Figure 10 presents results from a model calibrated to the United Kingdom. The model requires data along three main dimensions: income distribution, occupational characteristics, and macroeconomic aggregates. Before describing the data sources, two foundational concepts need to be defined.
AI Occupational Exposure (AIOE) This is the primary measure of how much an occupation’s functions overlap with current AI capabilities, as developed by Felten, Raj, and Seamans (2021). It is constructed by measuring the correspondence between 10 AI application domains (e.g., image recognition, language processing, predictive analytics) and 52 human ability requirements drawn from ONET — a US government database of occupational characteristics. The overlap is weighted by the degree of importance and complexity each skill plays in a given occupation. The resulting score is normalized between 0 and 1 and is interpreted as a relative* measure of the likelihood that AI will be integrated into the core functions of a job. Critically, the AIOE score is agnostic about whether AI will replace or enhance the worker — it only captures how deeply AI can penetrate that occupation.
AI Complementarity Developed by Pizzinelli and others (2023), this index adjusts the AIOE measure to distinguish between AI that would likely displace workers and AI that would likely complement them. The intuition is that even in occupations where AI could technically perform many tasks (i.e., high exposure), social, ethical, and physical constraints may require human oversight. For instance, a judge is highly exposed to AI (legal research, text analysis), but society is unlikely to delegate judicial rulings to unsupervised AI — so the judge has high shielding from displacement. The complementarity index draws on two additional ONET dimensions: (1) work contexts (social and physical aspects of the job — e.g., consequences of error, contact with others, decision-making responsibility) and (2) skills (the complexity and level of human expertise required). Occupations with high AI exposure and* high shielding have high complementarity: AI augments rather than replaces the worker. A technical expression of this adjustment is: C-AIOE = AIOE × (1 − θ − θ_MIN), where θ is the complementarity score; a higher C-AIOE value implies a higher risk of labor substitution.
Using these two dimensions, occupations are grouped into three categories: - High Exposure, High Complementarity (HEHC): Cognitive jobs with significant human responsibility — surgeons, lawyers, managers. AI boosts productivity but human oversight remains necessary. - High Exposure, Low Complementarity (HELC): Jobs where AI can autonomously perform core tasks — clerical workers, telemarketers. Higher risk of displacement. - Low Exposure (LE): Jobs with minimal AI applicability — manual and physical trades, performing arts.
| Data Source | Role in the Model |
|---|---|
| UK Wealth and Assets Survey (WAS) — Office for National Statistics | Core calibration dataset. Provides the actual UK income distribution, decomposed by source: (1) labor/wage income, (2) capital income (rents and investment income), and (3) transfers (government benefits, pensions). This allows the model to replicate the observed relationship between income level and income composition (Figure 9, Panel 1 of the paper). |
| UK Labour Force Survey (LFS) | Used to assign occupational codes (ISCO-08) to individual workers, and to compute the share of each worker’s hours in high-AI-exposure occupations. This enables Figure 9, Panel 2 — which shows how AI exposure and complementarity vary across the income distribution in the UK. |
| Felten, Raj, and Seamans (2021) AIOE Index | Occupation-level AI exposure scores, based on ONET. Applied to UK workers through their ISCO occupation codes via a crosswalk. Defines the displacement* channel. |
| Pizzinelli and others (2023) Complementarity Index | Occupation-level complementarity scores, also based on ONET (published as IMF Working Paper 2023/216). Defines the complementarity* channel and adjusts which workers benefit vs. lose from AI. |
| UK Historical National Accounts / Macro Data | Provides the calibration target for the magnitude of AI-induced labor share decline — the 5.5 percentage point fall in the UK labor share observed between 1980 and 2014 (the largest historical episode of automation-driven labor share decline). This anchors the displacement effect in all three scenarios. |
The model is developed in Rockall, Pizzinelli, and Tavares (forthcoming) and combines frameworks from Drozd, Taschereau-Dumouchel, and Tavares (2022) and Moll, Rachel, and Restrepo (2022). It is a heterogeneous-agent, task-based model in continuous time, designed to trace AI adoption’s distributional consequences across both labor and capital income.
