One in four US-trained STEM PhDs leave within 15 years—AI and quantum doctors are most likely to go—yet their emigrant work still disproportionately benefits the United States: even after moving abroad, these scientists’ patents are cited in the US five times more than in destination countries.

Main Finding

Using a near-population dataset of US-trained STEM PhDs (1980–2024) linked to their subsequent publications and patent citations, the authors find that about 25% of scientifically active US-trained STEM PhD graduates emigrate within 15 years of graduation (≈10% within 5 years). Emigration rates vary by field (higher in AI, quantum, and many engineering fields; lower in life/biomedical sciences), have been stable for decades, and have recently concentrated more on Asian destinations (notably China). Even when graduates leave, the US continues to be the single largest beneficiary of their work in commercial innovation: the US share of global patent citations to an emigrant’s science falls from ≈70% pre-emigration to ≈50% post-emigration, but remains much larger than any single destination country’s share and as large as the rest of the world combined.

Key Points

• Aggregate mobility
  • ~10% of US-trained, publishing STEM PhDs publish abroad within 5 years of graduation; ~20% by 10 years; ~25% by 15 years.
  • Foreign nationals make up ~40–45% of US STEM PhD cohorts since the 1990s; emigration rates track the foreign-share closely.
  • Emigration rates have been broadly stable across cohorts since the 1980s.
• Field and technology heterogeneity
  • Higher 15-year leave rates (~30% or more) in mathematics, computer science, and many engineering fields; AI and quantum-science-related PhDs show particularly high leave rates among OSTP "critical and emerging" areas.
  • Lower leave rates (<20% in 15 years) in biology, biomedical engineering, and health sciences.
• Destinations and trends
  • Growing share of emigrants move to Asia (China has been the top destination since mid-2000s); Europe receives about one-third of emigrants; Middle East share has risen.
• Impact on commercial uptake
  • US assignees account for ≈70% of patent citations to emigrants’ pre-emigration publications; this share drops to ≈50% for post-emigration publications.
  • Destination-country assignees’ share of citations to a given scientist’s work rises from ≈2% (pre-immigration) to ≈10% (post-immigration).
  • Stayers are positively selected on early career outcomes: conditional on publishing, stayers have slightly more publications and higher citation and patent linkage rates (e.g., 10–17% more patent citations/patent-paper pairs per publication).
• Interpretation caveats
  • Observational design: selection (better scientists staying) and causal effects (US environment raising impact) are both possible; authors document positive selection but cannot fully disentangle causality.
  • Patent citations and patent-paper pairs are proxies (useful but imperfect) for commercial value and knowledge flows.

Data & Methods

• Core sample
  • STEM PhD dissertations from ProQuest Dissertations & Theses Global (PQDT), cohorts 1980–2019, validated against NSF Survey of Earned Doctorates totals.
• Classification
  • Titles/abstracts classified into SED fields and 18 OSTP-defined "critical and emerging" technology areas using a custom LLM-based classifier.
• Longitudinal linkage
  • Linked PhD records to subsequent publications using a custom two-stage supervised ML model trained and validated on ORCID self-reported careers; model reported high precision and recall across ethnicities (details and performance in Supplementary Materials).
  • Institutional affiliations inferred from publications to determine country of employment and timing of emigration.
• Measuring commercial uptake
  • Merged publication-affiliation data with patent citation data (Marx & Fuegi style front-page patent–article citation measures and patent–paper pair linking) to compute national shares of patent citations to each graduate’s science pre- and post-emigration.
• Robustness & limitations
  • Conditioning analyses on "scientifically active" (publishing) PhDs.
  • Acknowledge modest PQDT attrition in recent years, imperfect causal identification, and the limitations of patent citations as measures of economic impact.

Implications for AI Economics

• Talent supply and localization of AI capabilities
  • AI-related PhDs have relatively high leave rates; losing a sizable fraction of trained AI researchers abroad reduces the onsite human capital available for US firms and labs that translate AI research into products.
  • Because US assignee citation share to emigrant AI science drops materially after emigration (~70% → ~50%), firm-level absorptive capacity in the US for that scientist’s work weakens when the scientist relocates.
• Diffusion vs. concentration trade-offs
  • Even after emigration the US remains the largest single consumer of graduates’ science, so training foreign AI PhDs delivers global spillovers that still disproportionately benefit the US. But the increase in destination-country patent citations (≈2%→10%) signals growing localized industrial absorption in destinations—accelerating AI capability building abroad.
• Policy levers and economic strategy
  • Immigration and retention policies matter: easing pathways to remain (jobs/visas/permanent residency) can increase the local commercial translation of AI science.
  • Complementary policies—industry–university linkages, sponsored postdocs, start-up/scaleup incentives, targeted R&D procurement, and mentorship/access to local labs—may help convert training into retained domestic productive capacity.
  • Given selection effects (stayers are on average more commercially linked), policies that both attract high-potential AI PhDs and provide attractive onshore commercialization ecosystems could have outsized returns.
• Firm strategy and multinational R&D
  • Multinational firms may respond to talent outflows by relocating R&D or establishing labs where talent clusters, amplifying cross-border production of AI systems. Monitoring and incentivizing onshore R&D investment can shape whether retained talent yields domestic AI industrial capacity.
• Measurement and research needs in AI economics
  • For AI specifically, mapping talent flows to measurable downstream economic outcomes (deployment of deployed AI systems, firm formation, market shares, and country-level AI capabilities) should be a priority; patent citation proxies are informative but incomplete for AI (where software, datasets, and services matter).
  • Future work should aim to better separate selection from causal effects of location on AI research productivity and commercialization, and to study how different retention policies alter the geography of AI innovation.

Bottom line: US training of AI and other STEM PhDs remains a major source of scientific and commercial inputs for US innovation even when some graduates emigrate, but higher emigration in AI and related fields reduces the degree to which onshore firms directly internalize and commercialize that scientific knowledge—making retention and domestic commercialization capacity important levers for AI economic competitiveness.