Vahid Ahmadi · PolicyEngine · September 2026
This report estimates the macroeconomic effects of AI-driven automation on the UK over 2026–2030 using OG-UK, the UK calibration of the OG-Core overlapping-generations model. Existing projections differ widely. Korinek et al. (2026) find US growth of 5.4% a year by 2030 in their “substantial” scenario, while Moll and Imas (2026) argue that diffusion lags, physical and relational work, and shifts in spending make double-digit growth unlikely within 10–15 years. The OBR’s March 2026 Economic and fiscal outlook projects GDP growth of 1.6% a year over 2027–30 and, in Box 2.2, sets out a displacement scenario with a lower labour share, 5.5% unemployment and unchanged GDP.
We model AI as a gradual rise in the capital share of output over 2026–2030, lowering the labour share by 3.9 percentage points, with or without a 3.1% rise in total factor productivity. Relative to the no-AI baseline, GDP is 4.9% higher in 2030 with productivity gains and 1.4% higher with automation alone, driven by investment rather than consumption. The technology paths are imposed, so the results show the general-equilibrium consequences of given scenarios, not their likelihood.
GDP, investment, consumption, wages and the public finances under each AI scenario, from OG-UK, 2026–2030.
The GDP gain is investment-led, not consumption-led.
OG-UK output. Government spending is a fixed share of GDP (caveat 4: spending held at baseline levels). Definitions and caveats: Scenario design.
| Variable | Automation only2030 · path 2026–2030 | Automation + productivity2030 · path 2026–2030 |
|---|---|---|
| Output and demand | ||
| GDP | +1.4% | +4.9% |
| Consumption | −1.3% | +0.3% |
| Investment | +6.9% | +13.1% |
| Factors | ||
| Capital stock | +5.0% | +6.1% |
| Labour | −0.8% | −0.8% |
| Average wage | −3.9% | −0.6% |
| Public finances | ||
| Tax revenue | +0.1% | +4.0% |
| Public debt | +2.2% | +1.4% |
| Shares and growth | ||
| Labour share | −3.9pp | −3.9pp |
| GDP growth | +0.2pp | +1.3pp |
Index, 2026 no-AI baseline = 100
The AI scenarios start away from 100 because of anticipation (caveat 8).
Real GDP (£bn, 2023 prices); scenarios from 2026
Sources: OBR March 2026 EFO (economy tables, public finances databank) and ONS. Scenario paths apply the model’s changes to the OBR’s 2026 level (caveat 10).
Each country against its own no-AI path; timing and model differences: caveat 9
Automation + productivity lies between US Modest and US Substantial, and Automation only below US Modest; GDP growth in 2030 rises by +3.4pp in the comparable US case (Substantial).
Hollow dots: Korinek et al. (2026), Table 3, US, 2030; US Extreme is off scale. The two UK AI scenarios share one labour-share path. The labour-share axis does not start at zero.
How OG-UK, the UK calibration of the OG-Core overlapping-generations model, represents households, firms and government.
The effects of automation run through saving, investment, incomes and the public finances together, and OG-UK determines all of them in one general equilibrium.
OG-UK has no unemployment or job search, and the runs use one sector. The consequences are set out in the caveats under Scenario design.
The central insight of an overlapping-generations model is that the economy is not populated by a single “representative agent.” At any moment in time, many generations coexist: young people just entering the labour force, middle-aged workers at peak earnings, and retirees living off savings and pensions. Each generation has a different remaining lifespan, a different stock of accumulated wealth, and therefore a different response to any given policy change.
OG-Core models S economically active age cohorts (typically 80, representing ages 21 to 100). Each period a new cohort enters, households face age-specific mortality, and the oldest cohort dies with certainty. The population distribution by age evolves over time with fertility, mortality, and immigration, and eventually reaches a stationary distribution in the long run.
Within each age cohort, households split into J ability types (seven in OG-UK), each with a distinct permanent labour productivity. High-ability types earn more per hour of work and typically save more. In OG-UK every type has the same discount factor. This heterogeneity captures how a tax reform affects the income distribution, not just average outcomes.
Every period, each household makes three decisions: how much to consume, how many hours to work, and how much to save for the future. The household chooses these to maximise lifetime utility, subject to a budget constraint, with perfect foresight about future prices and policy.
Consumption utility follows a CRRA (constant relative risk aversion) function, where the risk aversion parameter governs how strongly households prefer smooth consumption over time. Before entering this function, a Stone-Geary Cobb-Douglas form aggregates the individual consumption goods, accounting for minimum subsistence levels.
