## wpiea2025076-print-pdf

## Source details

**Canonical URL:** [wpiea2025076-print-pdf](https://www.imf.org/-/media/files/publications/wp/2025/english/wpiea2025076-print-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2025/english/wpiea2025076-print-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2025/english/wpiea2025076-print-pdf.pdf.json)

---

### 1. Introduction — Overview, approach, and headline projections
- Objective: Link three drivers—AI exposure, AI preparedness, and AI access—to country-level total factor productivity (TFP) growth and feed these into the IMF’s Global Integrated Monetary and Fiscal model (GIMF) to evaluate global and regional macroeconomic impacts.
- Model and sectoral setup:
  - Macro model: IMF’s Global Integrated Monetary and Fiscal model (GIMF), multi-sector, multi-region dynamic general equilibrium model.
  - Sectoral decomposition: three sectors per region—non-tradable, tradable, and AI-intensive.
  - Transmission: AI-related productivity shocks enter through a TFP channel, scaled by country-level AI preparedness and access parameters interacting with sector-specific AI exposure parameters.
- Micro-to-macro mapping:
  - Combine task- and sector-level exposure estimates with the IMF AI Preparedness Index (AIPI) and AI access assumptions to construct country TFP shocks; run GIMF simulations to quantify impacts on growth, inflation, exchange rates, and current accounts.
- Key baseline quantitative projections (from GIMF simulations):
  - Global GDP expansion: nearly 4 percent under a high TFP growth scenario; 1.3 percent under a low TFP growth scenario.
  - Underlying TFP gains over the next decade: 0.8-2.4 percent.
  - United States projected output increases: 5.6 percent (high TFP scenario) and 1.9 percent (low TFP scenario).
  - AEs potentially realize up to twice the income gains of LICs.
- Near-term dynamics:
  - Inflation: rises in the near term due to higher investment and stronger aggregate demand, then stabilizes as TFP growth lowers prices.
  - Exchange rates & current accounts: AI-driven productivity concentrated in non-tradable sectors can lead to depreciation of AE currencies relative to EMs and LICs; example: U.S. non-tradable sector gains could lower non-tradable prices and the value of the U.S. dollar, improving the U.S. current account despite larger AI investments.
- Principal limitations:
  - Stylized labor market representation (no explicit unemployment dynamics or labor force participation shifts).
  - Productivity shocks depend on adoption assumptions and coarse sectoral aggregation.
  - Distributional effects within countries (wage inequality, wealth concentration) not addressed.

### 2. Drivers of AI Adoption — Exposure, Preparedness, and Access
- Conceptual framework: three critical elements—Exposure, Preparedness, Access.
- Exposure (occupation- and sector-level):
  - Cross-country shares of “high-exposure” jobs: up to 60 percent in AEs; 42 percent in EMs; 26 percent in LICs.
  - Implication: AEs have larger immediate potential for productivity gains and higher displacement risk; EMs and LICs face subdued integration but could gain with appropriate infrastructure and skills.
- Preparedness (AI Preparedness Index, AIPI):
  - Dimensions: digital infrastructure; innovation and economic integration; human capital and labor market policies; regulation and ethics.
  - Coverage: measures for 174 countries; each dimension normalized and averaged across sub-indicators.
  - Stylized pattern: AEs and some EMs show stronger preparedness; many LICs remain severely underprepared.
- Access (hardware, data, geopolitical constraints):
  - Key inputs: semiconductor manufacturing, high-capacity data centers, robust data ecosystems, global partnerships.
  - Observed concentration: United States (with partners) and China possess extensive capabilities.
  - Policy/geopolitical actions: export restrictions and outbound investment reviews in semiconductors, AI, and quantum technologies affect future access.
  - Implication: limited access to chips, computing clusters, or data repositories can constrain EMs and LICs and widen global inequality.

