011) show that textual information in corporate disclosures can capture economically meaningful firm characteristics. Gentzkow et al. (2019) provide a broader methodological foundation for text-as-data approaches in economics. Hassan et al. (2019) demonstrate that textual disclosures can be used to measure firm-level political risk. In the Chinese context, Yao et al. (2024) construct firm-level AI indicators from annual reports using a machine learning-based AI dictionary.

This paper builds on this text-based measurement literature by constructing a province–year measure of regional AI orientation. The measure aggregates firm-level annual report AI signals from A-share listed companies to the provincial level. Conceptually, this variable captures the extent to which the listed corporate sector within a province displays strategic attention to AI-related technologies. It is not intended to measure the total stock of AI capital or the exact amount of AI investment. Rather, it captures a broader orientation toward AI, including managerial attention, organisational readiness, and technology absorption capacity.

This distinction between AI orientation and AI investment is important. Direct AI investment may be highly concentrated in a few large firms or technology-intensive industries. It may also reflect financial capacity rather than the breadth of AI engagement across the regional firm population. By contrast, annual report-based AI orientation can capture whether AI-related ideas and technologies have diffused across a wider set of firms. A province where many listed firms discuss AI-related technologies may have a broader base of AI awareness and organisational readiness than a province where AI investment is concentrated in a small number of firms. This breadth is especially relevant for regional industrial upgrading, which depends on a collective adjustment across sectors rather than isolated technological investment by a few leading firms.

At the regional level, AI orientation may strengthen industrial upgrading through several channels. AI-oriented firms are more likely to use digital information for prediction, risk assessment, and market analysis. They may be more capable of identifying suitable technologies in technology markets, adopting digital financial tools, and demanding sophisticated producer services such as data analytics, intelligent logistics, and IT consulting. These capabilities can improve the effectiveness of digital infrastructure. In a province with strong digital finance but weak AI orientation, digital infrastructure may exist but may not be fully exploited by firms. In a province with both strong digital infrastructure and broad AI orientation, firms may be better able to transform digital resources into upgrading outcomes.

This complementarity is especially important under high EPU. Policy uncertainty makes future market conditions more difficult to predict and increases the risks of structural adjustment. AI-oriented firms may be better able to process uncertain signals, evaluate alternative strategies, and adapt to changing policy and market conditions. Yet these advantages depend on access to digital platforms, financial data, technology markets, and producer-service ecosystems. Therefore, regional AI orientation should amplify the digital economy’s uncertainty-buffering role rather than replace it.

The expected empirical pattern follows directly from this logic. If AI orientation is a stand-alone driver, then the interaction between AnnualReportAI and HighEPU should be positive and significant even without considering digital economy development. However, if AI orientation is a complementary capability, its effect should appear mainly through the triple interaction among digital economy development, AI orientation, and high EPU. In other words, AI orientation should matter most when it is combined with a strong digital infrastructure during periods of elevated policy uncertainty.

This leads to the second hypothesis:

H2: 

Regional AI orientation amplifies the positive association between digital economy development and industrial structure upgrading during high-EPU periods but does not function as an independent buffer on its own.

2.4. Observable Implications: Producer Services and Regional Upgrading

The digital-AI complementarity should be visible not only in the aggregate upgrading indicator but also in service-oriented dimensions of regional structural transformation. Industrial upgrading in China is increasingly linked to producer services such as finance, logistics, information services, business services, technology consulting, and other knowledge-intensive intermediate inputs. These activities connect manufacturing firms to data, technology, professional services, and market information, and therefore provide a useful alternative outcome for assessing whether the main result reflects a broader upgrading process rather than a mechanical movement in the tertiary–secondary ratio.

Digital infrastructure can support producer-service expansion by lowering the cost of service delivery, improving cross-regional matching, and making specialised services more accessible to firms. Regional AI orientation can increase the demand for such services because AI-oriented firms are more likely to require data analytics, intelligent logistics, IT consulting, digital finance, technology brokerage, and other knowledge-intensive inputs. Under high EPU, external producer services may also reduce the irreversibility of adjustment by allowing firms to access specialised capabilities without building all of them internally.

