## AI and Economic Divergence in Asia — Working Paper No. WP/2026/166

## Source details

**Canonical URL:** [AI and Economic Divergence in Asia — Working Paper No. WP/2026/166](https://www.imf.org/-/media/files/publications/wp/2026/english/wpiea2026166-source-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2026/english/wpiea2026166-source-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2026/english/wpiea2026166-source-pdf.pdf.json)

---

### Households (Section 4.1)
- Household composition and demographics:
  - 100 overlapping generations indexed by g ∈ {0,1,...,100}; agents live up to age 100.
  - Three skill classes s ∈ {L,M,H} with population shares φ_{s,t} evolving over time.
  - Agents begin working at age 21 and retire at the country’s official retirement age.
  - Survival probabilities π_{g,t} depend on country-specific demographic projections.
  - Households derive utility from own consumption c_{g,t,s}, leisure l_{g,t,s}, and descendants’ consumption through H_{j,t}^{j,s}.
- Preference specification (lifetime utility):
  - Lifetime utility:
    - U_{g,t,s} = Σ_{j=g}^{100} β^{j−g} π_{j,t} [ (1/(1−γ)) ( c_{j,t}^{ρ}_{j,s} + α_{s} l_{j,t,s}^{ρ} )^{(1−γ)/ρ} + θ· H_{j,t}^{j,s} ]
  - Parameters/objects listed: β, γ, ρ, α_{s}, θ, H_{j,t}^{j,s}.
  - Utility from descendants:
    - H_{j,t}^{j,s} = D_{j,t}^{j,s} · [ c_{deps_{j,t}}^{j,s} · (1 + α_{s}^{ρ} · (w_{t}^{j,s} (1− τ_{total_{j,t}^{j,s}})/p_{t}^{j})^{1−ρ} / (1 + α_{s}^{ρ} · (w_{t}^{j,s} (1− τ_{total_{j,t}^{j,s}})/p_{t}^{j})^{−ρ} ))^{(ρ−γ)/(1−ρ)} ]
    - D_{j,t}^{j,s}: counts dependents; c_{deps_{j,t}}^{j,s}: consumption per dependent.
- Leisure choice (optimality condition):
  - Optimal leisure:
    - l_{g,t,s} = c_{g,t,s} · α_{s}^{ρ} · ( w_{t,s} (1− τ_{L_{g,t,s}}) / p_{t} )^{−ρ}
  - p_{t} is the consumption price.
- Budget constraint for working-age households:
  - p_{t} · c_{g,t,s} + a_{g+1,t+1,s} = (1− τ_{L_{g,t,s}}) w_{t,s} h_{g,t,s} + (1 + r_{t} (1− τ_{K_{g,t,s}}) ) a_{g,t,s} + T_{g,t,s} + I_{g,t,s}
  - Definitions: a_{g,t,s} (assets), h_{g,t,s} = 1 − l_{g,t,s} (hours worked), τ_{L_{g,t,s}} (labor tax), τ_{K_{g,t,s}} (capital tax), p_{t} (consumer price), T_{g,t,s} (transfers), I_{g,t,s} (inheritances).