### 3. Model, Calibration, and Scenarios — GIMF-GVC, calibration, and TFP shock construction
- Model features:
  - Uses GIMF-GVC—an extension adding a second tradable sector featuring global value chains to capture spillovers and network effects.
  - Short term: nominal and real rigidities amplify demand effects. Medium/long term: permanent productivity changes drive capital and labor effects.
  - Households: two types per region (OLG forward-looking and liquidity-constrained).
  - Firms: monopolistic competition; produce non-tradable, tradable, and AI-intensive goods; AI-intensive sector uses sector composite output as intermediate consumption domestically or via global value chains.
  - Investment: real adjustment costs, financial accelerator mechanism, endogenous corporate risk premia.
  - International finance & trade: long-term real global interest rate set by global savings-investment balance; bilateral trade flows tracked by four types (two final, two intermediate).
  - Nominal rigidities: menu costs following Ireland (2001) version of Rotemberg (1982).
  - Macroeconomic policy: fiscal policy targets debt sustainability and output stabilization; monetary policy follows inflation forecast-based targeting (affects dynamics first five to ten years).
- Calibration—regions (seven):
  - China (CHN); EMA; EML; EU and Switzerland (EUS); Other advanced economies (OAD); the United States (USA); Rest of the World (ROW, mainly LICs).
- Sectoral shares (selected exact values from Table 1):
  - Share in World GDP (percent) by region: CHN 16.8, EMA 8.5, EML 7.1, EUS 18.2, LIC 1.5, OAD 17.4, USA 25.1, ROW 5.5.
  - GDP by production (percent of regional GDP): Non-tradable: CHN 60.0, EMA 55.7, EML 61.3, EUS 64.2, LIC 51.6, OAD 66.1, USA 68.0, ROW 56.3. Tradable: CHN 27.9, EMA 33.6, EML 30.0, EUS 22.9, LIC 36.5, OAD 19.7, USA 15.6, ROW 37.3. AI-intensive: CHN 12.1, EMA 10.7, EML 8.7, EUS 12.9, LIC 11.9, OAD 14.1, USA 16.3, ROW 6.4.
  - Value added note: the US has 16.3 percent of GDP from the AI-intensive sector.
- Elasticities and markups (exact calibration points):
  - Elasticities (all regions): Consumption Domestic / imported 1.50, Different regions 1.50. Investment Domestic / imported 1.50, Different regions 1.50. Tradable Domestic / imported 1.50, Different regions 1.50. AI-intensive Domestic / imported 0.60, Different regions 0.60, Capital-labor / intermediate inputs 0.60.
  - Markups (Domestic / Imported) for AE/EM and LIC groups: Consumption: 1.10 / 1.05 / 1.01. Investment: 1.05 / 1.05 / 1.01. Tradable: 1.20 / 1.05 / 1.01. AI-intensive: 1.25 / 1.05 / 1.01. Non-tradable: 1.20 (domestic only listed).
- TFP shock construction (exact formulas cited):
  - Region-level formula: TFP_shock_i,t = (AI_preparedness_i × AI_exposure_i) / (AI_preparedness_b × AI_exposure_b) × AI_Access_i × TFP_shock_b,t
  - Sectoral allocation: TFP_shock_i,j,t = (AI_exposure_i,j / AI_exposure_i) × TFP_shock_i,t
- AI exposure by region and sector (Table 4 exact values):
  - CHN: Tradable 0.30, Non-tradable 0.43, AI-intensive 0.58.
  - EMA: Tradable 0.18, Non-tradable 0.41, AI-intensive 0.72.
  - EML: Tradable 0.24, Non-tradable 0.46, AI-intensive 0.72.
  - EUS: Tradable 0.40, Non-tradable 0.60, AI-intensive 0.82.
  - LIC: Tradable 0.17, Non-tradable 0.47, AI-intensive 0.72.
  - OAD: Tradable 0.36, Non-tradable 0.56, AI-intensive 0.77.
  - USA: Tradable 0.43, Non-tradable 0.67, AI-intensive 0.85.
  - ROW uses LIC values.
- AI preparedness (AIPI) calibration notes:
  - AIPI is the simple average of normalized components: digital infrastructure, human capital and labor market policies, digital innovation and economic integration, regulation and ethics.
  - Preparedness values listed: 0.77, 0.64, 0.50, 0.50, 0.68, 0.35, 0.73 (region-level preparedness entries as presented in source).
- Benchmark TFP growth assumptions and literature comparisons:
  - Literature: Acemoglu (2025) → TFP gains around 0.7 percent over next 10 years. Aghion and Bunel (2024) → cumulative median 6.8 percent TFP growth over 10 years.
  - Paper’s benchmarks: rescaled UK-based simulations yield economy-wide values of 1.5 and 4.3 percent for the US under low and high benchmarks.
- Scenario definitions:
  - Baseline: AI_Access_i = 1 for all i; heterogeneity from exposure and preparedness.
  - Limited AI Access: AI_Access_i = 1 for AEs and China, AI_Access_i = 0.5 for other EMs and LICs.
  - Enhanced AI Preparedness: reforms and support raise EMs and LICs preparedness to peer-top levels.
  - Baseline analyzed under two TFP growth alternatives: Low and High.

### 4. Baseline Results — Productivity, GDP, inflation, exchange rates, and scenarios
- Impact on productivity (exact outcomes):
  - High-TFP growth baseline: Global TFP increases by 1.8 percent in five years, and 2.4 percent in ten years.
  - Low-TFP growth baseline: Global TFP increases by 0.6 percent in five years, and 0.8 percent in ten years.
  - Sectoral ordering (both assumptions): AI-intensive > non-tradable > tradable.
- Impact on GDP (deviations from steady state, exact values):
  - World GDP:
    - High-TFP growth: 2.4 percent in five years; nearly 4 percent in ten years.
    - Low-TFP growth: 0.8 percent in five years; 1.3 percent in ten years.
  - Cross-country heterogeneity after 10 years (high TFP / low TFP):
    - U.S.: 5.4 (1.8) percent.
    - EU and Switzerland: 4.4 (1.5) percent.
    - OAD: 4.7 (1.6) percent.
    - EML: 3.2 (1.1) percent.
    - EMA: 3.0 (1.0) percent.
    - LIC: 2.7 (0.9) percent.
    - ROW: 2.2 (0.7) percent.
    - China: 3.5 (1.2) percent.
- Impact on inflation:
  - Short term: modest increase in inflation due to higher aggregate demand from expected income gains; central banks tighten policy rates modestly.
  - Long term: supply capacity increase and capital accumulation lead to a decline in inflation.
  - Low TFP scenario: qualitatively similar but smaller magnitudes.
- Impact on exchange rates and current accounts (high TFP, 10-year horizon):
  - AE currencies, notably the US Dollar, experience a modest real depreciation.
  - Mechanism: relatively higher productivity gains in non-tradables lower relative prices of non-tradables (inverse Balassa-Samuelson effect); tradable sector faces higher wages and international competitiveness maintained via real exchange rate depreciation.
  - Current accounts: US and other AEs experience moderate increases in current account balances; depreciation-induced export effect dominates higher investment-induced deficit effect.
- Limited AI Access scenario (assumptions and impacts):
  - Assumption: EMs (excluding China) and LICs constrained (AI_Access_i = 0.5).
  - Impacts after 10 years (high TFP baseline comparison): EMs and LICs’ real GDP increase falls on average by around 1 percentage point relative to baseline; slight changes to AEs and China.
  - Output losses more pronounced in EMA and EML due to higher exposure and preparedness that cannot be utilized under access constraints.
- Enhanced AI Preparedness scenario (exact preparedness changes and effects):
  - Assumption: preparedness index lifted for EM and LIC groups to peer-top levels; no access restrictions.
  - Specific assumed increases in preparedness indices:
    - EMA: increase to 0.63 from 0.50.
    - EML: increase to 0.59 from 0.50.
    - LIC and ROW: increase to 0.48 from 0.38.
  - Results: cross-region inequality persists but is partially offset; global GDP level marginally benefits; additional output gains concentrated in AI-intensive sector in EM and LIC groups.
- Relation to historical technological episodes:
  - ICT contribution example: IT sector contribution to TFP growth in the U.S. ≈ 0.47 percentage points per year in 1995-2006 (Jorgenson et al. (2008)) — comparable magnitude to high TFP baseline.
  - Caveat: simulations do not explicitly incorporate possible initial productivity slowdowns during adoption phases.