This reasoning leads to an observable implication: if the digital-AI complementarity supports industrial upgrading under high EPU, the triple interaction should also be positively associated with producer-service expansion. Accordingly, the empirical section reports producer-service share as the main alternative dependent variable. Other mechanism-related indicators and exploratory outcomes are not treated as confirmatory evidence and are available from the authors upon request.

This mechanism framework also helps distinguish the paper’s argument from a simple “digital economy promotes services” explanation. The central claim is not only that digital finance is correlated with a higher tertiary sector share, but that digital infrastructure and AI orientation jointly support the institutional and market conditions needed for upgrading under uncertainty. In this sense, technology markets, producer services, and digital finance usage depth are not merely outcome variables; they are components of the regional capability system through which provinces adapt to high-EPU environments.

3. Data, Variables, and Methodology

3.1. Sample and Data Sources

The empirical analysis uses an annual province–year panel for 2015–2023. The main sample contains 30 provinces and 270 observations because the provincial digital economy index does not cover Tibet. To ensure that the exclusion of Tibet does not drive the result, the model is also estimated using the Peking University Digital Financial Inclusion Index as an alternative digital economy proxy; the triple interaction remains positive and statistically significant.

The outcome variable is lnISU, defined as the logarithm of the ratio of tertiary industry value added to secondary industry value added. Control variables include lnPGDP, lnPopulation, Fiscal, Urbanisation, HumanCapital, TechExp, and Internet. Province and year fixed effects are included in all baseline specifications, and standard errors are clustered at the province level.

3.2. Variable Construction

The digital economy variable captures broader dimensions than financial inclusion alone. Following the entropy-weighting approach widely used in digital economy studies, and drawing on the multidimensional measurement framework of digital infrastructure, digital industrialisation, and industrial digitisation, this study uses a provincial digital economy index as the main explanatory variable (Zhao et al., 2020; Peng & Dan, 2023). Similar entropy-based index construction has been adopted in English-language studies on China’s digital economy, including Peng and Dan (2023). Recent Chinese studies further apply related multidimensional frameworks to policy contexts such as common prosperity; these studies are treated as supplementary background rather than as the sole measurement anchor. Because the minimum observed value is 0.0049101 and no province–year observation is zero or negative, no artificial shift such as 1 × 10⁻⁶ is added before taking logs.

Table 1. Variable definitions.

Variable	Definition/Construction	Role
lnISU	ln(tertiary VA/secondary VA)	Dependent variable
Digital	Standardised log provincial digital economy index; min Score = 0.0049101, no shift term added	Core explanatory variable
AnnualReportAI	Province–year AI orientation aggregated from listed firms’ annual report AI scores	Moderator
HighEPU	Indicator equal to 1 if the annual China EPU index (Baker et al., 2016) ≥ the sample median	Uncertainty state
Controls	lnPGDP, lnPopulation, Fiscal, Urbanisation, HumanCapital, TechExp, Internet	Covariates
ProducerServiceShare	Share of productive services in the regional economy; used in log form	Alternative dependent variable

summarises the definitions, construction, and roles of all variables.

Table 2. Descriptive statistics for core variables.

Variable	N	Mean	SD	Min	P25	Median	P75	Max
lnISU	270	0.328	0.354	−0.286	0.128	0.282	0.385	1.739
Digital	270	0.000	1.002	−2.618	−0.718	−0.004	0.695	1.864
AnnualReportAI	270	1.275	1.500	−2.731	0.203	1.332	2.275	4.584
HighEPU	270	0.556	0.498	0.000	0.000	1.000	1.000	1.000
lnPGDP	270	11.052	0.420	10.164	10.750	10.987	11.303	12.207
lnPopulation	270	8.217	0.742	6.358	7.816	8.287	8.756	9.450
Fiscal	270	0.258	0.107	0.105	0.187	0.229	0.311	0.753
Urbanisation	270	0.624	0.106	0.400	0.555	0.612	0.663	0.895
HumanCapital	270	0.023	0.006	0.010	0.019	0.022	0.026	0.044
TechExp	270	0.005	0.003	0.002	0.003	0.004	0.007	0.013
Internet	270	0.315	0.104	0.104	0.231	0.314	0.396	0.558
ProducerServiceShare	270	0.126	0.031	0.086	0.108	0.115	0.129	0.233

reports descriptive statistics for the core variables in the main sample (30 provinces, 2015–2023).

