How the Futur model works
Futur is a game, but the economy underneath it is a real, published, tested model. This page shows what it does, where its numbers come from, how it was checked, and where it is weak.
It is a stylised model of the UK economy from 2025 to 2035. It is a thinking tool, not a forecast. It is more reliable at showing the difference AI makes (your Britain against the same world without AI) than at predicting any single number.
1. What the model does each year
- AI spreads. Adoption follows an S-curve. It speeds up when AI is cheap compared with the pay of the workers it could replace. The cost of AI halves every four years by default.
- AI adds to productivity. How much depends on the hidden world you are playing in (section 4).
- Gains tilt towards capital. As machines do more, owners of capital take a larger share of income. A tax on capital pulls that back.
- Three groups of workers feel it differently. Junior workers (about 24% of the workforce) are the most exposed, mid-career workers (about 56%) are mixed, and senior workers (about 21%) gain the most from working alongside AI. The shock to each group is split between lost jobs and lower pay according to how flexible that group is.
- The books must balance. Total pay is always forced to equal the labour share of the economy, so the model cannot invent money.
- Inequality moves. It rises with a slow background drift, with machines taking over more income, with a widening pay gap between senior and junior workers, and with youth unemployment. A basic income compresses it a little.
- The public finances update. Revenue and spending start at 40.5% of the economy. A basic income, a capital tax and AI profits change them. Debt grows with interest.
- Stability is scored. It combines inequality, unemployment, youth unemployment, the pay share, years of flat wages, and wealth concentration.
The crisis lines that end a game early are: debt above 130% of the economy, youth unemployment above 12%, the pay share below 55%, or stability below 0.55.
Starting point (2025): economy £2,750bn, pay share 59.5%, debt 95% of the economy, richest 10% hold 43% of wealth, inequality (Gini) 0.329.
2. Where the numbers come from
| Part of the model | Based on |
|---|---|
| Baseline growth, population and public finances | Office for Budget Responsibility (OBR) November 2025 forecast; ONS 2024 income inequality; ONS Labour Force Survey April 2026; OECD Economic Outlook 2025/2 |
| The three worker groups and how flexible each is | Stanford / ADP evidence on entry-level work; Brynjolfsson and Li (QJE 2025); Bloom and Prettner on the skill premium; Hamermesh (1993) on labour demand; Manning (2021) on elasticities |
| Productivity, capital share and adoption | Acemoglu (2024) on productivity; IMF Working Paper 2025/068 on capital share; the Acemoglu-Restrepo task framework for adoption; Turchin for the stability idea |
Checks made on figures people often quote
When this project was reviewed in October 2026, several widely repeated figures were checked against the primary source. What was found:
- The OBR's March 2026 outlook does include a 1.5% a year productivity upside case, linked to AI. Its downside case (0.5%) is not framed as AI-related, so the "£90bn gap" sometimes quoted is not an AI gap.
- The "10 to 30% of UK jobs are automatable" line is a pre-2020 consensus repeated in a Department for Education paper. It is not a recent official estimate.
- The Department for Science, Innovation and Technology (January 2026) reports about 70% of UK workers are in jobs with tasks AI could perform or enhance. That measures exposure, not job losses.
- No PwC figure for the UK economic boost from AI by 2035 could be found, so none is used.
A second review of a separate research library (October 2026) supported several of the settings, notably that young people's job-finding has fallen in AI-exposed work while overall unemployment has not yet risen, and that firms' own reports of AI productivity gains are much higher than measured gains so far.
3. How it was tested: the backtest
The model was run on the UK for 2015 to 2019, a period before today's AI, and compared with what actually happened.
| Measure | Average miss, first version (v4) | Average miss, current calibrated engine (v5) |
|---|---|---|
| Size of the economy | 3.9 percentage points | 0.5 percentage points |
| Inequality (Gini) | 0.029 | 0.009 |
| Unemployment | 0.85 percentage points | 0.5 percentage points |
Three fixes closed the gap: a small steady background rise in inequality (housing wealth and top incomes), faster recovery of unemployment after a shock (matching how quickly UK unemployment actually adjusted), and a correction to how labour enters the production function.
What the backtest does and does not show. It shows the model reproduces the UK's ordinary behaviour in a normal period. It cannot show that the AI effects are right, because there was no AI shock in 2015 to 2019 to test against. The AI settings rest on the published research above and on the range of views in section 4.
The game adds extra mechanisms on top of the tested engine: a hiring freeze for young workers, new tasks created by AI, capital moving abroad when capital tax is high, higher borrowing costs when debt is high, and a social-tension score. These are stylised and not calibrated to measured data, because no source gives the sizes. They were tuned so that no single policy always wins. The backtest does not cover them.
3b. The game on top of the model
The game adds a few things to the tested engine. They are stylised choices, not measured effects:
- Decision cards. Between rounds you may face a situation (for example "firms stop hiring juniors"). Each answer changes a real model setting for one or two rounds and states its cost up front. The sizes of the effects are illustrative and not calibrated to data.
- The forecast shown while you move the sliders runs the next two years once for each of the four possible worlds, using the middle of each world's range. It never uses the real hidden world.
- Goals and the debrief. Each world has a goal. At the end, the debrief replays this very world with other policies (doing nothing new, and the best steady policy) so the comparison is exact.
- The picture. The landscape exaggerates each worker group's change from 2025 threefold so that small changes are visible. The figures in text are exact.
- Background headlines (a market correction, trade barriers and so on) are labelled on screen as not part of the model.
One known weakness: in the model a faster rollout is almost never worse than a slower one, because scores are judged in 2035, after any transition cost has passed. This is recorded as an open calibration question.
4. The hidden worlds
You never know how much productivity AI will really deliver. Each game secretly picks one of four worlds:
| World | Extra productivity growth a year | Anchored on |
|---|---|---|
| Sceptic | 0.1 to 0.3% | Acemoglu-style low estimates; measured studies so far find close to no effect on pay or hours |
| Steady | 0.6 to 0.9% | Task-based mid estimates; in line with OECD estimates of 0.4 to 1.2% for the UK |
| Boom | 1.0 to 1.5% | Historical-analogy estimates; the OBR's 1.5% upside case |
| Surge | 1.5 to 2.0% | Upper estimates in the style of investment-bank scenarios |
The ranges are modelling choices that span the published debate. They are not predictions about which world is likely.
5. What the model leaves out
- Sectors (finance, law, health, creative work), regions and housing.
- Interest-rate policy and inflation.
- Geopolitics, trade, energy and computing limits, and climate. The background headlines in the game (a market correction, trade barriers, a power squeeze, deepfakes in an election) are labelled on screen as not part of the model and change no numbers.
- Concentration of the AI industry in a few companies.
- Migration, and the UK acting alone: the model has one economy and no other countries.
- Sudden shocks. It cannot represent a crisis like 2008 or 2020.
The unrest score is a rough indicator, not a measured quantity.
6. See for yourself
Every game can be replayed exactly: the same hidden world and the same choices give the same result. The model has an automated test suite, including a check that the original calibrated results are reproduced exactly, and every change to the engine is documented with what it affects.