### 5. Policy implications and “Mind the AI gap” — Priorities for EMs and LICs
- Core policy takeaways:
  - Preparedness matters: institutions, digital infrastructure, and skilled workforce significantly influence ability to realize AI gains.
  - Access matters: availability of chips, data, and computing infrastructure and geopolitical controls are decisive.
  - Sectoral focus: AI gains concentrated in non-tradable sectors can produce macro outcomes that differ from tradable-sector-driven shocks.
- Policy recommendations:
  - For well-prepared economies (most AEs and some EMs): strengthen digital innovation capacity; adapt legal and ethical frameworks to govern and foster AI.
  - For weakly prepared economies (some EMs and most LICs): prioritize investment in digital infrastructure and human capital to build toward second-generation preparedness.
  - Public investment priorities: fundamental research; necessary infrastructure (particularly in EMDEs); public sector applications with high social returns (education, health care, government administration).
  - Technology access and sharing: enhancing access to hardware, software, and relevant data can reduce the gap, but geopolitical and national security considerations may constrain sharing.
  - Potential for leapfrogging: breakthroughs requiring less computing capacity (example cited in source) and open-source code/data access could help less-developed EMs and LICs narrow the gap.
- Illustrative development example:
  - Kenya’s M-Pesa cited as a technology leapfrogging case that broadened financial inclusion and catalyzed a digital economy ecosystem.

### 6. Conclusions and research avenues
- Macro growth and distributional summary:
  - AI-induced productivity gains could lift global GDP by up to 4 percent in a high TFP growth scenario over the next decade.
  - AEs benefit disproportionately; estimated growth impact in AEs could be more than double that in LICs.
  - Improved access by EMs and LICs to AI data and technologies can generate positive spillbacks to AEs via higher global growth.
- Role of preparedness and infrastructure:
  - Enhanced AI preparedness highlights importance of institutions and digital infrastructure; policy action can reduce the cross-country AI gap but not fully eliminate it.
- Exchange rate and sectoral considerations:
  - AI’s cross-sector effects (especially in non-tradables) can produce real exchange rate depreciation for AEs and bolster external competitiveness beyond directly AI-boosted sectors.
- Caveats and suggested future research directions:
  - Refine labor market representation (unemployment dynamics, skill mismatches).
  - Deepen analysis of within-country distributional consequences.
  - Explore policy-varying scenarios (infrastructure investments, fiscal settings) that alter pace and breadth of AI adoption.

*IMF Working Paper: The Global Impact of AI: Mind the Gap (Introduction, Drivers of AI Adoption, Model and Calibration, Baseline Results, Policy Implications, Conclusions).*

### 1. Introduction ........................................................................................................

### 1. Introduction

### Overview and objectives
- AI holds the promise of raising productivity, stimulating innovation, and altering the nature of work globally, but adoption to date is uneven across AEs, EMs, and LICs.
- The paper links three drivers—AI exposure, AI preparedness, and AI access—to country-level total factor productivity (TFP) growth and feeds these into the IMF’s Global Integrated Monetary and Fiscal model (GIMF) to evaluate global and regional macroeconomic impacts.
- The analysis covers AEs, EMs, and LICs and uses microdata on task- and sectoral-level job exposures combined with country-specific measures of AI preparedness and assumptions on AI access.

### Model structure and approach
- Macro model: IMF’s Global Integrated Monetary and Fiscal model (GIMF), a multi-sector, multi-region dynamic general equilibrium model.
- Sectoral decomposition: each region’s economy is divided into three sectors—non-tradable, tradable, and AI-intensive—to capture cross-country and within-country heterogeneities in AI adoption.
- Transmission: AI-related productivity shocks enter through a TFP channel, scaled by the interaction of country-level AI preparedness and access parameters with sector-specific AI exposure parameters.
- Two-pronged strategy:
  - Micro-to-macro mapping: combine task- and sector-level exposure estimates with country AIPI (AI Preparedness Index) and AI access assumptions to construct TFP shocks by country.
  - GIMF simulations: use these TFP shocks to quantify global and regional impacts on growth, inflation, exchange rates, and current accounts.

### Key quantitative baseline projections (from GIMF simulations)
- Global GDP expansion:
  - nearly 4 percent under a high TFP growth scenario.
  - 1.3 percent under a low TFP growth scenario.
- Underlying TFP gains over the next decade: 0.8-2.4 percent.
- Country highlights:
  - United States projected output increases: 5.6 percent (high TFP scenario) and 1.9 percent (low TFP scenario).
  - China ranks near the middle in growth outcomes—high AI preparedness but relatively lower exposure due to the greater relative importance of manufacturing versus services in comparison to AEs.
  - AEs potentially realize up to twice the income gains of LICs.