### Cross-Country Divergence (Section 6.1)
- Adoption timing and scenarios:
  - Scenarios: capital share rises at 2×, 5×, and 10× the historical US rate (mild, moderate, accelerated).
  - Advanced economies (AEs) adopt immediately in all scenarios.
  - EMDEs adopt with delays increasing with faster AI progress; longer lags under 5× and 10× than under 2×.
  - Mechanism: early adopters demand more capital → global interest rates rise → late adopters face higher financing costs that impede capital accumulation and switching to capital-intensive technology.
- Determinants of early adoption:
  - Initial capital intensity example: capital accounts for 40 to 50 percent of output in Japan, Korea, and Singapore under the baseline.
  - Skill composition example: high-skilled workers comprise 10 to 15 percent in Singapore and Korea versus 2 to 5 percent in Bangladesh and Cambodia.
  - TFP: higher TFP amplifies capital and labor productivity and, with high capital shares, yields higher capital-to-labor ratios (example: Singapore and Japan).
  - Demographics: working-age population declines of 25 to 30 percent by 2050 in Japan and Korea favor early adoption; young, growing populations in Bangladesh, India, Philippines delay adoption.
- Clarifications and sensitivity:
  - Dates indicate economy-wide AI adoption at scale (profitability threshold), not marginal or experimental AI use.
  - Timing sensitive to calibration; read results as ordinal/illustrative.
  - Cross-country ordering can change across scenarios (example: Philippines vs Vietnam ordering reverses across scenarios).
  - Faster AI progress can widen and prolong cross-country gaps absent policies accelerating diffusion and strengthening complementarities.
  - Some EMDEs delayed into the 2040s and low-income countries into the 2050s–60s under faster development scenarios.
- Growth impacts — country examples and aggregates:
  - Korea (early adopter): growth rises above baseline immediately; gains scale with AI progress; source: capital deepening and higher output per unit of capital.
  - Vietnam and Bangladesh (late adopters): growth below baseline during pre-adoption due to higher global interest rates raising domestic cost of capital → slower capital accumulation and reduced investment.
  - Channel strength depends on capital account openness and pass-through from global to domestic interest rates.
  - Aggregate EMDE impacts (pre-adoption vs no-AI baseline):
    - Pre-adoption annual growth lower by about 0.5 to 0.6 percentage points in the mild and moderate AI scenarios than the no-AI scenario.
    - Investment-to-GDP ratio lower by 5 to 6 percentage points in the mild and moderate AI scenarios during pre-adoption.
  - Long-term average annual growth boosts from AI adoption:
    - AEs: 0.5 to 1 percentage point.
    - EMDEs in the region: 0.2 to 0.8 percentage points.
- Modeling caveats:
  - Simulations use country-specific structural fundamentals and a balanced-budget assumption for government revenues/expenditures.
  - Do not incorporate policy responses to accelerate diffusion or mitigate adverse effects; results illustrative and conditional.
  - Accelerated (10×) scenario is a stylized upper bound; in some specifications it generates materially lower investment and growth reflecting frictionless capital-retooling assumption and 10× pace.
  - Adjustment costs, financing frictions, and policy responses would dampen extreme simulated swings in investment-to-GDP.
- Structural reforms to accelerate adoption (2025–2035 reform scenarios):
  - Reform scenarios: moderate and ambitious applied over ten years.
  - Parameter changes:
    - High-skilled labor share: +2.5 percentage points (moderate); +5 percentage points (ambitious).
    - Labor-specific productivity (ΦL,s): +25 percent (moderate) and +50 percent (ambitious) for high-skilled; +12.5 percent (moderate) and +25 percent (ambitious) for low- and medium-skilled.
    - Capital-specific productivity (ΦK): +25 percent (moderate); +50 percent (ambitious).
  - Policy domains: higher education and vocational training (high-skilled share); labor market reforms (labor productivity); financial deepening, reduced misallocation, improved infrastructure (capital productivity).
  - Scale and comparability: ambitious magnitudes likened to Korea’s 1970s transformation or China’s 2000s; chosen to illustrate potential impacts.
  - Impacts under accelerated (10×) scenario:
    - Countries that would adopt in the 2050s or 2060s could advance adoption by one to two decades with ambitious reforms.
    - Reforms deliver a double dividend: direct output/productivity boosts and earlier readiness for AI adoption, compounding growth gains.
  - Policy implications:
    - Comprehensive, coordinated reforms across education, labor markets, capital allocation, and infrastructure are critical.
    - Piecemeal measures yield smaller benefits due to cross-dimension interactions.
    - Ten-year reform horizon ambitious but achievable with sustained political commitment; more modest reforms still produce positive, smaller effects.