Table A1. Indicator system for the provincial digital economy index.

Primary Dimension	Secondary Dimension	Measurement Item	Direction
Digital infrastructure	Telecommunication infrastructure	Mobile phone penetration rate	+
Digital infrastructure	Telecommunication infrastructure	Length of optical cable lines	+
Digital infrastructure	User infrastructure	Internet broadband access users	+
Digital infrastructure	User infrastructure	Internet broadband access ports	+
Digital infrastructure	Software infrastructure	Number of webpages	+
Digital infrastructure	Software infrastructure	Number of domain names	+
Digital industrialisation	Digital industry	Number of employees in information transmission, computer services, and software industries	+
Digital industrialisation	Digital industry	Software business revenue	+
Digital industrialisation	Digital innovation	Number of patent applications in the current year	+
Digital industrialisation	Digital innovation	Number of patents granted in the current year	+
Industrial digitisation	Digital transactions	Share of enterprises with e-commerce transaction activity	+
Industrial digitisation	Digital finance	Digital inclusive finance index	+

Notes: The table translates the indicator system shown in the source document. All indicators are positive. The final provincial digital economy score is constructed using the entropy value method after annual normalisation.

reports the full indicator system for the provincial digital economy index based on the measurement framework provided in the source document. The system covers digital infrastructure, digital industrialisation, and industrial digitisation. All measurement items enter the entropy value index as positive indicators.

AnnualReportAI is constructed from two province–year components: the mean annual report AI score of listed firms and the share of AI-positive listed firms. AI-positive firms are defined as firms with a positive annual report AI score in a given year. The province–year index is the sum of the standardised mean AI score and the standardised AI-positive firm share. Because each firm’s AI score is computed as the mean cosine similarity of its AI-relevant sentences rather than a raw keyword count, it does not mechanically increase with report length or total disclosure volume; a longer report receives a higher score only if its AI-related content is semantically more salient, not simply because the document contains more words. The baseline aggregation is equal-weighted across listed firms because total-asset information is not available in the current provincial panel. Mean-only and firm-share-only variants are retained as sensitivity checks and are available from the authors upon request. Annual reports are sourced from the CSMAR database. Sentence segmentation uses the jieba Chinese text segmentation library, and sentence embeddings are computed using the paraphrase-multilingual-MiniLM-L12-v2 model from the sentence-transformers library.

To assess the external validity of the annual report-based AI indicator,

Table A2. External Validity of AnnualReportAI.

External Proxy	N	Pearson r	Spearman rho
ln invention patent applications	270	0.572	0.554
ln invention patent grants	270	0.664	0.660
ln AI-related investment	270	0.619	0.650
AI-related firm count	270	0.507	0.593
information-service firms	270	0.598	0.705
research-service firms	270	0.568	0.679

Notes: Correlations are calculated on the main sample (N = 270). Patent variables are in logarithmic form. The table is used as convergent validity evidence for AnnualReportAI, not as causal evidence.

reports its correlations with innovation and AI-related regional proxies. AnnualReportAI is positively correlated with invention patent applications and grants, AI-related investment, AI-related firm counts, and information/research-service firm counts. These correlations support the interpretation that the text-based measure captures meaningful regional AI orientation, while not implying causality.

3.3. Econometric Specification

The baseline model estimates the interaction among digital economy development, regional AI orientation, and the high-EPU state. The main coefficient of interest is Digital × AnnualReportAI × HighEPU. Since EPU varies at the national-year level, the main effect of EPU is absorbed by year fixed effects; identification comes from cross-provincial differences in digital economy development and AI orientation interacted with the common high-EPU state.