### Near-term macroeconomic dynamics
- Inflation: projected to rise in the near term because of higher investment and stronger aggregate demand, before stabilizing as TFP growth exerts downward pressure on prices.
- Exchange rates and current accounts:
  - Conventional expectation (Balassa-Samuelson): productivity gains in tradables lead to real appreciation.
  - GIMF result: AI-driven productivity shocks concentrated in non-tradable sectors can be accompanied by a depreciation of AEs’ currencies relative to EMs and LICs due to downward pressure on non-tradable prices.
  - Example: AI-driven productivity gains in U.S. non-tradable sectors (education and healthcare) would lower the relative price of non-tradable services and the value of the U.S. dollar, improving the U.S. current account even with larger AI investments.

### Limitations and caveats
- Labor market representation in GIMF is stylized:
  - Lacks explicit modeling of unemployment dynamics and shifts in labor force participation.
- Productivity shocks rely on assumptions about AI adoption patterns that may not capture heterogeneous diffusion speeds across sectors, regions, and timeframes.
- Sectoral aggregation is coarse and may overlook within-sector heterogeneity and organizational/managerial complexities that determine actual AI deployment.
- Distributional effects within countries—such as wage inequality or wealth concentration—are not addressed.

---

### 2. Drivers of AI Adoption

### Conceptual framework
- Three critical elements shape a country’s capacity to benefit from AI:
  - Exposure: how extensively AI can affect a country’s workforce and industries (occupation- and sector-level overlap with AI capabilities).
  - Preparedness: readiness of institutions, digital infrastructure, workforce, regulatory frameworks, and governance structures to adopt AI (measured by the IMF’s AI Preparedness Index, AIPI).
  - Access: availability of specialized technologies, data, and infrastructure (chips, data centers, data ecosystems) and global partnerships necessary to harness AI.

### Exposure — occupational and sectoral differences
- Measurement approach: task-based analyses mapping overlap between AI capabilities and occupational requirements; complementarity potential incorporated to identify likelihood of benefiting from AI adoption.
- Cross-country shares of “high-exposure” jobs (from cited micro evidence):
  - up to 60 percent in AEs.
  - 42 percent in EMs.
  - 26 percent in LICs.
- Implication: AEs have larger immediate potential for productivity gains and higher displacement risk; many EMs and LICs face more subdued integration but could gain substantially with appropriate digital infrastructure and skills.

### Preparedness — AI Preparedness Index (AIPI)
- AIPI dimensions: digital infrastructure; innovation and economic integration; human capital and labor market policies; regulation and ethics.
- Data sources and construction: measures cover 174 countries and draw on official data and perception surveys, with each dimension normalized and averaged across sub-indicators (e.g., digital connectivity, labor policies, R&D investment, regulatory adaptability).
- Stylized pattern: AEs and some EMs generally exhibit stronger digital infrastructure, skilled labor, ecosystems for innovation and economic integration, and governance frameworks; many LICs remain severely underprepared.

### Access — hardware, data, and geopolitical constraints
- Key inputs: semiconductor manufacturing, high-capacity data centers, robust data ecosystems, global partnerships.
- Observed concentration: United States (with partners) and China possess extensive capabilities enabling leadership in AI research and deployment.
- Policy actions and geopolitical factors: export restrictions and reviews of outbound investments in semiconductors, AI, and quantum technologies are cited as affecting future access.
- Implication: limited access to chips, large-scale computing clusters, or robust data repositories can constrain EMs and LICs, potentially cementing advantages of leading economies and widening global inequality.

---

### 3. Scenarios, International Policy Findings, and Robustness Notes

### Scenarios analyzed
- Baseline high and low TFP growth scenarios (quantitative results summarized above).
- Two alternative scenarios to probe policy and access:
  - “Limited AI Access” scenario: restrictions or limited diffusion of AI-specific inputs across countries.
  - “Enhanced AI Preparedness” scenario: policy-driven improvements in readiness (AIPI) that raise countries’ capacity to translate exposure into productivity gains.

### Scenario outcomes and policy insights
- Limited AI Access:
  - Constrains AI diffusion and magnifies cross-country disparities.
  - In this scenario, the impact on U.S. growth exceeds that on LICs by more than threefold.
- Enhanced AI Preparedness:
  - Policy-driven improvements in readiness can materially mitigate disparities.
  - However, even with enhanced preparedness, cross-country disparities persist—policy interventions can reduce but are unlikely to fully eliminate uneven impacts of AI adoption.

### Policy-relevant takeaways
- Preparedness matters: strong institutions, sophisticated digital infrastructure, and a skilled workforce significantly influence countries’ ability to realize AI-driven productivity gains.
- Access matters: availability of chips, data, and computing infrastructure (and geopolitical controls over them) can be decisive for countries’ ability to innovate and compete.
- Sectoral focus: AI productivity gains concentrated in non-tradable sectors can produce macro outcomes (inflation, exchange rates, current accounts) that differ from canonical tradable-sector-driven productivity shocks.
- Policy implications (derived from scenarios and model structure):
  - Investments to raise AIPI components (digital infrastructure, human capital, regulatory frameworks) can materially improve outcomes for EMs and LICs.
  - International cooperation and policies that broaden global access to AI-specific inputs could reduce divergence; conversely, access restrictions risk amplifying global inequality.