### Within-Country Distributional Impacts (Section 6.2)
- Inequality across skill levels (6.2.1):
  - Mechanisms and empirical context:
    - Digital technologies substitute routine tasks and complement non-routine cognitive tasks.
    - Occupational heterogeneity in AI exposure; large share of workers have some tasks exposed, often in information-intensive and higher-paid occupations.
    - Demand for new AI-related skills associated with higher wage offers but can reinforce job polarization—benefiting high-skilled and indirectly low-skilled while limiting middle-skilled gains.
    - AI can raise productivity more for less-experienced or lower-performing workers, potentially narrowing within-occupation dispersion.
  - Model representation:
    - Task displacement mapped to declining output elasticities for each skill group; low-skilled shares fall fastest.
    - Simulations present wage ratios (AI scenario vs no-AI baseline) for Japan and Singapore under 2×, 5×, 10× scenarios; a value of 2 means wages twice the no-AI baseline.
  - Key findings:
    - High-skilled workers see the largest wage gains across all scenarios (their output elasticity declines least and they complement capital strongly).
    - Low- and medium-skilled wage effects vary with AI speed:
      - Under 2×: low-skilled wages initially dip below baseline during transition, then recover as capital accumulates.
      - Under 5× and 10×: faster capital deepening offsets declining labor share sooner; in 10× low-skilled wages rise above baseline initially but gains are much smaller than for other groups.
    - Under 10× in Japan: low-skilled wage ratios eventually overtake medium-skilled ratios; similar but milder pattern in Singapore—driven by labor supply: low-skilled stronger leisure preference → withdraw labor as consumption rises → scarcity pushes wages up.
  - Sensitivities and clarifications:
    - Adverse low-skilled effects arise from assumed capital-skill complementarity (low-skilled output elasticity falls fastest); workers do not switch skill categories in the model.
    - If low- and middle-skilled shares fell more symmetrically, or AI substituted more for routine middle/high-skilled tasks, transitional low-skilled losses would be smaller or could reverse.
    - Education-based skill groups are stylized; real exposure varies by occupation and task.
- Inequality across age groups (6.2.2):
  - Cohorts analyzed: age 25 (young), age 45 (mid-career), age 65 (older/retirees).
  - Aggregate pattern:
    - Older cohorts see earlier and larger income gains than younger cohorts, widening cross-sectional income gaps between generations.
    - Figures referenced: Figure 8 (total income ratios by age under 2×, 5×, 10×) and Figure 9 (decomposition of additional income by source circa 2050).
  - Mechanism:
    - Young and mid-career gain via higher wages building with capital deepening.
    - Older cohorts benefit largely through capital income: accumulated savings earn higher returns as AI-driven investment demand pushes up interest rates.
    - Faster AI progress produces larger interest rate increases and wider intergenerational gaps.
  - Scenario-specific nuances:
    - 10× in Singapore: retiree (age 65) income ratio rises sharply then declines after peaking around 2070 due to initial interest rate spike on retirement savings followed by dissaving and declining asset stocks as interest rates stabilize.
    - Mid-career (age 45) in Japan under 10×: income ratios dip below baseline driven by low-skilled workers—strong leisure preferences reduce labor supply; mid-career low-skilled are net borrowers and higher interest rates amplify borrowing costs, lowering measured total income even if lifetime welfare higher.
  - Interpretation: results highlight potential widening of wage and income dispersion across generations under rapid AI adoption absent redistributive policy responses.
- Redistributive policies and trade-offs (6.2.3):
  - Policy experiments: two transfer schemes for early adopters under mild (2×) scenario, each costing 5 percent of GDP and financed through the endogenous tax system:
    - Targeted transfer to low- and medium-skilled workers.
    - Universal transfer to all households.
  - Main results (summarized from Figure 10):
    - Both schemes reduce income inequality relative to no-transfer AI scenario; targeted transfers produce somewhat larger Gini reductions.
    - Both schemes entail modest output costs (annual average percent deviation in GDP).
    - Targeted transfers generate larger output losses than universal transfers because:
      - Tax burdens concentrated on high-skilled workers (receiving no transfer) create stronger work disincentives.
      - Low- and medium-skilled recipients reduce labor supply as non-labor income lowers marginal value of working.
    - Output costs larger in Japan (higher baseline tax rates amplify marginal distortions) and smaller in Singapore (more favorable fiscal starting point).
  - Limitations and design caveats:
    - Transfers modeled as unconditional lump sums; real designs (means-testing, phase-outs, conditionality) alter marginal tax rates and work incentives.
    - Skill composition treated as exogenous—transfers could reduce upskilling incentives; higher taxes on high-skilled could discourage human capital investment—these effects not modeled.
    - Active labor market policies (retraining, job matching, transition support) not modeled but could mitigate distributional pressures.
    - Financing mix matters: simulation uses household taxes (consumption, wage, capital); alternative instruments (corporate or wealth taxes) would have different distortionary effects.
    - Results focus on long-run aggregate GDP under a fixed fiscal envelope and abstract from institutional design, political economy, and social objectives.
  - Practical considerations: program design features (conditionality, time limits, active labor-market policies) can contain work disincentives and make targeted programs less distortionary.
- Model scope, assumptions, and interpretive caveats:
  - Small-open-economy assumption: perfect capital mobility, immobile labor; domestic interest rate follows externally-disciplined global path—may understate feedbacks for large economies.
  - AI modeled as frontier production technology used domestically, not as an innovation sector producing exports or accruing IP rents.
  - Aggregation: AI modeled broadly as capital-biased technological change; single homogeneous capital stock—does not distinguish robotics, LLMs, GPUs, data centers, or intangible investments.
  - Transition dynamics: adopting economies face sustained incentives to deepen capital and build complementary ecosystems while frontier moves; late adopters may face large transition investments.
  - Skill heterogeneity: represented by three education-based groups; model does not endogenize schooling or re-skilling decisions.
  - Omitted channels: trade, exchange rates, multi-sector structure, firm dynamics, financial frictions, labor market frictions, and policy credibility.
  - Sectoral staggering of adoption could smooth adjustment or concentrate gains/dislocations.
  - Interpretation: results illustrate mechanisms and directions of macroeconomic adjustment and distributional trade-offs under rapid AI progress, not precise country-level forecasts. Central message: preparedness and complementary fundamentals—capital deepening capacity, skills, productivity—and demographic structure shape timing and magnitude of AI dividends; redistributive and active labor-market policies matter for mitigating within-country inequality.

*Source: wpiea2026166-source-pdf — selected sections 4.1, 6.1, 6.2, References and Appendices.*

### 4.1    Households

### 4.1    Households

### Household structure and demographics
- The household sector consists of 100 overlapping generations indexed by g ∈ {0,1,...,100}, where agents live up to age 100.
- Each generation contains three skill classes s ∈ {L,M,H} representing low-, medium-, and high-skilled workers, with population shares φ_{s,t} evolving over time.
- Agents begin working at age 21 and retire at the country’s official retirement age.
- Survival probabilities π_{g,t} depend on country-specific demographic projections.
- Households derive utility from their own consumption c_{g,t,s} and leisure l_{g,t,s}, and also derive utility from descendants’ consumption through H_{j,t}^{j,s}.