3.4. Predictive Machine Learning Validation

Machine learning models are used only for predictive validation. This predictive use of machine learning follows the applied machine-learning tradition in economics (Varian, 2014; Kleinberg et al., 2015; Mullainathan & Spiess, 2017; Athey, 2019), and the feature-importance diagnostics reported in Appendix A draw on feature-attribution approaches to model interpretation (Lundberg & Lee, 2017). The target variable is High_ISU, and the feature set includes the digital economy variable, AnnualReportAI, HighEPU, their interactions, and controls. To reduce province-level information leakage, model performance is evaluated using province GroupKFold and a province-level holdout split. These results are not interpreted as causal evidence. All econometric analyses are conducted in Python 3.11 using statsmodels (0.14) and linearmodels (6.0) for panel regressions, and scikit-learn (1.4) for the machine learning validation exercises.

4. Empirical Results

4.1. Baseline Results

Table 3. Baseline fixed-effects results. Dependent variable: lnISU.

Model	Digital × HighEPU	AI × HighEPU	Digital × AI × HighEPU	p(triple)	N	R2
M0 FE only	−0.0477 (0.0142)	0.0476 (0.0192)	0.0252 (0.0097)	0.0093	270	0.9580
M1 basic controls	−0.0318 (0.0171)	0.0324 (0.0211)	0.0156 (0.0090)	0.0835	270	0.9684
M2 non-overlap controls	−0.0235 (0.0137)	0.0214 (0.0126)	0.0193 (0.0089)	0.0307	270	0.9747
M3 full controls	−0.0355 (0.0151)	0.0329 (0.0115)	0.0215 (0.0080)	0.0071	270	0.9787

Notes: Province and year fixed effects are included. Standard errors clustered by province are in parentheses.

reports the baseline fixed-effects estimates. In the full M3 specification, the triple interaction is positive and statistically significant (coefficient = 0.0215, t = 2.69, p = 0.0071; N = 270; R2 = 0.9787). This indicates that the association between digital economy development and industrial structure upgrading under high EPU depends on the region’s AI orientation.

4.2. Marginal Effects

The positive triple interaction does not imply that digital economy development is universally beneficial under high uncertainty. The marginal effect analysis shows that the digital economy effect under high EPU becomes positive only when AnnualReportAI exceeds approximately 1.65. About 58.5% of sample observations have a conditional marginal effect that is less than or equal to zero. The result, therefore, supports a threshold-like, state-dependent interpretation.

4.3. Robustness Checks

Table 4. Robustness Checks: Coefficient on Digital × AnnualReportAI × EPU.

Specification	Coef.	SE	t	p	N
Exclude DFI component from digital economy	0.0217	0.0077	2.8392	0.0045	270
Drop Internet control	0.0197	0.0076	2.6044	0.0092	270
Drop Internet and TechExp controls	0.0193	0.0089	2.1604	0.0307	270
Continuous lnEPU, standardised	0.0042	0.0030	1.4183	0.1561	270

summarises the key robustness checks. The triple interaction remains significant after removing the digital financial inclusion component from the digital economy index and after dropping controls that may overlap with components of the index. The continuous lnEPU specification is not statistically significant, suggesting that the mechanism is better characterised as a high-uncertainty state effect rather than a linear response to continuous EPU changes.

Additional data-processing checks (

Table 5. Additional data-processing robustness checks.

Check	Coef.	SE	t	p	N
Winsorised continuous variables at 1/99 percentiles	0.0222	0.0079	2.816	0.0049	270
Winsorised continuous variables at 5/95 percentiles	0.0198	0.0082	2.409	0.0160	270
HighEPU: Median threshold	0.0215	0.0080	2.692	0.0071	270
HighEPU: Top 40 per cent threshold	0.0198	0.0078	2.547	0.0109	270
HighEPU: Top tertile threshold	0.0216	0.0084	2.572	0.0101	270
HighEPU: Above historical mean threshold	0.0198	0.0078	2.547	0.0109	270

Notes: All models include province and year fixed effects, the baseline controls, and province-clustered standard errors. The table is reported as supplementary robustness evidence.

) further show that the main result is not driven by extreme observations or by an arbitrary high-EPU cutoff. Winsorising continuous variables at the 1/99 and 5/95 percentiles leaves the triple interaction positive and significant. Alternative high-EPU definitions based on the top 40%, top tertile, and historical mean thresholds also yield positive and statistically significant coefficients.

4.4. Alternative Dependent Variables

The main alternative dependent variable is the logarithm of producer-service share, which captures a service-upgrading dimension closely related to productive specialisation. The triple interaction is positive and significant for this alternative outcome (

Table 6. Alternative dependent variables.

