The resource based view argues that sustainable advantage arises when actors access and deploy resources that are valuable, rare, difficult to imitate, and not easily substituted (31). Public policy can reshape the distribution of such resources by converting scarce technological capabilities into usable inputs for specific populations (32). In the context of youth entrepreneurship, AI policy operates as a resource provision and coordination device (33). It lowers the access cost to data, compute, and algorithmic services, standardizes interfaces to digital tools, and connects young founders to complementary assets such as mentoring, market intelligence, and certification. These policy functions directly address early stage constraints that young adults face, including limited analytical capacity, narrow information channels, and weak data support for decision making (34). By building inclusive access platforms and regulating service quality, AI policy turns advanced analytics, digital process tools, and data validation into common capabilities rather than privileges of incumbents. This equalization aligns with sustainable development objectives that emphasize opportunity equality and diffusion of productivity gains across regions and social groups.

AI policy also reorganizes the entrepreneurial ecosystem from a fragmented landscape of isolated services into an integrated system that combines technology and support services (35). This system reduces trial and error costs, lowers perceived risk, and improves the stability of project quality. Critically, it weakens traditional barriers that channel opportunities toward resource rich groups, thereby broadening the feasible set of ventures for young entrepreneurs from diverse backgrounds. On this basis I expect a positive relationship between AI policy exposure and youth entrepreneurial entry.

Therefore, research hypothesis H1 can be obtained:

H1. AI policy implementation increases the probability that young adults choose entrepreneurship.

The effect of AI policy on youth entrepreneurship does not arise solely from generic resource provision. It works through targeted relief of binding constraints that suppress entrepreneurial entry among young adults. Two constraints are central for both inclusion and sustainability. The first is high non-productive transaction cost associated with reliance on informal relations—often referred to as Guanxi in the Chinese context—to obtain permits, information, and partnerships. The second is limited access to finance due to collateral based lending practices that disadvantage young people with thin credit files and few tangible assets.

AI policy reduces the first constraint by digitizing and standardizing public services and market intermediation (36). By digitizing administrative approvals and standardizing market matching, AI policy reduces the necessity of Guanxi. Drawing on the theory of Institutional Voids (37), informal networks often serve as imperfect substitutes for weak market infrastructure. By establishing transparent digital platforms, AI policy reduces the ambiguity that necessitates relational spending, effectively lowering the transaction costs of entry. In many emerging economies, interpersonal connections serve as an informal substitute for weak institutions. AI portals replace this ‘relational tax' with transparent, rule-based digital interfaces, allowing young founders to obtain key inputs through formal channels. Resources can thus be reallocated from costly relationship maintenance to core productive activities.

AI policy alleviates the second constraint by encouraging AI enabled credit assessment and risk management in financial institutions (38). From the perspective of Information Asymmetry theory, traditional lending relies heavily on collateral because lenders lack verifiable information about borrowers. AI-enabled algorithms bridge this gap by processing ‘soft information' (e.g., behavioral footprints and transaction histories) to generate predictive risk profiles. This allows lenders to substitute digital information for physical collateral, directly benefiting youth with ‘thin' credit files (39). AI policy alleviates the second constraint by promoting the adoption of digital credit scoring technologies that utilize alternative data. Traditional lending relies heavily on fixed assets as collateral, disadvantaging youth with ‘thin credit files.' AI-enabled credit scoring translates intangible signals (e.g., technical skills, online behaviors) into measurable risk assessments. This effectively substitutes physical collateral with information capital, relaxing the binding liquidity constraints that typically deter entry for asset-poor entrepreneurs.

Therefore, research hypothesis H2 can be obtained:

H2. AI policy promotes youth entrepreneurship by reducing spending on relationship maintenance and by relaxing credit constraints.

Figure 1 presents the conceptual framework that links AI policy to youth entrepreneurial entry through environmental optimization, with the two mechanisms operating as mediators.

I draw on the China Family Panel Studies³ administered by Peking University. Five survey waves are used: 2014, 2016, 2018, 2020, and 2022. The target population is young adults. I therefore restrict the age range to 18–44 in every observed wave. I remove observations with missing values on the outcome, the policy indicator, or any control variables. I then construct a balanced panel by retaining individuals observed in all five waves after cleaning. I acknowledge potential concerns regarding attrition bias inherent in using a balanced panel over five waves. While this approach ensures I track the same individuals over time, it may exclude those who drop out due to migration or other unobserved factors. However, I mitigate this by employing individual fixed effects to control for time-invariant heterogeneity that might correlate with attrition, and prior studies using CFPS suggest core estimates remain robust across balanced and unbalanced samples.