Labour disutility uses an elliptical function rather than a standard power function. This ensures that the marginal cost of working goes to zero when hours are zero (so households always choose to work at least a little) and goes to infinity at the upper bound (so they never hit the ceiling). This avoids computationally expensive occasionally-binding constraint methods.
Bequest utility captures the “warm glow” from leaving wealth behind upon death. The age-specific mortality probability weights it, so it matters most for older households. The strength of the bequest motive can vary by ability type.
The household's budget constraint says that each period, income from savings returns, labour earnings, bequests received, government transfers, and pensions must cover consumption (including consumption taxes), income and wealth taxes, and savings carried into the next period.
Two Euler equations characterise the optimal decisions; the panel shows them in simplified form. The labour first-order condition says: work until the after-tax wage equals the marginal disutility. The savings Euler equation says: give up one unit of consumption today only if the discounted, after-tax return makes tomorrow’s consumption sufficiently valuable.
The production side of the economy consists of M industries, each populated by perfectly competitive firms. Each firm combines three inputs: private capital, public capital (government-provided infrastructure), and labour. A constant elasticity of substitution (CES) technology combines them. In OG-UK the public-capital share is zero and there is no public investment, so public capital is inactive in these runs.
The elasticity parameter governs how easily firms substitute between capital and labour. When it equals one, the function reduces to the Cobb-Douglas form. Total factor productivity varies across industries and over time, so the model can capture sector-specific technological change.
Profit maximisation yields the standard first-order conditions: the wage equals the marginal product of labour, and the rental rate of capital equals the after-tax marginal product of capital, accounting for depreciation, tax deductions, and investment tax credits.
When it is active, public capital generates economic rents: firms cannot deduct the cost of using government infrastructure, so its returns flow to private capital owners through an augmented rate of return. With a zero public-capital share in OG-UK, this channel is switched off.
The government in OG-Core collects revenue from five tax instruments: individual income taxes, consumption taxes, wealth taxes, corporate income taxes, and bequest taxes. It spends on public goods, infrastructure, pensions, lump-sum transfers, and universal basic income. Issuing debt finances the difference.
The income tax carries the most structure. OG-Core supports several functional forms for effective and marginal tax rates, fitted to detailed tax-benefit calculations. For OG-UK, we use the Gouveia-Strauss specification, which estimates the effective tax rate as a smooth, monotonically increasing function of income and derives the marginal tax rate analytically. The estimation runs on PolicyEngine-UK output drawn from the Enhanced FRS — the Family Resources Survey enhanced with HMRC’s Survey of Personal Incomes, the Living Costs and Food Survey, and the Wealth and Assets Survey.
The wealth tax uses a progressive three-parameter function that produces anything from zero taxation to smoothly increasing marginal rates. Consumption taxes take linear rates that vary by good, covering VAT and excise. Corporate income taxes take flat rates by industry.
Since debt cannot grow without bound, a fiscal closure rule activates from a set period (tG1) to bring the debt-to-GDP ratio gradually towards a target. In OG-Core 0.17.0 only government consumption G adjusts between tG1 and the end of the adjustment (tG2); transfers stay a fixed share of GDP. The setting used in these runs is under Scenario design.
Market clearing conditions connect all the pieces above — household decisions, firm production, and government policy. For the economy to be in equilibrium, supply must equal demand in every market simultaneously.
In the labour market, the total efficiency-weighted labour supply from all households across all ages and ability types must equal total firm demand across all industries. In the capital market, the total stock of savings (both domestic and foreign) must equal the capital firms demand plus government debt. The model runs as an open economy: parameters control foreign capital and foreign debt holdings, setting the degree of international capital mobility.
For each goods market, industry output must equal consumption demand. The final industry clears residually, absorbing investment, government purchases, and any remaining demand. The debt market splits government debt between domestic and foreign holders.
An equilibrium is a set of prices (the interest rate, wage, and goods prices) such that all households optimise, all firms optimise, the government budget constraint holds, and all four markets clear. Finding one is a large-scale nonlinear fixed-point problem.
Finding equilibrium is a two-stage computational problem. The model first solves for the long-run steady state, then computes the transition path from today's economy to that steady state.
The model first finds the long-run equilibrium where all variables stay constant (after removing trend growth). It uses a nested fixed-point algorithm. The outer loop guesses a vector of aggregate prices and quantities. The inner loop takes these as given and solves the Euler equations for every combination of ability type and age. Because each type is independent, the inner loop solves them in parallel. After solving all households, the algorithm aggregates their decisions, computes firm demands, checks the government budget, and derives new guesses. It iterates until convergence.