### Preference specification (lifetime utility)
- Lifetime utility:
  - U_{g,t,s} = Σ_{j=g}^{100} β^{j−g} π_{j,t} [ (1/(1−γ)) ( c_{j,t}^{ρ}_{j,s} + α_{s} l_{j,t,s}^{ρ} )^{(1−γ)/ρ} + θ· H_{j,t}^{j,s} ]
  - Parameters and objects:
    - β: discount factor
    - γ: risk aversion
    - ρ: intratemporal substitution parameter
    - α_{s}: leisure preference for skill class s
    - θ: weight on descendants’ utility
    - H_{j,t}^{j,s}: utility from descendants’ consumption
- Utility from descendants:
  - H_{j,t}^{j,s} = D_{j,t}^{j,s} · [ c_{deps_{j,t}}^{j,s} · (1 + α_{s}^{ρ} · (w_{t}^{j,s} (1− τ_{total_{j,t}^{j,s}})/p_{t}^{j})^{1−ρ} / (1 + α_{s}^{ρ} · (w_{t}^{j,s} (1− τ_{total_{j,t}^{j,s}})/p_{t}^{j})^{−ρ} ))^{(ρ−γ)/(1−ρ)} ]
  - D_{j,t}^{j,s}: counts dependents (children and grandchildren)
  - c_{deps_{j,t}}^{j,s}: consumption allocated to each dependent

### Leisure choice (optimality condition)
- Households choose leisure optimally given wages and taxes.
- Optimal leisure choice:
  - l_{g,t,s} = c_{g,t,s} · α_{s}^{ρ} · ( w_{t,s} (1− τ_{L_{g,t,s}}) / p_{t} )^{−ρ}
  - p_{t} is the consumption price.

### Budget constraint for working-age households
- Budget constraint:
  - p_{t} · c_{g,t,s} + a_{g+1,t+1,s} = (1− τ_{L_{g,t,s}}) w_{t,s} h_{g,t,s} + (1 + r_{t} (1− τ_{K_{g,t,s}}) ) a_{g,t,s} + T_{g,t,s} + I_{g,t,s}
- Definitions:
  - a_{g,t,s}: assets
  - h_{g,t,s} = 1 − l_{g,t,s}: hours worked
  - τ_{L_{g,t,s}}: combined tax rate on labor income
  - τ_{K_{g,t,s}}: tax rate on capital income
  - p_{t}: consumer price inclusive of consumption taxes
  - T_{g,t,s}: transfers and social benefits
  - I_{g,t,s}: inheritances

*Source: wpiea2026166-source-pdf - 4.1    Households (PDF chapter).*

### 6.1    Cross-Country Divergence

### 6.1 Cross-Country Divergence

### Adoption Timing
- Scenarios modeled: capital share in the production function rises at 2×, 5×, and 10× the historical US rate (referred to as mild, moderate, and accelerated scenarios).
- Advanced economies (AEs) adopt immediately in all scenarios.
- Emerging markets and developing economies (EMDEs) adopt with delays that increase with faster AI progress (longer lags under 5× and 10× scenarios than under 2×).
- Mechanism for longer EMDE delays under faster AI progress:
  - Early adopters demand more capital → global interest rates rise higher and faster → late adopters face higher financing costs that impede capital accumulation and switching to more capital-intensive technology.
- Three primary determinants of early adoption:
  - Initial capital intensity: e.g., capital already accounts for 40 to 50 percent of output in Japan, Korea, and Singapore under the baseline.
  - Skill composition: high-skilled workers comprise 10 to 15 percent of the workforce in Singapore and Korea (model definition) versus 2 to 5 percent in Bangladesh and Cambodia.
  - Total factor productivity (TFP): higher TFP amplifies capital and labor productivity and, combined with high capital shares, yields higher capital-to-labor ratios (e.g., Singapore and Japan).
- Demographics:
  - Population aging and slow labor force growth raise capital-to-labor ratios, promoting early adoption (Japan and Korea face working-age population declines of 25 to 30 percent by 2050).
  - Young, growing populations (Bangladesh, India, Philippines) keep labor relatively abundant and capital scarce, delaying profitability of AI.
- Clarifications on adoption dates:
  - Dates indicate economy-wide AI adoption at scale (profitability threshold), not marginal or experimental AI use.
  - Timing is sensitive to calibration; results are best read as ordinal/illustrative rather than precise calendar forecasts.
- Cross-country ordering can change across scenarios (example: the Philippines adopt earlier than Vietnam under the moderate scenario but later under the accelerated scenario).
- Faster AI progress can widen and prolong cross-country gaps absent policies that accelerate diffusion and strengthen complementary factors.
- Delays described: some emerging markets delayed into the 2040s and low-income countries into the 2050s–60s under faster development scenarios.

### Growth Impacts
- Country examples illustrate divergent dynamics: Korea (early adopter), Vietnam (middle adopter), Bangladesh (late adopter).
- Korea (early adopter):
  - Growth rises above baseline immediately upon adoption; magnitude scales with AI progress speed.
  - Source of gain: capital deepening and higher output per unit of capital as capital plays larger role in production.
- Vietnam and Bangladesh (late adopters):
  - Do not adopt immediately; experience growth below baseline during the pre-adoption period.
  - Cause: early adopters’ increased investment raises global interest rates → late adopters (small open economies) face higher cost of capital → slower capital accumulation, reduced investment, and lower growth relative to baseline.
- Integration caveat:
  - Channel strength depends on capital account openness and pass-through from global to domestic interest rates; weaker where pass-through is limited.
- Aggregate EMDE impacts (comparison to no-AI baseline and across scenarios):
  - Pre-adoption annual growth is lower by about 0.5 to 0.6 percentage points in the mild and moderate AI scenarios than the no-AI scenario.
  - Investment-to-GDP ratio is lower by 5 to 6 percentage points in the mild and moderate AI scenarios during pre-adoption.
  - Upon eventual adoption, late adopters’ growth recovers and exceeds baseline reflecting rapid catch-up capital accumulation.
- Long-term average annual growth boosts from AI adoption (depending on pace of AI advance):
  - AEs: 0.5 to 1 percentage point.
  - EMDEs in the region: 0.2 to 0.8 percentage points.
- Modeling caveats and interpretation:
  - Simulations use country-specific structural fundamentals and a balanced-budget assumption for government revenues/expenditures.
  - They do not incorporate policy responses to accelerate diffusion or mitigate adverse effects; results are illustrative and conditional on assumptions.
  - The accelerated (10×) scenario serves as a stylized upper bound; in some specifications it generates materially lower investment and growth (reflecting frictionless capital-retooling assumption and 10× pace).
  - In practice, adjustment costs, financing frictions, and policy responses would dampen extreme simulated swings in the investment-to-GDP ratio.

