Across 15 OECD economies, higher AI spending is linked to lower exports in the short term but to more stable export performance; meanwhile, R&D boosts exports primarily in low- and medium-performing exporters, with effects fading at the top end.

Main Finding

AI investments and R&D expenditures affect export performance in heterogeneous, non‑linear ways across OECD countries (2013–2023). Specifically, increased AI investment is associated with short‑run suppressing effects on export levels but with reduced export volatility (a stabilizing role). R&D spending has stronger positive effects for low‑ and mid‑level exporters and weaker effects at high export performance quantiles. The results imply that one‑size‑fits‑all, mean‑based policy prescriptions are inappropriate; AI and R&D policies should be country‑ and performance‑sensitive.

Key Points

• Sample and scope: 15 OECD countries (France, Greece, Spain, Germany, Italy, UK, Sweden, Norway, Canada, Belgium, Luxembourg, Portugal, Türkiye, Poland, USA), annual panel 2013–2023. Dependent variable: exports of goods and services (reported relative to GDP). Main regressors: AI investment (OECD measure of AI & data startup investment) and R&D expenditures (% GDP). Controls: GDP growth and inflation.
• Estimator: Method of Moments Quantile Regression (MMQR, Machado & Silva 2019) with location–scale decomposition; estimation reported across quantiles (25th to 90th) to reveal heterogeneous impacts on both levels and volatility of exports.
• Pre‑estimation diagnostics: strong cross‑sectional dependence (Pesaran CD and CDw+), significant slope heterogeneity (Pesaran & Yamagata), panel unit‑root evidence consistent with I(1) behavior prior to differencing (Pesaran 2007). These motivated heterogeneous, quantile‑based methods.
• Main empirical patterns:
  • AI_inv: tends to depress export levels in the short run at several quantiles (interpreted as short‑term adjustment/transition costs), but is associated with lower conditional volatility (stabilizing exports).
  • R_D: positive and significant effects concentrated at lower and middle export quantiles; effects decline or become weak at higher export quantiles.
  • Heterogeneity: coefficients vary substantially across countries and across conditional export distribution — average (mean) estimates would mask these differences.
• Policy prescription emphasized by the paper: AI and R&D policies should be integrated, differentiated by country export performance and capacity, and designed to manage short‑term adjustment costs while leveraging complementarities between R&D and AI.

Data & Methods

• Data sources:
  • Exports of goods & services (% of GDP): World Bank.
  • AI investments: OECD dataset on investments in AI and data ecosystems (proxy: investment flows into AI/data start‑ups).
  • R&D expenditures (% GDP), GDP growth, inflation: World Bank.
• Summary statistics (paper):
  • I_Expo mean ≈ 1.60 (with sd ≈ 0.255); AI_inv mean ≈ 2.32 (sd ≈ 1.27); R_D mean ≈ 3.03 (sd ≈ 0.81).
• Econometric strategy:
  • Use MMQR to estimate quantile‑specific effects and to decompose impacts into location (level) and scale (volatility) components, allowing coefficients to vary across the conditional distribution of exports.
  • Estimation performed for quantiles from the 25th to the 90th percentile.
  • Tests performed for cross‑sectional dependence (Pesaran CD, Fan et al. CDw+), slope heterogeneity (Pesaran & Yamagata), and unit roots (Pesaran).
• Limitations noted or implied: relatively small cross‑section (15 countries), potential measurement issues with the AI investment proxy, limited ability to establish causal identification (possible endogeneity), and the study focuses on country‑level aggregates rather than firm/sector microdata.

Implications for AI Economics

• Theory and modeling:
  • AI effects on trade are distributionally heterogeneous — models should allow heterogeneous treatment effects (quantile methods, heterogeneous panels) rather than relying solely on mean effects.
  • Important trade‑off: AI can impose short‑run adjustment costs (reducing export levels initially) while improving stability/volatility of export performance. Economic models of technology adoption and trade should incorporate both level and variance channels.
  • Complementarity with R&D: AI amplifies returns to prior knowledge/capacity and R&D accelerates the productive use of AI — empirical and theoretical work should model interaction effects and lags more explicitly.
• Policy:
  • Design nuanced, country‑specific AI/trade policies: countries with lower export performance may benefit more immediately from R&D support, while higher exporters may face transitional disruption from rapid AI adoption and thus need policies to manage adjustment (retraining, financing, phased adoption).
  • Stabilization role of AI suggests policy objectives should include export predictability and resilience (not only growth); public support can target smoothing costs (e.g., adjustment funds, targeted credit, digital infrastructure).
  • Complementary investments matter: digital infrastructure, skills, regulatory frameworks, and innovation ecosystems can strengthen positive effects and reduce transition costs.
• Empirical research agenda:
  • Move towards causal identification (instrumental variables, difference‑in‑differences with credible shocks, firm‑level panel IV) to unpack short vs. long‑run effects of AI on exports.
  • Disaggregate analyses: sectoral and firm‑level studies to see where AI raises product quality/complexity vs. where it causes substitution/disruption.
  • Explore interaction terms and dynamic specifications (lags of R&D and AI, diffusion dynamics) to capture delayed innovation benefits.
  • Refine AI measurement: combine investment flows with adoption/use indicators, patent activity, AI‑enabled product shares, and skills measures.

Caveats: results rely on a country‑level AI investment proxy and an observational panel (2013–2023); endogeneity and measurement error remain concerns and are opportunities for follow‑up work.