Venture capital and R&D in AI shift jobs away from industry but are linked to higher aggregate employment in a 14-country 2013–2023 panel; system-GMM estimates point to sectoral reallocation, though findings rely on imputed data and internal instruments.

Main Finding

Corporate venture-capital investment in AI and higher R&D intensity (2013–2023, 14 advanced and South American economies) are associated with a decline in industrial employment shares but with an increase in overall employment — implying a heterogeneous sectoral reconfiguration (task- and sector‑level shifts) rather than net job destruction.

Key Points

• Study period and sample: balanced panel 2013–2023 for 14 countries (advanced economies + South America).
• Dependent variables: sectoral employment shares (industry, services, agriculture) as percent of total employed.
• Key regressors: VC investment in AI (OECD.AI), R&D spending (% of GDP), plus controls (FDI inflows % GDP, GDP level).
• Main econometric approach: dynamic panel estimators to address persistence and endogeneity — Anderson–Hsiao and system GMM (Blundell–Bond).
• Validity checks: instrument and serial-correlation diagnostics via Sargan–Hansen and Arellano–Bond tests.
• Principal empirical result (Blundell–Bond): VC in AI and R&D are negatively associated with industrial employment share and positively associated with aggregate employment — consistent with task reallocation (destruction of some routine industrial jobs, creation of jobs elsewhere, often higher-skilled).
• Results interpreted as heterogeneous reallocation across sectors rather than aggregate employment loss.
• Data limitations noted: missing observations handled by linear interpolation imputation; potential measurement issues using VC flows as a proxy for AI adoption.

Data & Methods

• Data sources: World Bank (employment shares, R&D, FDI, GDP), OECD.AI (VC investments in AI), ILO referenced for labor metrics.
• Panel: balanced country–year panel, 14 countries, 2013–2023.
• Variables:
  • Dependent: employment shares in services, industry, agriculture (% of employed).
  • Independent: VC investment in AI (USD billions), R&D (% GDP).
  • Controls: FDI inflows (% GDP), GDP (USD billions).
• Missing data treatment: linear interpolation in R to construct balanced panel.
• Estimation strategy:
  • Static description and time-series plots + descriptive statistics.
  • Dynamic panel estimation with Anderson–Hsiao (instrumenting lagged dependent) and system GMM (Blundell–Bond) to control for:
    • persistence in employment shares,
    • endogeneity of regressors (reverse causality from employment to investment),
    • unobserved country fixed effects.
  • Instrument validity and serial correlation assessed with Sargan–Hansen and Arellano–Bond tests; reported validation consistent with model specification.
• Robustness: use of two dynamic estimators; primary inference from Blundell–Bond system GMM.

Implications for AI Economics

• Heterogeneous effects matter: AI investment reallocates employment across sectors — declines in industrial employment coexist with net employment gains elsewhere (likely services and high‑skill occupations). Aggregated employment metrics alone can mask important sectoral shifts.
• Measurement: VC flows to AI (OECD.AI) capture private investment signals but may understate public, incumbent firm, or informal-sector adoption — future work should combine firm‑level adoption measures and public R&D.
• Policy priorities:
  • Active labor-market policies (reskilling/upskilling) to manage transition from routine industrial tasks to higher‑skill roles.
  • Social-protection and transition-support mechanisms, especially in South America where informality and weaker digital infrastructure raise displacement risk.
  • Strengthening digital infrastructure and education to allow South American economies to capture employment gains from AI rather than only suffer job displacement in routine sectors.
• Research design guidance: dynamic panel GMM is appropriate for studying temporal and endogenous links between AI investment and employment structure, but results hinge on instrument quality and data completeness — researchers should test alternative proxies for AI adoption and exploit more granular (occupation/task, firm) data.
• Equity and distributional concerns: sectoral reconfiguration can exacerbate inequality (middle-skill job losses, high-skill gains); economics of AI should integrate distributional outcomes (wages, employment stability, informality) into welfare assessments.

Limitations noted by the authors that bear on interpretation: interpolation of missing data, limited sample size/coverage (14 countries), and the use of VC AI flows as a partial proxy for broader AI adoption. Future work should extend country coverage, use firm‑ or task‑level adoption indicators, and examine heterogeneous impacts by worker characteristics (skill, gender, age).