AI activity lifts demand for China’s education and training market primarily in bullish conditions, while policy uncertainty tends to depress it during calmer periods. Risk spillovers from AI to training are larger and more volatile than those from policy uncertainty, implying policymakers should balance AI-driven opportunity with market stability.

Main Finding

The paper finds quantile-dependent, nonlinear interactions among China's artificial intelligence (AI) activity, economic policy uncertainty (EPU), and the education & training market (ETM). Key results: (1) unidirectional quantile causality links exist between AI→EPU, AI→ETM, and EPU→ETM; (2) AI exerts a positive predictive effect on ETM mainly in bullish (upper‑quantile) market states, while EPU’s negative predictive effect on ETM concentrates in periods characterized as policy stability (particular quantiles); (3) AI and EPU interact asymmetrically across market states (AI tends to increase EPU in bullish markets; EPU tends to promote AI during periods of economic / policy stability); and (4) risk‑spillover analysis (GARCH‑CQR‑CoVaR) shows strong asymmetry: upside risk spillovers dominate downside ones, AI→ETM spillovers are high‑volatility and extreme, whereas EPU effects are milder but more persistent.

Key Points

• Causality-in-quantiles: Using a nonparametric quantile causality test, the authors document one‑way causal relations (in quantile sense) for AI→EPU, AI→ETM, and EPU→ETM (both mean and variance dimensions are considered).
• Quantile dependence: The cross‑quantilogram analysis reveals that predictive relationships depend on the quantile pairings:
  • Positive AI→ETM prediction is strongest in upper quantiles (bullish market states).
  • Negative EPU→ETM prediction is concentrated in certain lower/mid quantiles interpreted as periods of policy stability.
  • Cross-effects: AI increases EPU in bullish states; EPU boosts AI in more stable states—showing state-dependent feedback.
• Tail risk and asymmetry: The GARCH‑CQR‑CoVaR framework uncovers asymmetric risk transmission:
  • Upside (extreme positive) spillovers are substantially stronger than downside spillovers.
  • AI produces more volatile, extreme tail impacts on ETM; EPU’s tail impacts are steadier and longer lasting.
• Policy message: Policymakers and ETM providers must account for quantile/state dependence and asymmetric tail risks when designing interventions or business strategies.

Data & Methods

• Data (as described): Time series indices representing China’s AI activity, China economic policy uncertainty (EPU), and measures of the education & training market (ETM). (The paper applies methods compatible with daily/weekly/monthly lags; specific sample period and index construction are reported in the full paper.)
• Main empirical methods:
  • Nonparametric causality‑in‑quantiles test (Balcilar et al., 2016): tests whether one series has causal effects on different quantiles of another series’ conditional distribution (in mean and variance).
  • Cross‑quantilogram (Han et al., 2016): measures serial and cross‑quantile dependence between two series’ quantile‑hit processes across different lags; inference via smooth bootstrap.
  • GARCH‑CQR‑CoVaR (Tian et al., 2022) / ARMA‑EGARCH margins + copula quantile regression: fits ARMA‑EGARCH models to marginal series, links margins via copula to estimate conditional quantile curves and compute CoVaR/UCoVaR for downside/upside spillover analysis; captures asymmetric volatility and tail dependence.
• Inference procedures: kernel‑based/nonparametric test statistics, bootstrap resampling (1000 replications for cross‑quantilogram), and copula estimation for conditional quantiles.

Implications for AI Economics

• Policy design must be state‑aware: The effects of AI adoption and EPU on ETM vary by market state (quantile). Policies that are effective in bullish conditions (e.g., incentives for AI‑driven educational innovation) may be less effective or have different consequences in downturns or stable policy regimes.
• Regulatory clarity reduces harmful spillovers: Rapid AI expansion can raise policy uncertainty in bullish states. Clear, forward‑looking regulation of AI (privacy, labor impacts, data governance in education) can limit EPU amplification and stabilize ETM outcomes.
• Risk management for ETM providers: Education and training firms should incorporate asymmetric tail risks into planning. Because AI→ETM spillovers are extreme and volatile, providers need buffers (liquidity, diversified offerings, scalable delivery) and stress‑testing for upside shocks as well as downside shocks.
• Targeted support during policy shifts: EPU shows persistent effects even if moderate in intensity. During periods of policy change or uncertainty, targeted support (temporary subsidies, retraining programs, regulatory transition assistance) can sustain ETM provision and labor market upskilling.
• Research and measurement: The quantile‑based approach highlights heterogeneity missed by mean‑based studies. Future empirical work in AI economics should use quantile/time‑state methods, consider copula tail dependence, and attempt micro‑level linkages (firm/household) to unpack mechanisms.
• Strategic interplay between AI and EPU: Policymakers aiming to encourage AI adoption in education should balance innovation support with communication and policy certainty to avoid unintended increases in EPU that could destabilize the ETM.

Caveats: findings are quantile‑ and index‑specific and rely on chosen indices, copula specifications, and sample period. For operational use, replicate with updated data and, where possible, complement with microdata on firms, institutions, or households to identify mechanisms.