Outcome	Coef.	SE	p	Within R2	Placement
Main alternative DV: ln ProducerServiceShare	0.0152	0.0042	<0.001	0.5114	main_text
Appendix alternative DV: Upgrade_theil	0.0008	0.0005	0.1176	0.2980	appendix

). The Theil-type rationalisation measure is reported in Appendix A and is not statistically significant.

4.5. Endogeneity-Related Checks

The fixed-effects design mitigates time-invariant provincial heterogeneity and common annual shocks, but it does not establish definitive causal identification. We therefore report several checks that help reduce, but do not eliminate, endogeneity concerns (

Table 7. Endogeneity-related checks.

Specification	Triple Coef.	p	N	Clusters
Lagged log digital economy	0.0157	0.0205	240	30
Lagged AnnualReportAI	0.0180	0.0094	240	30
Lagged digital economy and AI	0.0185	0.0091	240	30
Add lagged dependent variable	0.0053	0.1761	240	30
Province-specific linear trends	0.0020	0.6665	270	30
Exclude four municipalities	0.0192	0.0188	234	26

). Lagged digital economy and lagged AI specifications remain significant, and the result also holds after excluding municipalities. Specifications with a lagged dependent variable or province-specific linear trends are weaker and are interpreted cautiously.

4.6. Policy Exposure Supplement

As supplementary evidence, we use digital-infrastructure policy exposure based on Broadband China, smart-city, and big-data pilot programmes. The most useful evidence comes from Broadband China (

Table 8. Policy exposure supplement.

Policy Check	Coef.	p	Interpretation
BigData pre-trend			pre p < 0.10: 2; pre p < 0.05: 2
Broadband pre-trend			pre p < 0.10: 0; pre p < 0.05: 0
SmartCity_log_count	0.0048	0.7223	Reduced-form DDD
BigData_log_count	0.0240	0.0887	Reduced-form DDD
Broadband_ever	0.0264	0.0783	Reduced-form DDD

): its event-study estimates show no significant pre-trends across the four pre-treatment leads, and the reduced-form interaction is positive, although only marginally significant. Big-data pilot exposure also shows a positive marginal association, but treated provinces were already more developed before the pilot, and their pre-trends are less clean. These policy results are therefore used only to mitigate endogeneity concerns, not as definitive causal identification.

4.7. Machine Learning Validation

The machine learning results, presented as a predictive validation exercise, show moderate performance under province GroupKFold (Figure 2;

Table 9. Machine learning performance under province GroupKFold.

Model	AUC Mean	AUC SD	F1 Mean	F1 SD	Accuracy Mean
GB	0.6162	0.0886	0.5587	0.0492	0.5852
Logit	0.5755	0.1830	0.4662	0.1790	0.5222
RF	0.5833	0.1251	0.5884	0.1208	0.6148

). Gradient boosting obtains the highest average AUC, while random forest obtains the highest average F1 score. These results indicate that the digital economy and AI-related features contain predictive information for High_ISU, but they are not used for causal interpretation.

5. Discussion

The evidence supports a state-dependent complementarity between the digital economy and regional AI orientation under policy uncertainty, extending the existing literature in several respects. First, the digital economy should be interpreted as a broad regional capability rather than only as digital financial inclusion. The main result remains significant when a broader provincial digital economy index is used, and it also survives the removal of the digital financial inclusion component. Second, the effect is conditional. Digital economy development does not uniformly promote upgrading under high uncertainty; its positive association emerges mainly when regional AI orientation is sufficiently high. This conditionality distinguishes the present findings from average effect studies that document a generally positive link between digital finance and industrial upgrading (H. Shen et al., 2024; Yang, 2023; Zheng et al., 2023). Those studies pool high- and low-EPU periods and treat the digital economy as uniformly beneficial, and the marginal effect analysis here shows that approximately 58.5% of sample observations have a non-positive conditional effect. The result is also consistent with Ye et al. (2024) and Zhu et al. (2022), who find that EPU constrains industrial upgrading at the provincial level; the present paper extends their work by showing that digital–AI complementarity may help mitigate this constraint under high-EPU conditions. The AI orientation finding complements Zhang (2025), who documents a positive AI–upgrading association at the provincial level with inclusive finance as a moderator, by showing that AI orientation matters primarily as a complement to digital infrastructure rather than as a stand-alone driver.