The final dataset constitutes a panel consisting of a total of 3,794 observations pooled across five survey waves (2014, 2016, 2018, 2020, and 2022). City level covariates are taken from the China City Statistical Yearbook for the corresponding years. I acknowledge that the selection of AI pilot cities is not random; it often correlates with pre-existing industrial and research capacities. This potential endogeneity could threaten validity if pilot cities were already on a different growth trajectory. However, our difference-in-differences design addresses this by controlling for time-invariant city characteristics (via fixed effects) and common time shocks. As long as the trends in entrepreneurship would have been similar in the absence of the policy—a condition supported by our parallel trends test—the non-random selection does not bias the estimator of the treatment effect.

Explained variable. Youth entrepreneurship (Job) is coded from the CFPS main work item. Respondents whose main work is private business owner, individual industrial and commercial household, or other self-employment are coded as entrepreneurs with a value of 1. All others are coded 0. This yields a binary indicator that captures entrepreneurial entry status in each wave.

Explanatory variable. Exposure to the AI pilot policy is coded at the city by year level and assigned to individuals by their city of residence. For cities designated as pilots, the indicator equals 1 in the years after designation and 0 in pre designation years. For non-pilot cities, the indicator equals 0 in all years.

Control variables. Following prior work on AI, entrepreneurship, and inclusive development (40, 41), I include rich individual, household, and regional covariates. Individual characteristics are gender (gender), marital status (married), residence location (resident), self-reported health (health), and education (edu). Gender equals 1 for male and 0 for female. Marital status equals 1 if married and 0 otherwise. Residence equals 1 for rural and 0 for urban. Health is measured on a five point scale where higher values indicate better health. Education is an ordered scale where 0 denotes illiterate or semi-literate, 1 = primary school, 2 = middle school, 3 = high school or equivalent vocational track, 4 = junior college, 5 = bachelor, 6 = master, and 7 = doctorate. Household characteristics are household size (size) and the logs of household income (lnincome) and consumption (lnconsume), both calculated from CFPS items. City level controls are the log of GDP per capita (lnpgdp) and the urbanization rate (urb) for the respondent's city and year.

Mediating variables. To study transmission channels that are central to inclusion and sustainability, I construct two mediators. Relational spending is measured as the natural log of social capital outlays (soc) as reported in CFPS. Credit constraint (limit) is a binary indicator that equals 1 if the household reports being credit constrained and 0 if not. Related variable assignment as shown in Table 2.

Building on the theoretical framework, I estimate a two way fixed effects difference in differences model to identify the effect of AI pilot exposure on youth entrepreneurial entry. The baseline specification is

Here i indexes individuals and t indexes survey years. The dependent variable JOB_{i, t} is a binary indicator of youth entrepreneurship coded from the CFPS main work item. The key explanatory variable AI_{i, t} indicates whether the respondent's city of residence i is designated as an AI pilot in year ttt after the adoption year. The vector Control_{i, t} collects individual, household, and city level controls defined above. The term μ_i captures time invariant individual heterogeneity and δ_t captures common shocks by year. The disturbance ε_{i, t} is idiosyncratic. I report heteroskedasticity robust standard errors clustered at the individual level to address serial correlation within persons. Under parallel trends, α is a constant term, the coefficient β identifies the average change in the probability of entrepreneurial entry associated with AI pilot exposure, conditional on covariates and fixed effects.

Descriptive statistics for all variables in the balanced panel are reported in Table 3.

Table 4 reports the two way fixed effects estimates linking AI pilot exposure to the probability of youth entrepreneurship. In Model 1, which includes individual and year fixed effects but no additional controls, the coefficient on the AI indicator is 0.137 with a robust standard error of 0.012, significant at the one percent level. In Model 2, which adds the full set of individual, household, and city covariates, the coefficient rises to 0.484 with a robust standard error of 0.079, again significant at the one percent level. Given the binary outcome, these coefficients can be read as percentage point changes in entry probability. The controlled estimate implies an increase of 48.4 percentage points for young adults residing in pilot cities relative to comparable peers in non-pilot cities. This substantial magnitude reflects the ‘big push' nature of the AI pilot policy. These zones do not merely offer a single subsidy but implement a comprehensive ecosystem overhaul—simultaneously lowering informational, financial, and administrative barriers. For young potential entrepreneurs facing high initial constraints, this multi-dimensional relief can trigger a non-linear ‘tipping point' in occupational choice, explaining the large estimated effect.

The increase in magnitude from Model 1 to Model 2 suggests negative confounding in the uncontrolled specification, where omitted factors that depress entrepreneurship are more prevalent in pilot locations and are absorbed once controls are included. Taken together, the baseline results provide clear support for H1 that AI policy promotes youth entrepreneurial entry.