Production side: The economy produces a final consumption good using a Cobb-Douglas aggregation over a continuum of tasks. Each task can be performed by either human labor or AI capital. When AI becomes productive enough to perform a task, it is assumed to do so more efficiently than labor — making displacement productivity-enhancing from an aggregate standpoint. The production function for a worker of skill type z takes the form:
Y ∝ K^{α_z} × (ψ_z × L_z)^{1−α_z} aggregated over all task types z with importance weight η_z
where K is the aggregate capital stock, L_z is labor of type z, ψ_z is labor productivity, α_z is the capital share in performing tasks of type z, and η_z is the weight of those tasks in total value added.
Household side: Agents (households/workers) are heterogeneous along two dimensions: (1) their labor skill level (which determines their wage in each sector) and (2) their asset holdings (which determine their capital income). Agents supply labor inelastically across sectors, and those who invest in capital markets earn a higher rate of return than those who hold bonds (risk-free rate). Each agent maximizes utility from consumption over their lifetime subject to a budget constraint and a natural borrowing limit. Agents also face idiosyncratic dissipation shocks — sudden, random events that destroy their existing capital or wealth — which introduce realistic wealth dispersion.
This heterogeneity across agents allows the model to simultaneously replicate the income and wealth distributions observed in UK data, which is essential for capturing how AI affects inequality across the full distribution.
The model identifies three distinct mechanisms through which AI adoption affects the economy:
The Displacement Channel — captured by changes in α_z (the capital share in task production). As AI becomes capable of performing previously human tasks, capital substitutes for labor in those tasks. This reduces the share of income accruing to labor and shifts it toward capital owners. Importantly, this channel is productivity-enhancing because AI capital is assumed to perform displaced tasks more efficiently than labor.
The Complementarity Channel — captured by changes in η_z (the relative importance of tasks by skill type). AI adoption increases the relative value of tasks performed in high-complementarity occupations, because demand and value-added shift toward those jobs where human oversight remains essential and AI enhances rather than replaces the worker. This reallocates labor demand (and thus wages) from low-complementarity to high-complementarity workers, without changing the overall labor share in the economy.
The Productivity Channel — captured by changes in ψ_z (labor productivity). Beyond displacing tasks and reallocating labor demand, AI can broadly augment the productivity of workers in high-complementarity occupations. This generates economy-wide output growth and raises labor demand across sectors, creating a positive general-equilibrium effect that can offset some displacement losses.
All scenarios are calibrated to the same 5.5 percentage point decline in the labor share (matching the UK’s 1980–2014 experience). They differ only in the strength of the complementarity and productivity channels:
Scenario 1 — Low Complementarity: AI modestly increases demand for high-complementarity jobs. The displacement effect dominates at the top of the income distribution (where exposure is highest), reducing labor income for higher earners. Result: labor income inequality falls, but capital income inequality rises. The share of workers negatively affected at the top reaches nearly 15%.
Scenario 2 — High Complementarity: AI strongly boosts demand for high-complementarity occupations, reallocating value-added and wages toward high-income workers. The complementarity gains more than offset displacement for top earners, reducing the share negatively affected at the top to under 5%. Low-income workers see labor income decline by around 2% as sectoral reallocation shifts resources away from low-complementarity occupations. Result: labor income inequality increases, and capital income inequality increases further.
Scenario 3 — High Complementarity + High Productivity: The productivity channel is also activated, calibrated to generate approximately a 1.5 percentage point increase in workers’ average annual productivity growth in the first decade — the low end of firm-level estimates of AI’s productivity effect. This drives output up by roughly 16% between steady states and raises TFP by nearly 4%. Labor income rises for all workers, ranging from +2% at the bottom to nearly +14% at the top. Result: total income increases broadly, but inequality still rises because gains are disproportionately concentrated at the top.
Capital income channel (present in all scenarios): In all three scenarios, AI adoption raises the return on capital (interest rates rise approximately 0.4 percentage points), increasing the value of asset holdings. Since wealthy households hold the most assets, capital income and wealth inequality always increase regardless of what happens to labor income.