---

*IMF Working Paper: “The Global Impact of AI: Mind the Gap” (Introduction and Drivers of AI Adoption).*

### 3. Model, Calibration and Scenarios

### 3. Model, Calibration and Scenarios

### Model
- The analysis uses the IMF’s Global Integrated Monetary and Fiscal Model (GIMF), a multi-region, multi-sector dynamic general equilibrium model.
- Applications employ GIMF-GVC—an extended version of the standard two-sector GIMF (tradable and non-tradable)—which adds a second tradable sector featuring global value chains to capture broader spillovers, including supply-side and network effects relevant for analyzing AI.
- GIMF focuses on dynamic effects of shocks:
  - Short term: nominal and real rigidities amplify effects on aggregate demand.
  - Medium to long term: effects mainly from permanent productivity changes on capital and labor.
- Tradeoff: dynamic and policy richness vs. a more stylized treatment of intra-temporal channels and less sectoral detail and full input-output relationships.
- Households:
  - Two types per region: (i) finitely lived overlapping generation (OLG) households—forward-looking on consumption, savings, labor supply; hold financial assets including government and corporate bonds, and net foreign assets; (ii) liquidity-constrained households—consume all income every period and follow labor supply set by OLG households.
- Firms:
  - Operate in monopolistically competitive markets, set profit-maximizing prices subject to nominal rigidities and a residual demand schedule.
  - In GIMF-GVC firms produce non-tradable, tradable, and AI-intensive goods and services. The AI-intensive category is a composite of tradable sectors most exposed to AI (exposure indicator from Cazzaniga et al. (2024)); these firms use the sector’s composite output as intermediate consumption sourced domestically or via AI-intensive global value chains.
  - Non-tradable and tradable sectors use Cobb-Douglas technology to combine labor and capital.
  - No foreign ownership in GIMF.
- Investment:
  - Chosen by firms to maximize profits subject to real adjustment costs; investment requires inputs sourced both domestically and from foreign regions which are imperfect substitutes.
  - Includes a financial accelerator mechanism as in Bernanke, Gertler and Gilchrist (1999); retained earnings are insufficient so firms borrow from financial intermediaries; corporate risk premia are determined endogenously.
- International finance and trade:
  - Long-term real global interest rate set by balancing global savings and investment.
  - Real exchange rate adjusts each country’s saving position (current account and net foreign assets) relative to the global pool.
  - Bilateral trade flows are tracked and separated into four types: two types of final goods (consumption or investment) and two types of intermediate goods and services (tradable or AI-intensive); trade reacts to demand, supply and pricing (terms of trade and bilateral real exchange rates).
- Nominal rigidities:
  - Modeled as menu costs following the Ireland (2001) version of Rotemberg (1982); adjustment costs apply to changes in both the price level and the inflation rate.
- Macroeconomic policies:
  - Fiscal policy in each region pursues twin objectives: debt sustainability in the long run (ensuring a non-explosive government-debt-to-GDP path) and output stabilization in the short run.
  - Monetary policy follows inflation forecast-based targeting; it shapes dynamics over the first five to ten years but has no implications for long-term real economy outcomes.
- Modeling of AI:
  - An AI shock is applied with region- and sector-specific sizes.
  - The approach evaluates changes in TFP across scenarios and countries/regions.
  - Unlike task-based models, this approach does not focus on worker replacement or augmentation; within-country inequality effects are not the paper’s focus.

### Calibration
- Regions in the GIMF calibration (seven regions):
  - China (CHN)
  - Emerging Market Economies Asia, Central Asia, Russia, etc. (EMA)
  - Emerging Market Economies Latin America, Middle East, Africa, etc. (EML)
  - EU and Switzerland (EUS)
  - Other advanced economies (OAD)
  - the United States (USA)
  - Rest of the World (ROW) — represents mainly LICs
- Data sources used for calibration:
  - OECD Inter-Country Input-Output database for input and trade shares, and value added by country and sector (Table 1).
  - IMF’s Government Finance Statistics database for fiscal ratios.
- Sectoral structure:
  - GDP produced in three sectors:
    - Non-tradables: largely construction and services.
    - Tradables: mainly agriculture, mining, manufacturing and transportation.
    - AI-intensive: tradable goods and services particularly exposed to AI (sectoral index per Cazzaniga et al. (2024)); examples include pharmaceuticals, computers, telecommunications and finance.
  - The AI-intensive sector features global value chains through roundabout production—firms use output from the sector, either domestic or foreign, as intermediate input.
- Calibration notes on sectoral importance:
  - Table 1 reports sector importance by region and Annex I provides granular definitions.
  - Value added by sector: the US has 16.3 percent of GDP coming from the AI-intensive sector.
  - The US trades less in AI-intensive goods and services as a share of GDP compared to other regions because the U.S. economy trades much less overall with the rest of the world relative to other regions.
  - Other advanced economies (OAD) and the EU-Switzerland region (EUS) also have a large share of total activity in the AI-intensive sector.

### Scenarios and Simulation Focus
- Framework purpose:
  - Simulate domestic and global macroeconomic impacts of AI by evaluating changes in TFP across different scenarios and countries/regions.
- Scope limitations:
  - The model emphasizes macroeconomic dynamics and policy interactions rather than intra-country labor market reallocations or inequality outcomes.
  - The AI-intensive sector is modeled to capture network externalities and cross-border spillovers via global value chains.

*Source: 3. Model, Calibration and Scenarios, wpiea2025076-print-pdf*

### 14.1 and 12.9 percent of GDP being produced in that sector, respectively. The AI-intensive sector in China

### wpiea2025076-print-pdf - 14.1 and 12.9 percent of GDP being produced in that sector, respectively. The AI-intensive sector in China