### Structural Reforms to Accelerate Adoption
- Reform scenarios simulated over a ten-year period (2025–2035): moderate and ambitious.
- Parameter changes in reform scenarios:
  - High-skilled labor share: increase by 2.5 percentage points (moderate) and 5 percentage points (ambitious).
  - Labor-specific productivity (ΦL,s): increase by 25 percent (moderate) and 50 percent (ambitious) for high-skilled workers; increase by 12.5 percent (moderate) and 25 percent (ambitious) for low- and medium-skilled workers.
  - Capital-specific productivity (ΦK): increase by 25 percent (moderate) and 50 percent (ambitious).
- Policy domains corresponding to parameter changes:
  - High-skilled share: expanded higher education and vocational training.
  - Labor productivity: labor market reforms.
  - Capital productivity: financial deepening, reduced misallocation, improved infrastructure.
- Scale and comparability:
  - Ambitious reform magnitudes are large—comparable to Korea’s transformation in the 1970s or China’s in the 2000s—and chosen to illustrate potential impacts of sustained comprehensive reforms.
  - Moderate reforms yield smaller but still meaningful advances.
- Impacts of reforms under the accelerated AI (10×) scenario:
  - Countries that would adopt in the 2050s or 2060s could advance adoption by one to two decades with ambitious reforms.
  - Reforms provide a double dividend: (1) direct boosts to output and productivity; (2) earlier readiness for AI adoption, compounding growth gains over time.
- Policy implications:
  - Comprehensive, coordinated reforms across education, labor markets, capital allocation, and infrastructure are critical to accelerate adoption and capture AI gains.
  - Piecemeal measures addressing only one dimension produce smaller benefits due to interactions across reform dimensions.
  - The ten-year reform horizon is ambitious but achievable with sustained political commitment and sequenced implementation; more modest reforms still produce positive, smaller effects.

*Source: Authors’ simulations in the chapter "6.1 Cross-Country Divergence."*

### 6.2    Within-Country Distributional Impacts

### 6.2    Within-Country Distributional Impacts

### 6.2.1    Inequality across Skill Levels
- Mechanisms and evidence:
  - Digital technologies substitute for routine tasks and complement non-routine cognitive tasks (Autor et al., 2003).
  - Task displacement by automation has explained a sizable share of changes in the U.S. wage structure (Acemoglu and Restrepo, 2021).
  - Occupation-level measures show wide heterogeneity in AI exposure (Felten et al., 2021); LLM estimates suggest a large share of workers have at least some tasks exposed, often concentrated in information-intensive and higher-paid occupations (Eloundou et al., 2023).
  - Demand for new (including AI-related) skills is associated with higher wage offers but can reinforce job polarization—benefiting high-skilled and indirectly low-skilled workers while offering limited gains for middle-skilled workers (Jaumotte et al., 2026).
  - Augmenting AI can raise productivity more for less-experienced or lower-performing workers, potentially narrowing within-occupation dispersion (Brynjolfsson et al., 2021).
- Model representation:
  - Task displacement across occupations maps into declining output elasticities for each skill group, with low-skilled workers’ shares falling fastest.
  - Simulations present wage ratios between AI scenarios and the no-AI baseline for Japan and Singapore under three AI progress speeds: 2×, 5×, and 10× (capital share increases at 2, 5, and 10 times the historical U.S. rate).
  - A value of 2 in Figure 7 means wages are twice as high as in the no-AI baseline.
- Key simulation findings:
  - Across all three scenarios, high-skilled workers see the largest wage gains (their output elasticity declines least and they benefit from strong complementarity with capital).
  - Low- and medium-skilled wage effects vary with AI progress pace:
    - Under 2×, low-skilled wages initially dip below the no-AI baseline during transition as falling output elasticity outweighs capital complementarity, then recover as capital accumulates.
    - Under 5× and 10×, faster capital deepening offsets declining labor share sooner; in the 10× scenario low-skilled wages rise above baseline initially, but gains are much smaller than for other groups.
  - Under 10×, low-skilled wage ratios eventually overtake medium-skilled ratios in Japan; a similar but milder pattern occurs in Singapore. This reversal operates through labor supply: low-skilled workers have a stronger leisure preference and withdraw labor as consumption rises, making low-skilled labor scarcer and pushing wages up.
- Important modeling clarifications and sensitivities:
  - The adverse low-skilled effect arises from assumed capital-skill complementarity—low-skilled labor’s output elasticity falls fastest—not from occupational downgrading (workers do not switch skill categories in the model).
  - If low- and middle-skilled shares fell more symmetrically, or if AI substituted more for routine middle/high-skilled tasks, transitional low-skilled wage losses would be smaller or could reverse.
  - Education-based skill groups are a stylized proxy; actual exposure and complementarity vary at occupation and task level (some low-skilled manual or in-person services may be less exposed).
  - Simulations are an illustrative mapping from a skill-biased, capital-deepening shock to relative wage dynamics, not point predictions for specific occupations or countries.