Two additional hypothetical scenarios in Annex 4 isolate the importance of how exposure and complementarity are distributed across the income distribution. They show that when displacement is uniform but complementarity is concentrated at the top, inequality rises; when exposure is concentrated at the top and complementarity is absent, inequality may fall as top earners face greater displacement than those at the bottom.
The direction of inequality change is not predetermined. Whether AI increases or decreases labor income inequality depends on the relative strength of displacement versus complementarity. The paper’s empirical data show that in the UK, both exposure and complementarity rise with income — a configuration that, under plausible scenarios, leads to rising inequality. This may not hold everywhere.
Capital income inequality unambiguously increases. In all scenarios, AI raises the return on capital. Because asset ownership is concentrated among higher-income households, this channel consistently amplifies inequality regardless of what happens to labor markets. This makes the distribution of wealth — and the policy framework governing AI ownership — a critical determinant of aggregate distributional outcomes.
The productivity channel is the key to broad-based gains. Only when AI generates sufficient productivity growth do all workers benefit, even those in low-complementarity jobs. Without it, AI adoption is redistributive rather than broadly beneficial. The magnitude and speed of this productivity effect remain empirically uncertain.
Complementarity is positively correlated with income in the UK, with the share of high-complementarity employment rising across most of the income distribution (peaking near the 75th percentile). This structural fact is the main reason the high-complementarity scenario raises labor income inequality — AI most benefits those who already earn more.
Fiscal and redistributive policies are excluded from the model — the simulations represent pre-redistribution outcomes. Policy choices around AI property rights, taxation, and social protection can substantially reshape distributional results.
The following data sources would be needed to calibrate and run an equivalent model for a non-UK country:
| Data Requirement | Description | Challenge |
|---|---|---|
| Household income survey with income source decomposition | Microdata distinguishing between labor income, capital income (especially investment income and rents), and transfers, across the full income distribution. Equivalent to UK’s Wealth and Assets Survey. | Most developing country household surveys report total income but not its source composition — particularly capital income, which is severely under-reported or absent. |
| Labour force survey with ISCO occupational codes | Panel or cross-sectional survey linking workers to detailed occupational codes (ideally 3- or 4-digit ISCO-08), with wage/earnings data. Used to map AI exposure and complementarity scores to workers. | Many emerging market surveys use national occupation classifications that require crosswalks to ISCO. Panel structure is rare outside advanced economies. |
| AI Occupational Exposure (AIOE) scores | The Felten, Raj, and Seamans (2021) index, applicable internationally via an ISCO crosswalk. Publicly available. | Scores are derived from US O*NET data and assume occupational task content is homogeneous across countries — a simplification the paper explicitly acknowledges. |
| Complementarity scores by occupation | The Pizzinelli and others (2023) index, published in IMF Working Paper 2023/216. Also based on O*NET, broadly applicable internationally. | Same cross-country homogeneity caveat as above. Local task variations within occupations could produce meaningfully different complementarity scores. |
| Historical labor share data | National accounts data showing the evolution of the labor share over time, used to calibrate the magnitude of the displacement shock. | Reliable long-run labor share series exist for most countries via the Penn World Tables or ILO, but quality varies. The appropriate historical analog for calibration in countries with different automation histories may differ from the UK’s 1980–2014 episode. |
| Wealth/asset distribution data | Data on financial and non-financial asset holdings (stocks, bonds, real estate) by income or wealth percentile, to calibrate the capital income channel. | This is the most demanding requirement and most commonly missing in emerging market and developing economies. Wealth surveys comparable to the UK WAS are rare; estimates can sometimes be constructed from administrative tax records or national financial accounts combined with distributional assumptions, but with much greater uncertainty. |
In summary, the model is feasible to adapt for advanced economies with rich household surveys (e.g., those participating in the ECB Household Finance and Consumption Survey, or countries with income and wealth panel data). For emerging markets, the capital income and wealth distribution data represent the binding constraint. Simplified versions of the model — using only labor income channels with stylized capital income distributions — may offer a practical intermediate approach.
Source: IMF Staff Discussion Note, “Gen-AI: Artificial Intelligence and the Future of Work” (2024). Model details in Annex 4; empirical inputs in Section II and Figure 9.