Starting from today’s economy, the model then solves for how the economy evolves towards the steady state over T periods. The Time Path Iteration (TPI) method guesses entire time paths for all prices and quantities, solves every cohort’s lifetime decisions given those paths (with perfect foresight about future prices), checks whether the implied paths match the guesses, and iterates. This step produces the year-by-year projections.
OG-Core provides the theoretical framework. Turning it into a model of the UK economy means calibrating every parameter to UK data. The OG-UK calibration layer handles that step.
Macroeconomic parameters come from the ONS national accounts and the OBR’s Economic and Fiscal Outlook: the debt-to-GDP ratio, the revenue-to-GDP ratio, government spending shares, and growth rates. The state pension age matches current UK rules. Demographics come from the UN World Population Prospects for the United Kingdom, read through the EAPD-DRB population-data mirror.
Tax functions come from PolicyEngine UK. Rather than hard-coding the UK income tax schedule, the model fits smooth Gouveia-Strauss functions that capture the effective relationship between income and taxes across the entire population. This approach folds in the interaction of income tax, National Insurance, and the personal allowance taper.
The model solves in abstract units. OG-UK can scale them to pounds by anchoring GDP to the ONS figure; this report presents indices and ratios instead.
Together with a static microsimulation engine, OG-UK supports two complementary views of any UK tax or benefit reform. Static analysis provides the detailed distributional picture: which households gain, which lose, and by how much. OG-UK provides the macroeconomic picture: what happens to GDP, investment, government revenue, interest rates, and wages as the economy adjusts over time.
The two perspectives can tell different stories. A reform that raises revenue in a static analysis may raise less in practice, because higher tax rates reduce labour supply and investment and shrink the tax base. The reverse also holds: a reform that costs revenue upfront might pay for itself over time through extra economic activity. OG-UK captures these dynamics.
The transition path carries the bulk of the answer. Rather than comparing two long-run steady states, it shows the year-by-year adjustment: how GDP dips in the first few years, how interest rates respond, when revenue stabilises. These time paths map onto OBR forecast horizons.
OG-UK is a structural macro model and inherits the simplifications of that class. The caveats for reading its outputs:
None of these are model bugs — they are choices that make the macro accounting tractable. They apply to every result under Economic effects.
How AI enters OG-UK: the two technology changes, the three scenarios, the fiscal convention and the caveats.
OG-UK has no AI parameter; AI enters through two changes to the production technology, and everything else is as in the no-AI baseline.
Of the seven AI channels in the table, the runs carry three fully, one partly and three not at all.
A channel marked “not modelled” has no counterpart in the runs, so no result here speaks to it. The channel-by-channel comparison with Korinek et al., Moll and Imas and the OBR is under Model comparison.
| Channel | What it means | How OG-UK represents it |
|---|---|---|
| Modelled (3) | ||
| Automationmodelled | Tasks move from labour to capital | γ, the capital weight in production, rises over the window and lowers the labour share one for one (step 1). |
| Productivitymodelled | AI makes production more efficient | Z, total factor productivity, rises from 1 to a value solved for each scenario’s target (steps 4–5). |
| Capital accumulationmodelled | Investment responds to the higher return | Endogenous: households save and the capital stock builds; not imposed. |
| Partly or not modelled (4) | ||
| Adoption and diffusionpartly | How fast firms take AI up | Not separate — folded into the ramps of γ and Z. |
| Unemployment and searchnot modelled | Displaced workers take time to find new jobs | Absent: no unemployment or search frictions. |
| New tasks for peoplenot modelled | AI creates work as well as replacing it | Absent. |
| Which goods get cheapernot modelled | Spending shifts toward what AI cannot do | Absent: the runs have one sector. |
ε = 1, γ = 0.35 and Z = 1 in every year. γ = 0.35 (a labour share of 65%) is OG-Core’s default, not a UK estimate.
No AI shock: OG-UK’s existing calibration and the current tax system. Labour-augmenting productivity grows at 1.0% a year, the OBR’s medium-term rate; the population follows the UN World Population Prospects projection for the UK.
Every other path is measured against this one. It is checked against the OBR’s forecast under Model comparison.
A shift of income from labour to capital with no output gain at fixed inputs.
Z offsets the output effect of the higher γ at baseline inputs: a pure factor-bias shift with no output gain at fixed K and L. Output per worker still rises in equilibrium, because capital deepens.
The automation-only case follows the OBR’s Box 2.2 in the March 2026 Economic and fiscal outlook: a lower labour share with no output gain at fixed inputs. In the OBR’s words, “higher trend productivity fully offsets lower employment to leave the level of GDP unchanged… As labour income faces a higher effective tax rate, this reduces the tax-richness of economic activity.”
