Aging OECD economies need minimum AI investment to sustain growth

Advanced economies facing population decline may need country-specific artificial intelligence (AI) investment benchmarks to prevent per capita growth from slipping into negative territory, according to a new economic analysis. Aging, as the study states, is not only a labor-market problem but also a research-capacity problem, because shrinking populations can weaken the base of workers and researchers needed to sustain long-run innovation.

Published in Economies, the study "Sustaining Growth Under Demographic Decline: A Minimum AI Investment Threshold for OECD Economies" utilizes a semi-endogenous growth model calibrated to 15 OECD economies with United Nations World Population Prospects 2024 and OECD research-and-development data to estimate that the midpoint minimum AI-investment threshold ranges from 0.236% to 0.275% of GDP when 10% of gross expenditure on R&D is assumed to be AI-designated.

Aging turns AI investment into a growth-stability question

As working-age populations shrink or grow more slowly, countries face weaker labor supply, rising dependency burdens and pressure on public finances. The research adds another dimension to that debate: a smaller demographic base can reduce the scale of research effort available for knowledge creation.

The model treats AI not simply as another technology sector, but as research capital that can help offset demographic pressure. In aging economies, AI-oriented research tools may raise the productivity of a smaller research workforce, allowing fewer workers and researchers to support continued knowledge creation. The issue is not whether AI improves productivity in general, but how much AI-designated investment is needed to keep per capita growth from turning negative under demographic decline.

The result is a minimum AI-investment threshold. In the benchmark model, the required AI share rises as population growth falls. Economies with faster population decline need a larger AI research effort to maintain non-negative per capita growth. That relationship is linear in the model, making demographic pressure directly translatable into an investment requirement.

This is a different policy question from the usual debate over AI competitiveness. Governments have increasingly promoted AI through infrastructure spending, subsidies, regulation and national AI strategies. But those efforts are often framed around winning technological races, attracting private investment or improving productivity. The research places AI policy inside a demographic-growth framework, where AI becomes one possible tool for stabilizing long-run growth in countries with shrinking populations.

The framework also separates direction from scale. Existing research has shown that aging countries often adopt more automation because labor becomes scarcer. The new contribution is a benchmark for the amount of AI-oriented investment needed to make that adjustment growth-sustaining. Automation may increase as countries age, but the key question is whether the increase is large enough to offset the demographic drag.

The model includes three main growth channels:

• Semi-endogenous knowledge production: It faces diminishing returns, meaning that sustaining growth becomes harder when the research base weakens.
• Human-capital deepening: It can partly compensate by making workers more productive, but it may not be enough on its own.
• AI research capital: It is introduced as an additional channel that can raise innovation capacity under demographic stress.

The analysis does not solve every question about household saving, capital markets, migration, education or government finance. Instead, it isolates the research-capital channel to show how demographic pressure changes the minimum AI-investment requirement. This makes the result useful as a policy reference point, but not as a fixed spending mandate.

Country rankings show why a single OECD-wide target may mislead

The 15-country application shows that the GDP-point range of the benchmark threshold is narrow, but the burden differs sharply once compared with each country's existing research base. Under the midpoint scenario, the required AI investment share ranges from 0.236% of GDP in Australia to 0.275% of GDP in Japan and Poland. South Korea and Italy are close behind at about 0.272% of GDP, while Germany, Spain, Finland and Portugal fall in the middle of the distribution.

The ranking is driven mainly by demographic pressure under a common normalization. Japan and Poland sit near the upper end because their projected population declines are among the steepest in the sample. Australia lies at the lower end because its population is projected to keep growing. The United States, the United Kingdom, Canada, Sweden, France and the Netherlands also face lower thresholds because their demographic outlooks are less severe than those of the fastest-declining economies.

The raw GDP-point figures, however, tell only part of the story. Research intensity varies widely across countries. South Korea and the United States have large research bases, so the midpoint threshold translates into about 5–7% of current gross expenditure on R&D. By contrast, Italy, Poland and Spain have thinner research bases, so a similar GDP-point threshold can require roughly 18–20% of current R&D. This makes the benchmark more demanding for countries with both demographic pressure and lower research intensity.

A research-rich economy can meet the benchmark partly by redirecting or improving the effectiveness of existing research spending. A country with a smaller R&D base may need to expand its research system itself before AI-designated investment can reach the necessary scale without displacing other priorities.

