What Next?

Machine intelligence is getting cheaper at a remarkable rate. When one thing becomes abundant, value moves to whatever is still scarce beside it. Across the twelve chapters the same four things keep running short.

• Proof. When anything can be generated, knowing what is real becomes valuable. Trust in news is at 37%, the lowest the Reuters Institute has recorded.
• People. From this year the UK is projected to record more deaths than births. Carers, builders, physics teachers and grid engineers are short now and will be shorter.
• Physical capacity. Software can be copied; substations cannot. Grids, water, homes and planning permission will set the pace of the 2030s far more than code.
• Coherence. Most failures in this report are failures to join things up: tools nobody has absorbed, targets with no delivery plan.

The caveat runs through every chapter: cheap is not the same as useful. The gap between what the machines can do and what organisations actually do with them is wide.

1. Will AI take our jobs? No. Not enough to break the jobs market. The bigger problem is too few workers and a broken first rung.
2. Will AI eat the power grid? Yes. Data centres head for more than 1,000 TWh a year.
3. Will we need a universal basic income? No. We will pay for a bigger benefits bill instead.
4. What happens to the NHS? It survives, but misses its own promise, and not by a little.
5. Will renewables win? Only just: about half the world's electricity by 2035.
6. Will we stop buying so much stuff? We already have, per head.
7. Will Britain make more of what it buys? No. The goods deficit stays wide.
8. Will we build the homes we need? No.
9. Will British farming survive? Yes. Fewer farms, same share of food.
10. Is university still worth it? For most, yes. Not for all.
11. Will we get our news from AI? Yes, nearly a third of us.
12. Is Britain shrinking? Yes, without migration.

Read together, most answers say that something people fear or hope for does not arrive on time. A few say a slow trend keeps going. Only three say something moves fast: the machines' appetite for electricity, the share of it from renewables, and the number of people who get their news from a machine. That is the argument in twelve lines. Intelligence is getting cheap fast. Almost everything it needs in order to matter is scarce and slow.

Two things decide which 2035 we get: whether people and institutions learn to use cheap intelligence widely, and whether the physical and public things get built and paid for. That gives four worlds, which we follow through one invented person, Sam, aged 27 in Stockport. In The Long Grind, nothing collapses and everything takes slightly longer than it should. In The Crossing, institutions redesign how they think and the gains reach people like Sam. In The Split, the remarkable things all happen, to someone else. In The Shock, something large from outside stops the money and materials everything else assumed. Our view is that Britain is in The Long Grind today, drifting towards The Split. The Crossing is available, and it is mostly a matter of design and delivery.

Every call names the official figure that will settle it. Our confidence levels imply we should get about eight of the twelve right: all twelve would mean we were too timid, five that we were overconfident. Before publishing we put the calls to a panel of 24 synthetic forecasters. It thought two of our calls too easy, so we made them harder: unemployment below 5.5% rather than 6%, and fewer than 75% of NHS patients seen within 18 weeks rather than 92%. Where it thought us too sure, on farming and university, we have held and said so. The scorecard comes every October.

• Treat late 2026 to mid-2028 as one project. Carbon borders, battery passports, one-day settlement, identity wallets and AI rules all land in that stretch.
• Keep hiring juniors. Judgement has to be grown, and an organisation that stops hiring 23-year-olds now will find in 2034 that it has no 31-year-olds who know how anything works.
• Measure AI gains with your own controlled comparisons. Feelings of speed are not data.

2035 is exactly as far ahead of us as 2017 is behind. Nine years. Most of us can remember 2017 without effort, which is a useful corrective to the idea that the mid-2030s is science fiction.

A surprising amount of 2035 already exists. The eighteen-year-olds of that year are in primary school now. Everyone who will draw a state pension in 2035 has been alive for more than half a century. The children who will start school in England and Wales in 2029 and 2030 were born last year, and there were fewer of them than in any year since 1977 (ONS). The grid connections, reservoirs and railway lines that will be finished by then have, for the most part, already been applied for. Several have already been delayed.

So this is not a book of guesses. It starts with what is already settled: the laws with dates on them, the people already born, the contracts already signed. Then it moves to what is likely, and says how sure we are. Then it stops at the questions nobody can answer yet, and shows the two ways each could go.

We have done this for twelve parts of life, from work and money to food and power. And we finish by doing something forecasters generally avoid. We make twelve specific calls about 2035, write down what would prove each one wrong, and promise to mark our own homework in public every year.

A report about everything is an absurd thing to attempt. We know. But most people do not live in a sector. They live in a house, with a job, a body, a bank account and a view about the news. The future arrives in all of those at once, and it seemed worth trying to describe it the same way.

One change runs under every chapter of this report, so it is worth stating once, plainly.

Machine intelligence is getting cheaper at a remarkable rate. Epoch AI found the price of a given level of AI capability falling by between 9 and 900 times a year, depending on the level. The length of task an AI system can complete on its own has been doubling roughly every seven months over the long run, and faster since 2024, according to METR. Four of the largest American technology companies, Amazon, Google, Meta and Microsoft, plan to spend around $735bn to $750bn on infrastructure this year, on one compilation of their filings.

Economics has a useful rule of thumb for moments like this. When one thing becomes abundant, value moves to whatever is still scarce beside it. Cheap printing made attention valuable. Cheap shipping made location matter less and brands matter more. So the useful question about the next decade is this: when intelligence is everywhere, what runs short?