### Sector sizes and shares (Table 1)
- Share in World GDP (percent) by region: CHN 16.8, EMA 8.5, EML 7.1, EUS 18.2, LIC 1.5, OAD 17.4, USA 25.1, ROW 5.5.
- GDP by expenditure (percent of regional GDP):
  - Consumption: CHN 69.7, EMA 71.8, EML 80.1, EUS 75.2, LIC 74.5, OAD 75.3, USA 77.3, ROW 74.2.
  - Investment: CHN 30.3, EMA 28.2, EML 19.9, EUS 24.8, LIC 25.5, OAD 24.7, USA 22.7, ROW 25.8.
  - Exports: CHN 16.6, EMA 23.3, EML 22.4, EUS 19.7, LIC 32.8, OAD 23.1, USA 10.7, ROW 29.3.
    - o/w AI-intensive exports (percent of regional GDP): CHN 2.1, EMA 2.7, EML 1.3, EUS 2.8, LIC 2.7, OAD 3.5, USA 1.9, ROW 1.2.
  - Imports: CHN 16.6, EMA 23.3, EML 22.4, EUS 19.7, LIC 32.8, OAD 23.1, USA 10.7, ROW 29.3.
    - o/w AI-intensive imports (percent of regional GDP): CHN 1.9, EMA 2.6, EML 2.4, EUS 3.0, LIC 2.5, OAD 3.0, USA 1.5, ROW 3.1.
- GDP by production (percent of regional GDP):
  - Non-tradable: CHN 60.0, EMA 55.7, EML 61.3, EUS 64.2, LIC 51.6, OAD 66.1, USA 68.0, ROW 56.3.
  - Tradable: CHN 27.9, EMA 33.6, EML 30.0, EUS 22.9, LIC 36.5, OAD 19.7, USA 15.6, ROW 37.3.
  - AI-intensive: CHN 12.1, EMA 10.7, EML 8.7, EUS 12.9, LIC 11.9, OAD 14.1, USA 16.3, ROW 6.4.
- Note: “These shares illustrate the relative size of the sector, encompassing industries that might be substantially impacted by AI. The actual adoption of AI is heavily influenced by the share of occupations exposed to AI in the sector, as well as AI preparedness.”

### Elasticities and substitutability (Table 2)
- Elasticities (all regions):
  - Consumption: Domestic / imported 1.50, Different regions 1.50.
  - Investment: Domestic / imported 1.50, Different regions 1.50.
  - Tradable: Domestic / imported 1.50, Different regions 1.50.
  - AI-intensive: Domestic / imported 0.60, Different regions 0.60, Capital-labor / intermediate inputs 0.60.
- Modeling notes:
  - Elasticities of substitution for the AI-intensive sector are set at a slightly lower level than other goods.
  - Substitutability between factors and intermediate inputs in the AI-intensive sector is set lower to reflect supplier-specific relations.
  - Elasticities are based on standard GIMF calibration.

### Markups calibration by sector and country group (Table 3)
- Markups (Domestic / Imported) for AE/EM and LIC groups:
  - Consumption: 1.10 / 1.05 / 1.01
  - Investment: 1.05 / 1.05 / 1.01
  - Tradable: 1.20 / 1.05 / 1.01
  - AI-intensive: 1.25 / 1.05 / 1.01
  - Non-tradable: 1.20 (domestic only listed)
- Modeling notes:
  - Domestic markups for AI-intensive firms set higher based on assumed higher market power from network effects and winner-take-all dynamics.
  - LICs assumed to have lower markups for imported goods (reflecting lower market power on global markets).
  - “Results are robust to alternative assumptions about markups for the AI-intensive sector.”

### Calibration of productivity (TFP) shocks and formulas
- Region-specific TFP shock formula (equation (1)):
  - TFP_shock_i,t = (AI_preparedness_i × AI_exposure_i) / (AI_preparedness_b × AI_exposure_b) × AI_Access_i × TFP_shock_b,t
- Sectoral allocation formula (equation (2)):
  - TFP_shock_i,j,t = (AI_exposure_i,j / AI_exposure_i) × TFP_shock_i,t
- Implementation notes:
  - Two benchmark TFP shocks (high and low) are used.
  - Table 4 reports AI exposure and preparedness by region and sector to rescale TFP shocks.
  - Most scenarios assume AI_Access_i = 1 for all i. A Limited AI Access Scenario sets AI_Access_i = 1 for AEs and China, and AI_Access_i = 0.5 for other EMs and LICs.

### AI exposure by region and sector (Table 4 values)
- China (CHN): Tradable 0.30, Non-tradable 0.43, AI-intensive 0.58.
- EMA: Tradable 0.18, Non-tradable 0.41, AI-intensive 0.72.
- EML: Tradable 0.24, Non-tradable 0.46, AI-intensive 0.72.
- EUS: Tradable 0.40, Non-tradable 0.60, AI-intensive 0.82.
- LIC: Tradable 0.17, Non-tradable 0.47, AI-intensive 0.72.
- OAD: Tradable 0.36, Non-tradable 0.56, AI-intensive 0.77.
- USA: Tradable 0.43, Non-tradable 0.67, AI-intensive 0.85.
- ROW uses LIC values.
- Sectoral pattern: AI-intensive sector has the highest exposure, non-tradable second, tradable lowest.

### AI preparedness (IMF AI Preparedness Index, AIPI)
- AIPI constructed as simple average of normalized components: digital infrastructure, human capital and labor market policies, digital innovation and economic integration, regulation and ethics.
- AIPI for model regions obtained by averaging country values.
- Ranking: US and other AEs at the top, followed by EMs and LICs.
- Table 4 also provides preparedness values (listed in the source text as trailing numbers): 0.77, 0.64, 0.50, 0.50, 0.68, 0.35, 0.73 (these correspond to region-level preparedness entries in Table 4 as presented).

### Benchmark TFP growth assumptions and literature comparisons
- Literature estimates:
  - Acemoglu (2025): TFP gains of around 0.7 percent over the next 10 years → annual TFP growth increase of about 0.07 percentage points. Based on 4.6 percent of tasks being replaced profitably by AI and a 27 percent increase in productivity due to labor cost savings.
  - Aghion and Bunel (2024): cumulative median 6.8 percent TFP growth over 10 years → 0.7 per year; range 0.7 to 12.4 percent cumulative over ten years. Based on: 60 percent of tasks exposed in developed countries, 50 percent of exposed tasks profitable to use AI, and labor saving costs of 40 percent.
- Paper’s benchmark choices:
  - Uses two alternative benchmark assumptions rooted in UK simulations of Cazzaniga et al. (2024) and Rockall et al. (forthcoming) finding TFP increases between 1.3 and 3.4 percent over ten years.
  - Rescaling yields economy-wide values of 1.5 and 4.3 percent for the US under low and high benchmarks (as stated: “Rescaling these impacts ... delivers economy-wide values of 1.5 and 4.3 percent for the US”).
  - “The low benchmark is about double the impact predicted by Acemoglu (2025). The high benchmark is closer to (but still significantly smaller than) the proposed median TFP gain of Aghion and Bunel (2024).”
- Resulting pattern:
  - Higher overall TFP shocks observed in regions with greater exposure and preparedness (notably the US and other AEs).
  - Sectoral ordering: AI-intensive largest impacts, followed by non-tradables, then tradables.