### 6.2.2    Inequality across Age Groups
- Cohorts analyzed:
  - Young workers (age 25), mid-career workers (age 45), older workers or retirees (age 65).
- Aggregate pattern across scenarios and countries:
  - Older cohorts see earlier and larger income gains than younger cohorts, widening cross-sectional income gaps between generations.
  - Figure 8 shows total income ratios (labor, capital, pension) under AI scenarios relative to the no-AI baseline for ages 25, 45, and 65 at 2×, 5×, and 10×.
  - Figure 9 decomposes additional income from AI by source for each cohort circa 2050.
- Mechanism:
  - Young and mid-career workers gain primarily via higher wages that build gradually with capital deepening.
  - Older cohorts benefit largely through capital income: accumulated savings earn higher returns as AI-driven investment demand pushes up interest rates.
  - Faster AI progress produces larger interest rate increases and wider intergenerational gaps.
- Scenario-specific nuances:
  - In the 10× scenario in Singapore:
    - The income ratio for retirees (age 65) rises sharply then declines after peaking around 2070.
    - Initial surge: interest rate spike on pre-existing retirement savings.
    - Subsequent decline: dissaving—later retiree cohorts consumed more during the AI boom and arrived at retirement with less wealth; as interest rates stabilize and asset stocks decline, income ratios fall. Effect strongest in Singapore (earlier retirement), mildest in Japan (retirees nearer peak accumulation).
  - Mid-career workers (age 45) in Japan under 10× see income ratios dip below baseline, driven primarily by low-skilled workers:
    - Low-skilled workers have strong leisure preferences and reduce labor supply as lifetime income rises.
    - Mid-career low-skilled workers are net borrowers; higher interest rates amplify borrowing costs, reducing measured total income even as lifetime welfare may be higher (they optimize over full lifetime).
- Interpretation:
  - The increased cross-sectional income gap between younger and older cohorts represents a real increase in income inequality.
  - Results are a transparent benchmark highlighting potential widening of wage and income dispersion under rapid AI adoption absent redistributive policy responses.

### 6.2.3    Redistributive Policies and Associated Trade-Offs
- Policy experiments:
  - Two alternative transfer schemes simulated for early adopters under the mild (2×) AI scenario, each costing 5 percent of GDP:
    - Targeted transfer directed to low- and medium-skilled workers.
    - Universal transfer paid to all households.
  - Both financed through the endogenous tax system.
- Main results (Figure 10 summary):
  - Both schemes reduce income inequality relative to the no-transfer AI scenario; targeted transfers achieve somewhat larger reductions in inequality (Gini coefficient).
  - Both schemes entail modest output costs (annual average percent deviation in GDP):
    - Targeted transfers tend to generate larger output losses than universal transfers because:
      - They create tax burdens concentrated on high-skilled workers who receive no transfer, creating stronger disincentives to work.
      - Low- and medium-skilled recipients reduce labor supply as non-labor income lowers the marginal value of working.
    - Universal transfers spread tax burden and benefit across households, dampening net distortion for any single group.
  - Output costs are larger in Japan (higher baseline tax rates amplify marginal distortions) and smaller in Singapore (more favorable fiscal starting point).
- Limitations and design caveats:
  - Transfers are modeled as unconditional lump sums; real program design (means-testing, phase-out rates, conditionality) alters effective marginal tax rates and can amplify or dampen work disincentives.
  - Skill composition is treated as exogenous—transfers could reduce incentives to upskill, and higher taxes on high-skilled workers could discourage human capital investment; these effects are not modeled.
  - Active labor market policies (retraining, job matching, transition support) are not modeled but could address distributional pressures.
  - Financing mix matters: the simulation uses household taxes (consumption, wage, capital taxes); alternative instruments (corporate or wealth taxes) would generate different distortionary effects.
  - Results focus on long-run aggregate GDP under a fixed fiscal envelope and abstract from institutional design, political economy, and social objectives; they illustrate potential impacts rather than prescribing specific policies.
- Practical considerations:
  - Policymakers may prioritize who is protected and the social/political benefits of shielding those most exposed to AI—dimensions not fully captured by the model.
  - Program design features (conditionality, time limits, robust active labor-market policies) can help contain work disincentives and may make targeted programs less distortionary in practice.