The OBR scenario also raises equilibrium unemployment to 5.5%; OG-UK has no unemployment, so that part has no counterpart (caveat 3).
Comparison with the OBR’s fiscal estimate. OG-UK carries part of the tax-richness channel: labour and capital income pass through one fitted income-tax function, plus a 27% corporation tax. In 2030 this scenario lowers tax/GDP by 0.55pp. The OBR (EFO paragraph 6.18) puts the effect of its scenario on receipts at −£6bn a year (about 0.2% of GDP), with borrowing +£9bn a year and debt +1.3% of GDP by 2030-31. The model’s debt result depends on the fiscal convention (caveat 4).
The same automation, plus a rise in total factor productivity of 3.1% at fixed inputs. This is the main scenario.
The same γ path as Automation only. Z is solved jointly with γ so that the combined change raises measured TFP by 3.1% at baseline inputs.
The 3.9pp fall is applied in percentage points to OG-UK’s 65%; the proportional equivalent (−6.5%) would give 60.8%.
The size of the labour-share fall and the productivity gain are set to match the “substantial” case of Korinek, Jones, Sacher, Cotter and McCrory (2026), Anthropic Institute Working Paper 2026-02, Table 3: a US labour share that falls from 60% to 56.1% by 2030, and measured TFP 3.1% higher. Their unemployment (4.5% for cognitive workers, 4.6% for all workers) and wage split have no counterpart in OG-UK.
The four US scenarios of Korinek et al. (2026), Table 3: their own model’s results for the United States, not anything run here.
Table 3 publishes 2030 endpoints only, so the US appears on the charts as single hollow dots. Each US scenario is compared with the paper’s own US no-AI path, as each UK scenario is with the UK no-AI baseline; the two no-AI paths grow at different rates, so levels are not compared (caveat 9).
Raising γ alone increases output at baseline inputs, so Z must offset part or all of that gain.
At the baseline’s capital-labour ratio, the rise in γ alone raises output by 6.06%. Z removes that gain in Automation only and all but +3.1% of it in Automation + productivity, so both solved values are below 1. The two AI scenarios differ only in Z.
The level of Z has no economic meaning (caveat 11).
Seven limits apply to every result on Economic effects.
| 2030, against the no-AI baseline | GDP | Tax/GDP | Debt/GDP |
|---|---|---|---|
| Automation + productivity Spending a fixed share of GDP (main runs) | +4.9% | −0.35pp | −3.00pp |
| Automation + productivity Spending held at baseline levels | +5.2% | −0.41pp | −5.72pp |
| Automation only Spending a fixed share of GDP (main runs) | +1.4% | −0.55pp | +0.68pp |
| Automation only Spending held at baseline levels | +1.5% | −0.58pp | −0.73pp |
Four points on how the charts are built.
Three problems in earlier versions are fixed in these results (PR #11): the AI scenarios set their own spending share (#2), the window stopped at 2029 because of an early switch to debt targeting (#3), and productivity growth was mis-sourced (#5; see Model comparison).
What OG-UK, Korinek et al. (2026), Moll and Imas (2026) and the OBR can each represent, and OG-UK’s baseline against the OBR forecast.
The four sources differ in what they are: two structural models, one argued model and one forecast scenario.
All four represent automation and a falling labour share; three model capital.
Diffusion, new human tasks and unemployment are in Korinek et al. only; OG-UK has none of them.
Korinek et al. model diffusion (the adopted share dt, from 0.10 to 0.20, 0.40 or 0.6 by 2030), task reinstatement (new human tasks, ρ of 0.50, 0.25 or 0), unemployment and search (matching, a search discount and posting speed) and an ideas stock. Moll and Imas share only the ideas channel, in their extension on AI in research. The OBR states an unemployment outcome (5.5%) without modelling search.
In OG-UK, adoption is folded into the ramps of γ and Z; reinstatement and search have no counterpart (caveat 3).
OG-UK is the only source with heterogeneous households; the OBR alone shares its cohorts and tax system.
Demand composition — spending shifting towards what stays scarce as automated goods get cheaper — is absent from every model as run. It is central to Moll and Imas’s argument, but argued rather than modelled; Korinek et al. have one good; OG-UK’s multi-sector build is not used (caveat 2).
In OG-UK, capability and adoption are folded into the ramps of γ and Z; reinstatement and search have no counterpart.
The labour-share fall in Korinek et al. comes from the affected mass, diffusion and automation share jointly (m·d·ψ), so adoption enters OG-UK’s γ as well as Z. OG-UK has no tasks to count separately.