Building on the theoretical framework, I evaluate two mediators that are central to inclusive and sustainable development: reductions in relational spending and alleviation of credit constraints. Structural equation models with individual and year fixed effects indicate statistically significant indirect effects for both channels. For relational spending as shown in Tables 9, 10, the estimated indirect effect is −0.004 with p-values around 0.007 to 0.008 across Delta, Sobel, and Monte Carlo tests. For credit constraints, the estimated indirect effect is −0.003 with p-values around 0.029–0.033. Negative signs indicate that AI policy reduces relational outlays and the incidence of binding credit constraints, which in turn increases the probability of entrepreneurial entry among young adults. Therefore, H2 is proven. Meanwhile, in terms of model validity, the structural equation modeling results demonstrate a satisfactory model fit, with a Comparative Fit Index (CFI) of 0.91 and a Root Mean Square Error of Approximation (RMSEA) of 0.048, validating the hypothesized paths.

These mediation results map closely to international evidence that AI enhances information production and credit risk assessment, which expands financial inclusion in emerging economies, improves peer to peer screening quality, and supports ethical and capability building dimensions of entrepreneurship education that reduce reliance on informal networking (46, 47). They also connect to findings that AI enabled risk management fosters entrepreneurship and performance in financial institutions and to evidence on the role of AI learning in strengthening student entrepreneurs' outcomes (48). The mechanisms therefore provide a micro level bridge between technology policy and the SDGs agenda on decent work and reduced inequality.

Based on this, H2 has been proven.

I examine heterogeneous impacts along three dimensions that are salient for inclusive and sustainable development: health status, digital access, and urban–rural residence. Results are summarized in Table 11 and visualized in Figure 4.

First, governments should leverage AI-driven diagnostic algorithms to upgrade entrepreneurship support services. Instead of merely digitizing application forms, policymakers should build intelligent service stacks that use predictive models to automatically match young entrepreneurs with suitable policy incentives based on their profiles. This approach replaces manual eligibility screening with algorithmic diagnostics, thereby reducing administrative friction and non-productive time costs for applicants.

Second, financial regulators should encourage the adoption of machine learning-based credit scoring models to broaden access to finance. Traditional collateral-based lending systematically excludes youth with “thin” credit files. By integrating non-traditional signals—such as digital transaction footprints and behavioral data—into AI risk assessment models, lenders can accurately evaluate creditworthiness without physical collateral. This shift from asset-based to data-based lending is crucial for relaxing the binding liquidity constraints facing young founders.

Third, support measures must be tailored to heterogeneous needs. For rural youth, AI-enabled remote mentoring and market-matching platforms can bridge the spatial divide. For those with health limitations, virtual incubation services ensure that physical constraints do not preclude entrepreneurial entry. Finally, a robust governance framework—incorporating “Explainable AI” (XAI) and algorithmic fairness audits—is essential to ensure that automated decision-making does not reinforce existing inequalities.

Fourth, institutionalize multi stakeholder governance with strong safeguards and continuous evaluation. Establish coordination mechanisms that link technology, labor, finance, taxation and education agencies with industry associations, universities and civil society. Encourage large AI firms to open technical resources through shared platforms and voucher schemes for compute, application programming interfaces and training, and to co run project studios with universities that translate research into venture ready products. Address potential risks by establishing ‘Explainable AI' (XAI) auditing mechanisms. Governance frameworks must ensure that automated approvals do not systematically disadvantage vulnerable groups through algorithmic bias. Privacy protection and clear documentation should be mandated for all AI systems used in credit and public intermediation. Implement specific Key Performance Indicators (KPIs) to monitor success beyond simple entry rates. Recommended KPIs include the ‘Credit Conversion Rate' (percentage of applicants without collateral receiving loans), ‘Digital Service Usage Rate' among rural youth, and the ‘Survival Rate' of AI-supported ventures over 3 years, including youth entry rates, survival, access to formal credit, usage of digital public services and distribution across health, access and location strata. Where feasible, embed randomized or staggered rollouts to enable credible impact assessment and course correction. This governance architecture ties innovation to accountability and ensures that AI policy advances the Sustainable Development Goals on decent work, industry innovation and reduced inequalities.

While this study provides robust micro-evidence, several limitations should be noted. First, regarding the identification strategy, I acknowledge that the selection of AI pilot cities is not random; these zones are often established in regions with stronger pre-existing industrial bases. While I employed a Difference-in-Differences (DID) design with comprehensive city fixed effects to control for time-invariant heterogeneity, and the parallel trend tests support the validity of the control group, unobserved time-varying shocks correlated with pilot designation could still theoretically pose identification challenges. Future research could explore natural experiments or instrumental variable strategies to further isolate causal effects from regional economic trends.

Second, the analysis relies on a balanced panel dataset from the CFPS. Although this allows for tracking individual-level changes, the sample size and the restriction to a balanced panel may introduce attrition bias if individuals who drop out of the survey differ systematically from those who remain. Future research utilizing administrative registration data with full population coverage would offer higher precision.

Third, the transmission mechanisms—relational spending and credit constraints—are measured using survey proxies. Entrepreneurial status is measured as a binary main work indicator, which may under-capture emerging forms of informal digital gig work. Future studies could benefit from cooperating with banks or digital platforms to access granular data on loan applications and approval logs, providing more direct evidence of how AI alters credit allocation mechanics.