### Model scope, assumptions, and interpretive caveats (implications for distributional results)
- Small-open-economy assumption and global factor prices:
  - Each economy is simulated as a small open economy with perfect capital mobility and immobile labor; domestic interest rate follows an externally-disciplined global path.
  - This simplifies cross-country comparison but may understate feedbacks for large economies that can influence global savings, investment, and technology.
- AI use versus AI production and rents:
  - The framework models AI as a frontier production technology adopted and used, not as an innovation sector producing AI exports or accruing IP rents; thus it does not capture rents to AI-producing economies.
- Aggregation over AI types and capital:
  - AI is modeled broadly as capital-biased technological change that increases capital intensity and reduces labor shares; the model does not distinguish robotics, LLMs, or other modalities.
  - The model features a single homogeneous capital stock; in reality, AI-related investment includes data centers, GPUs, cloud systems, software, and intangible investments—legacy capital may be only partially compatible with frontier AI.
- Transition dynamics and late adoption:
  - Even with falling unit costs, adopting economies face sustained incentives to deepen capital and build complementary ecosystems while the frontier moves; late adopters may still face large transition investments.
- Skill heterogeneity and task content:
  - Within-country heterogeneity is represented by three education-based skill groups; the model abstracts from occupation- and task-level variation in AI exposure and complementarity.
  - The model does not endogenize schooling or re-skilling decisions—these would likely mitigate inequality and accelerate adoption over long horizons.
- Omitted channels and sectoral heterogeneity:
  - The single-good framework abstracts from trade, exchange rates, prices, multi-sector structure, firm dynamics, financial frictions, labor market frictions, and policy credibility—each could affect divergence and distributional outcomes.
  - Sectoral staggering of adoption (capital-/information-intensive sectors adopting earlier than labor-intensive or in-person services) could smooth economy-wide adjustment or concentrate gains and dislocations.
- Interpretation for policy analysis:
  - Results are illustrations of mechanisms and directions of macroeconomic adjustment and distributional trade-offs under rapid AI progress, not precise country-level forecasts.
  - Central message: preparedness and complementary fundamentals—capital deepening capacity, skills, productivity—and demographic structure shape timing and magnitude of AI dividends; redistributive and active labor-market policies matter for mitigating within-country inequality.

*Source: IMF Working Paper — Chapter/section 6.2, "Within-Country Distributional Impacts" (authors’ simulations and discussion).*

### References

### References and Appendices (wpiea2026166-source-pdf - References)

### Key bibliographic references
- Selected studies and reports cited include works by Acemoglu & Restrepo (2019; 2021), Alonso et al. (2022), Autor et al. (2003), Benzell et al. (2021), Brynjolfsson et al. (2021), Budina et al. (2023), Cazzaniga et al. (2024), Cerutti et al. (2025), Chang et al. (2025), Comin & Hobijn (2010), Drozd et al. (2022), Eloundou et al. (2023), Felten et al. (2021), Gmyrek et al. (2023), Iida et al. (2026), IMF (2024, November), Jaumotte et al. (2026, January), Korinek & Stiglitz (2019), Moll et al. (2022), OECD/Boston Consulting Group/INSEAD (2025), Rockall et al. (2025), UNCTAD (2025), United Nations (2024), UNDP (2025, December).
- Data products and indices referenced include the Government AI Readiness Index 2025 (8th edition), the Frontier Technologies Readiness Index, World population prospects 2024, and multiple IMF Working Papers and Staff Discussion Notes.