The sensitivity exercises reinforce the need for country-specific benchmarks. Changing the assumed AI-designated share of R&D from 5% to 15% mechanically changes threshold levels, but does not change the cross-country ordering. Low-population scenarios raise the threshold for every country, while high-population scenarios lower it. The demographic ranking remains broadly stable because the underlying logic is tied to population pressure.

The model also shows that migration can reduce the AI-investment burden. If net migration, labor-force participation or skill retention improves the effective labor base, the required AI threshold falls. This does not mean migration can replace AI policy. It means demographic and innovation policies are connected. Countries under severe aging pressure may be able to reduce required AI spending by expanding the research-relevant workforce through immigration, education and labor-market participation.

Capital deepening also plays a role. If countries compensate for labor decline through higher capital per worker, the AI threshold can fall. For example, a modest capital-deepening contribution would lower the estimated benchmark requirement for Japan. That makes the main threshold an upper-leaning reference when capital accumulation helps cushion demographic decline.

The role of crowding out is more troubling. If public AI spending merely displaces private R&D or non-AI research rather than adding to net AI research capital, the gross spending requirement rises. Under a 25% crowding-out assumption, Japan's benchmark requirement rises from 0.275% to 0.367% of GDP. Under 50% crowding out, it rises to 0.551%. This highlights the importance of policy design. The issue is not only how much governments spend, but whether spending expands effective AI research capacity.

AI policy must be tied to research capacity, migration and demographic pressure

The findings point toward a more differentiated AI policy agenda for aging economies. A single OECD-wide AI spending target would fail to account for differences in demographic decline, research intensity and existing innovation capacity. Countries with similar GDP-point thresholds may face very different real burdens depending on the size of their R&D systems.

For Japan, Germany and South Korea, the policy challenge is largely about maintaining and redirecting strong research systems under demographic strain. These economies already have substantial research bases, but aging still raises the minimum AI effort needed to sustain growth. Their priority may be improving transmission from AI spending into effective research capital, supporting AI infrastructure, and ensuring that AI tools are integrated into scientific and industrial R&D.

The challenge is more structural for Italy, Poland and Spain. Because existing GERD intensity is lower, AI-designated investment equal to the benchmark would represent a much larger share of total research expenditure. These countries may need broader research-base expansion, stronger university-industry links, larger public-private R&D programs and more targeted support for AI research capacity.

For countries with more favorable demographic outlooks, such as Australia, Canada and the United States, the benchmark suggests less pressure from population decline. But that does not remove the need for AI investment. It means AI policy can focus more on competitiveness, productivity and diffusion, while aging-related growth stabilization is less urgent than in faster-shrinking economies.

The analysis also suggests that AI investment should not be treated in isolation from education and workforce policy. Human capital remains a core channel in the model. AI research capital can raise innovation capacity, but workers and researchers still need the skills to use AI tools effectively. Education, training and scientific workforce development become complements to AI spending rather than separate policy areas.

The same logic applies to migration. In countries where demographic decline is severe, immigration policies that attract skilled workers and researchers can lower the AI investment needed to maintain growth. Retaining older workers, increasing labor-force participation and improving the productivity of existing researchers can also reduce pressure on AI spending.

Public finance design is another key issue. Subsidies, tax credits, procurement programs and shared research infrastructure should be judged by their effect on net AI research capital, not only by headline spending. Programs that relabel existing R&D as AI investment or shift private spending onto public balance sheets may do little to close the threshold. Policies that crowd in private research, build usable data and computing infrastructure, and support AI tools for scientific discovery are more likely to increase effective research capacity.

The benchmark also gives international organizations a potential role. Instead of promoting one AI investment ratio for all advanced economies, institutions such as the OECD could develop country-specific dashboards that combine demographic projections, R&D intensity and AI-designated investment measures. Such benchmarks would help governments compare needs more transparently and update targets as demographic forecasts and AI measurement improve.

Speaking of the limitations, the authors note that the model is a simplified growth benchmark, not a full forecast of national economic performance. It relies on assumptions about AI-designated research shares and the transmission from spending into research capital. It uses total population growth as a proxy for the demographic base, even though working-age population or researcher counts might produce different estimates. It also does not fully model transition paths, private-sector behavior or institutional differences in innovation systems.