The models also measure substitution between different things. Korinek et al.’s elasticity of 0.5 is across tasks; Moll and Imas cite Charles Jones’s elasticity of 0.2 across goods; OG-UK’s ε is between capital and labour, set to 1 (Cobb-Douglas) so that the labour share is exactly 1−γ and the 3.9pp target is met by construction.
Box 2.2 states its outcomes; it does not derive them from a model.
The OBR has an overlapping-generations model of the same class as OG-UK (Working Paper No. 22, A new UK overlapping generations model, April 2025), but Box 2.2 is an assumption layered on the central forecast. Its AI-productivity scenarios are separate, in Briefing Paper No. 9, Forecasting productivity, Annex B. The OBR case can therefore be imposed in OG-UK but not tested against the OBR’s own model.
OG-UK’s no-AI baseline grows 1.68% a year on average over 2027–30, against the OBR’s 1.60%: a gap of +0.08pp.
Every scenario is measured against the no-AI baseline, so the baseline is checked against the OBR’s March 2026 Economic and fiscal outlook (paragraph 1.9). The model’s growth rates are computed from the committed results; the component values below come from the repository’s obr.py comparison.
Productivity growth now matches the OBR; the remaining gap is population growth against labour supply.
Before the re-run, g_y_annual was 1.1%: the OBR’s potential output growth entered as labour-augmenting productivity, which double-counted labour supply already carried by gn. It is now 1.0%, the OBR’s productivity figure (issue #5, PR #11).
Only growth has been compared with the OBR year by year.
The OBR also publishes borrowing (5.2% of GDP in 2024-25 and 4.3% in 2025-26, falling to 1.6% in 2030-31; paragraph 1.3), debt and receipts. OG-UK produces all three, and its own ratios to GDP differ from the data (caveat 10); a year-by-year comparison has not been run.
OG-UK is weakest where Korinek et al. are strongest (adoption, reinstatement, unemployment) and strongest on households, cohorts and the tax system, where only the OBR also covers cohorts and tax. OG-UK adds the general-equilibrium response of capital, saving, prices and the tax base to an imposed technology path; PolicyEngine UK adds who bears it.
How to reproduce every result, and how to use OG-UK for your own analysis.
Everything on this dashboard is produced by the uk-growth-with-ai repository, with the results committed so the charts can be rebuilt without re-running the model.
The re-run results live on the branch fix/rerun-2-3-5 (PR #11) until it is merged. The package itself needs only NumPy, pandas and Dask; report, obr and dashboard work without the model installed.
The results use OG-Core 0.17.0 and OG-UK at commit d0e6ae5.
OG-UK is installed from that commit rather than from PSLmodels/OG-UK. It reads PolicyEngine UK microdata from Hugging Face, so a token with read access is needed once.
Stock OG-Core holds γ fixed over time; the patch lets it vary by year.
The patch changes ogcore/firm.py in the installed package only; the design and its checks are in docs/firm_gamma_tv.md. run refuses to start if the firm block is unpatched.
run solves the steady state and the transition path for the baseline and each AI scenario.
--baseline-spending repeats the AI scenarios with government spending held at the baseline’s levels, the robustness run for the fiscal convention.A full run takes several hours; add --only anthropic or --only obr to run one scenario.
check runs the validation checks on a results file before anything is published.
checks.py tests the fiscal rule, the common spending share, the Z targets, the labour share and solution quality. report prints the year-by-year gaps for one scenario, and obr the baseline comparison with the OBR (obr.py).
dashboard writes the two JSON files this site reads; they are generated, never edited by hand.
dashboard_data.py builds aiScenarios.json (the model paths, including the fixed-spending run when present) and ukAiPaths.json (the OBR and ONS series). The test suite runs dashboard --check, so stale or hand-edited files fail.
Reforms are PolicyEngine policies: a parameter from the UK tax-benefit rules, a new value and a start date.
Anything PolicyEngine UK can represent — rates, thresholds, allowance tapers, new benefits — passes through; to change the reform, change the parameter path and value.
solve_steady_state finds the long-run equilibrium under a policy; run it for the baseline and the reform and compare.
A steady-state pair takes a few minutes on a laptop.
The transition path gives the year-by-year adjustment from today to the steady state.
It solves every cohort’s lifetime under perfect foresight, so it is heavier; OG-UK uses Dask to spread the work across CPU cores.
Results come back as NumPy arrays by year, in pounds (the mapping is described under Model).
Pass multi_sector=True for the eight-sector build (not used in this report). The full API and theory are in the OG-Core documentation and the OG-UK documentation.