The heterogeneity results are treated as exploratory. Formal group-difference tests do not support strong claims that one region has a systematically larger effect. In particular, the Eastern subsample has only ten clusters, and the wild-cluster bootstrap confirms that its triple interaction is not statistically significant. The paper, therefore, avoids presenting regional heterogeneity as a core confirmatory result.

The endogeneity checks also require caution. Lagged-variable and policy exposure results are consistent with the main interpretation, and the Broadband China pre-trend evidence helps address concerns about pre-existing trends. However, policy pilots were not randomly assigned, and the IV estimates are not strong. In addition, the sample period spans only nine years (T = 9), which limits the degrees of freedom available for estimating multiple interaction terms, mechanism regressions, and heterogeneity splits within a fixed-effects framework. With a short time dimension, the stability and precision of the estimated interaction effects should be interpreted cautiously, and readers should note that the reliability of these estimates may improve as future data extend the panel. Accordingly, the findings are interpreted as conditional associations supported by supplementary evidence, rather than as definitive causal estimates.

6. Conclusions and Policy Implications

This paper examines whether the digital economy helps provincial economies sustain industrial structure upgrading under economic policy uncertainty and whether regional AI orientation strengthens this role. Using a balanced panel of 30 Chinese provinces from 2015 to 2023 and the standardised logarithm of a provincial digital economy index, the results show a positive and statistically significant triple interaction among digital economy development, regional AI orientation, and high EPU. The evidence supports a state-dependent complementarity: digital economy development is more closely associated with industrial upgrading under high uncertainty when regional AI orientation is high.

The marginal effect analysis is central to this interpretation. The digital economy effect under high EPU turns positive only after AnnualReportAI exceeds approximately 1.65, and a majority of the observations remain at or below zero conditional marginal effect. This means that the policy implication is not simply to expand digital infrastructure in isolation, but to coordinate digital economy development with firms’ AI-related capabilities and organisational readiness.

Several tentative policy implications follow, subject to the observational nature of the evidence. First, digital economy infrastructure may serve as a resilience resource during high-uncertainty periods, but its effectiveness appears to depend on complementary regional capabilities; policies aimed at deepening digital infrastructure alone may yield weaker upgrading effects without parallel investments in AI-related organisational capacity. Second, AI policy may benefit from emphasising broad organisational adoption, managerial attention, and application capacity rather than concentrating support on a small number of leading AI firms. Third, producer-service ecosystems, technology markets, and digital platforms may be most effective when developed as connected institutions that allow digital and AI capabilities to translate jointly into structural upgrading. Finally, policymakers should recognise that the evidence is conditional rather than universal: the findings do not support a blanket claim that digital economy development automatically promotes upgrading in all regions or under all conditions.

This study has limitations. Province and year fixed effects, lagged specifications, and policy exposure checks mitigate but do not eliminate endogeneity concerns. The sample covers only nine years (T = 9), which constrains the degrees of freedom for estimating interaction effects and conducting subgroup analyses within a fixed-effects framework; this short time dimension means that the reliability and stability of the estimated effects should be assessed against future data as additional years become available. The dependent variable, ln(tertiary VA/secondary VA), captures the sectoral composition of output rather than the productivity, sophistication, or technological content of industrial activity; a province can exhibit apparent “upgrading” through arithmetic expansion of services without genuine productivity improvement. AnnualReportAI is based on A-share listed firms and may not fully represent unlisted enterprises, particularly small and medium-sized firms whose AI adoption patterns may differ substantially. The text-based measure captures disclosed strategic attention to AI, which may partly reflect managerial framing or signalling incentives rather than realised AI capabilities; it is therefore best understood as AI orientation rather than verified AI adoption. The policy exposure tests are supplementary because policy pilots were not randomly assigned. Future research could combine firm-level AI adoption data, more detailed policy shocks, and micro-level industrial upgrading outcomes to sharpen causal identification.