### Scenarios evaluated
- Baseline Scenario:
  - Assumes no direct impediments to AI adoption; AI_Access_i = 1, ∀i.
  - Heterogeneous productivity gains driven by differences in AI preparedness and exposure.
- Limited AI Access Scenario:
  - EMs (except China) and LICs face constraints; AI_Access_i = 0.5, ∀i = EMA, EML, LIC.
  - Shortage of advanced processors, data-sharing platforms, and supportive infrastructure halves TFP growth in these economies.
- Enhanced AI Preparedness Scenario:
  - Domestic reforms and international support allow EMs and LICs to raise AI preparedness to peer-top levels, increasing scope for AI-driven productivity gains.
- Baseline is examined under two TFP growth alternatives:
  - Low TFP growth simulation (conservative; akin to Acemoglu (2025)).
  - High TFP growth simulation (optimistic; akin to Aghion and Bunel (2024)).

*IMF WORKING PAPERS The Global Impact of AI: Mind the Gap — INTERNATIONAL MONETARY FUND*

### 4. Baseline Results

### 4. Baseline Results

### Impact on Productivity
- High-TFP growth baseline scenario:
  - Global TFP increases by 1.8 percent in five years, and 2.4 percent in ten years.
  - AI-intensive sector experiences the largest productivity gains, followed by the non-tradable and tradable sectors.
- Low-TFP growth baseline scenario:
  - Global TFP increases by 0.6 percent in five years, and 0.8 percent in ten years.
  - AI-intensive sectors still see the largest increase, followed by non-tradables and tradables.
- Under both TFP assumptions, sectoral ordering of productivity gains is: AI-intensive > non-tradable > tradable.

*Note: The limited AI access scenario and the enhanced AI preparedness scenario are examined under the high TFP growth assumption; quantitative magnitudes differ but qualitative implications are similar.*

---

### Impact on GDP
- World GDP (deviations from steady state):
  - High-TFP growth baseline scenario:
    - World GDP rises by 2.4 percent in five years.
    - World GDP rises nearly 4 percent in ten years.
  - Low-TFP growth baseline scenario:
    - World GDP rises by 0.8 percent in five years.
    - World GDP rises by 1.3 percent in ten years.
- Cross-country heterogeneity after 10 years (high TFP / low TFP in parentheses):
  - U.S.: output increases by 5.4 (1.8) percent.
  - EU and Switzerland: output increases by 4.4 (1.5) percent.
  - Other Advanced Economies (OAD): output increases by 4.7 (1.6) percent.
  - Emerging Market Economies Latin (EML): output increases by 3.2 (1.1) percent.
  - Emerging Market Economies Asia/Central/ Russia, etc. (EMA): output increases by 3.0 (1.0) percent.
  - Low-income countries (LIC): output increases by 2.7 (0.9) percent.
  - Rest of the World (ROW): output increases by 2.2 (0.7) percent.
  - China: output increases by 3.5 (1.2) percent.
- Drivers:
  - Productivity gains and endogenous capital accumulation.
  - Policy variables (monetary and fiscal policy) adjust endogenously in response to macroeconomic conditions.
- Interpretation:
  - Benefits disproportionately accrue to Advanced Economies (AEs), reflecting stronger AI preparedness and higher occupational exposure to AI.

---

### Impact on Inflation
- Simulation results (GIMF with nominal rigidities and monetary policy, open-economy):
  - Inflation increases modestly in the short term in response to initial AI-related productivity shocks due to higher aggregate demand from concurrent and expected income gains.
  - In the long term, increased supply capacity from productivity gains and capital accumulation more than offsets demand-driven inflationary pressure, leading to a decline in inflation.
  - Central banks tighten policy rates modestly in the short term in response to inflationary pressures.
- Qualitative similarity:
  - Results under the low TFP growth assumption are qualitatively similar but reduced in magnitude compared to the high TFP growth assumption.

---

### Impact on Exchange Rates and Current Accounts
- Exchange rates (high TFP growth, 10-year horizon):
  - Broad sectoral reach and characteristics of AI shocks trigger a modest real depreciation of AE currencies, notably the US Dollar.
  - Large increase in U.S. non-tradable sector productivity relative to other countries generates a real depreciation of the US Dollar.
  - Dollar depreciation is most significant relative to LICs and EMs, less pronounced versus other AEs.
- Mechanism:
  - Inverse Balassa-Samuelson effect: relatively higher productivity gains in non-tradables lower relative prices of non-tradables; wages rise and equalize across sectors (no inter-sectoral labor market frictions in model); tradable sector faces higher wages and international competitiveness is maintained via real exchange rate depreciation.
- Current accounts (high TFP growth, deviations from steady state after 10 years):
  - US and other AEs experience a moderate increase in current account balances.
  - Two opposing forces:
    - Depreciation of AE currencies increases net exports and current account balances.
    - Higher investment driven by AI shock pushes AEs’ current account toward a deficit.
  - The depreciation-induced export effect quantitatively dominates the investment-induced deficit effect.