### Appendix A — Benchmark variables: construction and coverage
- Baseline sample of countries: Australia (AUS), Bangladesh (BGD), China (CHN), Indonesia (IDN), India (IND), Japan (JPN), Cambodia (KHM), Korea (KOR), Sri Lanka (LKA), Malaysia (MYS), Nepal (NPL), New Zealand (NZL), the Philippines (PHL), Singapore (SGP), Thailand (THA), and Vietnam (VNM).
- General conventions:
  - Figures originally reported as percentages of GDP are converted to unit-free ratios by dividing by 100, and results are presented to two decimal places unless otherwise noted.
  - Country names are matched by ISO-3 code on first mention.
  - Table?? consolidates benchmark variables reporting target values from external sources alongside model-implied values; all series are expressed as unit-free ratios.
  - Debt-service ratios smaller than 0.01 are shown to four decimals.
- A.1 Capital-to-output ratio
  - Inputs: RNNA (capital stock at constant 2017 national prices) and RGDPNA (real GDP at constant 2017 national prices) from Penn World Table (PWT) v10.01, combined with World Bank Wealth Accounts estimates of produced capital and non-renewable natural capital (both valued in chained 2019 US dollars).
  - Produced capital covers dwellings and other structures, machinery and equipment, and inventories.
  - Non-renewable natural capital comprises subsoil energy and mineral assets valued at the present value of net resource rents.
  - Reference year: 2019 for all countries unless stated otherwise.
  - Removal of non-renewable component:
    - NaturalShare = Nonrenewable / (Produced + Nonrenewable)
    - K_adj = RNNA × (1 − NaturalShare)
  - Capital–output ratio: K_adj / Y = K_adj / RGDPNA
  - Note: Because ratios are unit-free, differing valuation bases do not affect the measure.
- A.2 Consumption-to-GDP ratio
  - Source: World Bank World Development Indicators (indicator NE.CON.PRVT.ZS).
  - Reference year: 2023 for all countries, with Japan taken from 2022 as the most recent year available.
  - Published percentages are converted to ratios and reported to two decimals.
- A.3 Investment-to-output ratio
  - Source: IMF World Economic Outlook (April 2025), series NIDNGDP.
  - Reference year: 2023 for the baseline sample.
  - Values converted from per cent to ratios and reported to two decimals.
  - Series label retained as published by the IMF.
- A.4 Projected output ratio (long-run real output expansion 2024–2050)
  - Use staff-prepared median real GDP growth rates from the WEO for 2030, 2035, and 2040.
  - Fit linear trend g_t = α + β · t to these three points to extrapolate a terminal rate for 2050.
  - Annual growth path held piecewise constant:
    - 2025–2034 take the 2030 rate;
    - 2035–2039 the 2035 rate;
    - 2040–2044 the 2040 rate;
    - 2045–2050 the extrapolated 2050 rate.
  - Converting rates from per cent to decimals, cumulative expansion factor:
    - Y_2050 / Y_2024 = ∏_{t=2025}^{2050} (1 + g_t / 100)
  - Coverage follows the baseline sample unless otherwise noted.
- A.5 Health expenditure-to-GDP ratio
  - Source: World Bank WDI (indicator SH.XPD.CHEX.GD.ZS).
  - Reference year: 2022 for all 16 countries.
  - Percentages converted to ratios and reported to two decimals.
- A.6 Education expenditure-to-GDP ratio
  - Source: World Bank WDI (indicator SE.XPD.TOTL.GD.ZS).
  - Reference year: 2022 for all countries except Korea, for which 2021 is the latest available observation.
  - Percentages converted to ratios and reported to two decimals.
- A.7 Individual and corporate income tax-to-GDP ratios
  - Source: IMF Government Finance Statistics: TaxIncI (taxes on income and profits of individuals) and TaxIncC (taxes on income and profits of corporations), both as share of GDP.
  - Reference year: 2022 as the latest broadly comparable year across the sample.
  - Converted to ratios and reported to two decimals.
- A.8 Consumption tax-to-GDP ratio
  - Source: IMF World Revenue Longitudinal Database series TaxSal (aggregates VAT, general sales taxes, excises, and other taxes on final consumption).
  - Reference year: 2022 for the baseline sample.
  - Published percentages converted to ratios and reported to two decimals.
- A.9 Debt-service (interest payments)-to-GDP ratio
  - Source: IMF indicator “Interest paid on public debt, percent of GDP”.
  - Measure reflects interest actually paid within the year and abstracts from principal amortization.
  - For each country select the most recent strictly positive observation and report the resulting ratio to two decimals when at least 0.01, and to four decimals otherwise.
  - Reference year predominantly 2023 for Australia, Bangladesh, Cambodia, India, Indonesia, Japan, Malaysia, New Zealand, the Philippines, Thailand, and Vietnam. Sri Lanka uses 2022 as the latest available observation.

### Appendix B — Data for structural parameters: construction and harmonization
- Purpose: document data sources and construction methods for structural parameters used in model calibration: total factor productivity (TFP), education-based composition of the labor force (skill shares), and statutory retirement ages.
- Presentation conventions:
  - Table?? summarizes calibrated structural parameters combining TFP, skill shares, and statutory retirement age.
  - TFP values expressed as unit-free ratios relative to Japan (= 1.00).
  - Skill shares sum to one by construction.
  - Retirement ages reflect the most recent statutory ages reported by the International Social Security Association (ISSA) and are expressed in years.
  - All entries rounded to two decimals for presentation.
- B.1 Total Factor Productivity (TFP)
  - Source: Penn World Table (PWT) variable ctfp from version 10.01.
  - Series reports TFP level at current purchasing power parities (PPPs), in constant prices, benchmarked to the United States equalling one.
  - Use latest common year available in the dataset: 2019.
  - Normalize each country’s TFP level to Japan:
    - TFP_{i,rel} = ctfp_i / ctfp_JPN
    - Japan placed at one; other countries expressed as proportion of Japan’s level.
  - Missing 2019 observations imputed using the mean of the country’s predefined income group—advanced economies, emerging-market and middle-income economies, or low-income developing countries—computed from countries in that group with available data.
  - Resulting series reported as unit-free index relative to Japan, rounded to two decimal places.
- B.2 Skill Shares (education-based composition)
  - Data: World Bank WDI educational attainment indicators:
    - Proportion of adults 25+ completed upper-secondary education: indicator SE.SEC.CUAT.UP.ZS.
    - Proportion completed bachelor’s degree or equivalent: indicator SE.TER.CUAT.BA.ZS.
  - Values converted from percentages to proportions.
  - Benchmark to Japan baseline shares: (0.65, 0.25, 0.10) for (Low Skill Share, Middle Skill Share, High Skill Share).
  - For each country compute:
    - High Skill Share_i = 0.10 × (Graduate_i / Graduate_JPN)
    - Middle Skill Share_i = 0.25 × (UpperSecondary_i / UpperSecondary_JPN)
    - Low Skill Share_i = 1 − Middle Skill Share_i − High Skill Share_i
  - To prevent implausible upscaling where Graduate_i exceeds Japan’s level, replace offending attainment rate(s) with the mean for the country’s pre-assigned income group (advanced economies, emerging-market and middle-income economies, or low-income developing countries). Group means computed from available observations within each group for the relevant year. Advanced economies are not subject to this cap.
  - Resulting shares form a unit-sum triplet for each country. Presentation rounding: two decimals in general, extended to four decimals where a value would otherwise round to 0.00.

*AI and Economic Divergence in Asia — Working Paper No. WP/2026/166*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2026/english/wpiea2026166-source-pdf.pdf_