---

### Relation with Past Episodes of Technological Change
- Historical analogues:
  - Electricity, Internet/ICT, and other general-purpose technologies had wide-reaching macroeconomic impacts.
  - Literature on ICT in AEs finds ICT-related productivity gains and capital accumulation explained about 30-50 percent of aggregate labor productivity growth in the 1990s and early 2000s.
  - Example: Jorgenson et al. (2008) estimate the contribution of the IT sector to TFP growth in the U.S. at about 0.47 percentage points per year in 1995-2006—of similar magnitude to the TFP growth used in the high TFP growth baseline scenario.
- Caveats:
  - Simulations do not explicitly incorporate potential initial productivity slowdowns during early adoption phases of new technologies (transition costs and adjustment).
  - Model embeds countries' readiness for AI adoption in TFP calculations but does not model initial user-unfriendliness or transition-period slowdowns.

---

### Limited AI Access Scenario
- Assumption:
  - EMs (excluding China) and LICs face constraints in access to AI-specific technologies (advanced processors, computing clusters, comprehensive data repositories).
- Impacts after 10 years (high TFP growth baseline comparison):
  - EMs and LICs’ real GDP increase falls on average by around 1 percentage point relative to the baseline scenario.
  - Slight changes to output growth in AEs and China.
  - Output losses relative to baseline more pronounced in EM groups (EMA and EML) due to higher AI exposure and higher AI preparedness that cannot be utilized under access constraints.

---

### Enhanced AI Preparedness Scenario
- Assumption:
  - For countries in EM and LIC groups, preparedness index is lifted to the best prepared country in their respective peer group.
  - No AI access restrictions (as in baseline).
  - Specific assumed changes in preparedness indices:
    - EMA: increase to 0.63 from 0.50.
    - EML: increase to 0.59 from 0.50.
    - LIC and ROW: increase to 0.48 from 0.38.
- Results:
  - Significant cross-region inequality persists but is partially offset relative to the baseline.
  - Global GDP level marginally benefits.
  - Additional output gains are more pronounced in the AI-intensive sector in EM and LIC groups, reflecting higher sectoral AI exposure.

---

### Mind the AI gap: The Way Forward for EMs and LICs
- Policy priorities by country preparedness level:
  - For economies with strong foundational AI preparedness (most AEs and some EMs):
    - Strengthen digital innovation capacity.
    - Adapt legal and ethical frameworks to govern and foster AI advances.
  - Where foundational preparedness is weak (some EMs and most LICs):
    - Prioritize investment in digital infrastructure and human capital to reap early AI gains and build toward second-generation preparedness.
- Public investment focus:
  - Prioritize areas with positive externalities and low private-sector investment incentives:
    - Fundamental research.
    - Necessary infrastructure (particularly in EMDEs).
    - Public sector applications with high social returns (education, health care, government administration).
- Technology access and sharing:
  - Enhancing access to hardware, software, and relevant data can reduce the gap between frontrunners and underperformers, though geopolitical and national security considerations may constrain technology sharing.
- Potential for technology leapfrogging:
  - Example cited: an advanced LLM developed by DeepSeek reportedly requires less computing capacity and powerful chips, suggesting breakthroughs can reduce resource barriers.
  - Open-source code and relevant data access could help less-developed EMs and LICs narrow the AI gap and apply AI to local challenges.
- Illustrative case:
  - Kenya’s M-Pesa as an example of leapfrogging: mobile banking and payments broadened financial inclusion, improved tech infrastructure (mobile network coverage, internet access, data services, e-commerce), and catalyzed fintech startups and a vibrant digital economy.

*Source: IMF Working Paper — "The Global Impact of AI: Mind the Gap", Section 4 Baseline Results and Section 5 A Challenging Outlook for EMs and LICs.*

### 6. Conclusions

### 6. Conclusions

### Macro growth impacts and distributional outcomes
- AI-induced productivity gains could lift global GDP by up to 4 percent in a high TFP growth scenario over the next decade.
- AEs benefit disproportionately, reflecting stronger AI preparedness and exposure, and greater access to high-tech infrastructure and data.
- The estimated growth impact in AEs could be more than double than in LICs.
- Among AEs, the US emerges as a clear frontrunner; other AEs and China also see notable boosts to output.
- Improved access by EMs and LICs to AI data and technologies can generate positive spillbacks to AEs through higher global growth.

### Role of preparedness, institutions, and infrastructure
- An enhanced AI preparedness scenario highlights the critical role of stronger institutions and digital infrastructures in reaping AI growth benefits.
- Policy action can help reduce the cross-country AI gap.
- Improved access by EMs and LICs to AI data and technologies is characterized as a win-win with global spillovers.

### Real exchange rates and sectoral patterns
- AI’s impacts could differ from textbook productivity shocks because AI will affect productivity in both non-tradable and tradable sectors.
- Traditional Balassa-Samuelson theory posits productivity gains occur disproportionately in tradable sectors, leading to real exchange rate appreciation.
- The analysis shows significant AI-driven productivity increases in AE non-tradable sectors—such as education and healthcare—could exert downward pressure on non-tradable prices, leading to moderate currency depreciation relative to EMs and LICs.
- This real exchange rate movement can bolster AEs’ external competitiveness beyond sectors directly benefited by AI and generate global spillovers.

### Caveats and avenues for future research
- Results should be read as suggestive scenarios rather than definitive estimates.
- Important avenues for future research include:
  - Refining the representation of labor markets—including unemployment dynamics and skill mismatches—to capture more nuanced impacts of AI-driven structural change.
  - Deepening analysis of within-country distributional consequences, such as the channels through which different income groups are affected.
  - Exploring scenarios in which varying policy settings, such as infrastructure investments and fiscal policy, alter the pace and breadth of AI adoption.

*IMF Working Paper: The Global Impact of AI: Mind the Gap, Working Paper No. WP/2025/076*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2025/english/wpiea2025076-print-pdf.pdf_