Across the twelve domains, the same four answers keep turning up.

Proof. When anything can be generated, the valuable thing is knowing what is real: who made this, where it came from, whether the person on the call is a person. Trust in news across 48 countries stands at 37%, the lowest the Reuters Institute has recorded since 2015. Meanwhile the law is quietly building a proof layer: identity wallets, battery passports, labels on synthetic media. Chapter after chapter, the scarce thing is a claim you can check.

People. The decade's deepest shortage is human. From this year, more people die in the UK than are born, on the ONS's latest projections. Sixty-three countries holding 28% of the world's population have already passed their peak (UN). Carers, builders, teachers of physics and engineers who can connect things to the grid are all short now and will be shorter. The human hour gets more valuable just as the machine hour approaches free.

Physical capacity. Software can be copied. Substations cannot. More than 2,500 gigawatts of energy projects are waiting in grid connection queues worldwide, says the International Energy Agency. England faces a public water supply shortfall of 5 billion litres a day by 2055 (Environment Agency). British construction needs about 41,000 extra workers every year to 2030 (CITB). The pace of the 2030s will be set by concrete, copper, water and planning permission far more than by code.

Coherence. The fourth scarcity is the ability to join things up. Most of the failures in this report are failures of connection: brilliant tools nobody has absorbed, laws nobody enforces, targets with no delivery plan. Advantage goes to whoever can make people, machines, money and rules pull in one direction, whether that is a company, a city or a country.

An honest caveat. Cheap is not the same as useful. In a careful trial, experienced software developers given AI tools took 19% longer to finish their work while believing they had been faster. Three in ten UK businesses use AI at all, a share the ONS describes as broadly stable. The gap between what the machines can do and what organisations actually do with them is wide, and it is the subject of much of what follows.

The Future Collective has spent several years building a vocabulary for this shift, much of it set out in The Great Rewiring. Four ideas from that work do a lot of lifting in this report, so here they are in one place.

Three Minds. There are three ways a person can stand in relation to machine intelligence. The Isolated Mind thinks alone. The Augmented Mind thinks in partnership, which is where most of us are heading, often without noticing. The Integrated Mind thinks as a system, no longer separating the human contribution from the artificial one. The book's claim is that the crossing is one-way: a mind that has learned to think with AI does not go back, any more than a reader can return to not reading. Much of the next decade is the story of people, teams and institutions making that crossing at very different speeds, and not all of them completing it.

The Human Wall. This is the point at which an organisation's technology outruns its people's capacity to absorb it. Every organisation hits it, and buying more technology does not get you over. The wall explains the caveat at the end of the last section, and it appears in almost every chapter: in classrooms where AI raised practice scores and lowered exam results, in hospitals where only a third of clinicians used a tool heavily, in businesses whose adoption figures have stopped rising.

Cognitive Capital. This is the sum of our capacity to understand, to connect and to decide. It grows through use, compounds when shared, and lives in minds, networks and cultures. It is built from attention, curiosity, insight and purpose. If intelligence is becoming cheap, cognitive capital is the scarce human counterpart, and it sits behind the second of our four scarcities. Several chapters ask whether a society is building it or spending it.

Intelligence Architecture. This is the deliberate design of how an organisation thinks: how intelligence moves through it, where it sticks, how people and machines reason together, how judgement improves with each cycle. It sits a layer above strategy. When intelligence is abundant, the scarce skill is structuring it, which is our fourth scarcity, coherence, seen from inside a single organisation.

Put together, they give a simple reading of the decade. People are crossing from Isolated to Augmented. Institutions keep hitting the Human Wall. What gets them over is cognitive capital and deliberate architecture, with proof and physical capacity as the outside constraints no amount of cleverness removes.

Some of the next nine years is already in the diary. Everything below is in law, in formal regulation, in a published government plan or fixed by arithmetic, and the dates were checked in October 2026 against official pages.

Two things stand out. The first is how much lands in 2027 and 2028. Carbon borders, battery passports, one-day settlement, identity wallets and AI rules all arrive between late 2026 and mid-2028. Anyone running a business should treat that stretch as a single project.

The second is how often the diary slips. The EU's high-risk AI rules, its carbon price on heating, its deforestation law, Britain's junk-food advertising ban and the analogue landline switch-off have all been postponed at least once. HS2 was meant to be on this list and now opens no earlier than May 2036. A date in law is the strongest kind of prediction available, and it is still only a prediction. Only the eclipse and the asteroid are guaranteed to turn up on time.

Each of the twelve chapters in part two follows the same five steps, so you can read one or all of them.

1. Already locked in. Laws, plans, contracts and demographics that are settled.
2. Likely. Where the evidence points, and how much weight it will bear.
3. The open question. The one uncertainty that matters most, with the best case for each way it could go.
4. What breaks. What is already going wrong, including at least one finding that cuts against the excitement.
5. What to do. Three things for someone who leads an organisation, and three for anyone.

Between steps four and five sits a short field report, "Seen in the wild". Each chapter ends with a call: one prediction for 2035, how confident we are, and what would prove us wrong. Part four gathers all twelve.

Three notes on method. First, figures are linked to their sources wherever we could, and the sources are official statistics and named research wherever they exist. Where a number comes from a company or lobby group with an interest in it, we say so. Second, the lens is British, because that is where we stand, but each chapter looks outward, and several of the most important facts in this report are about India, China and Africa. Third, the evidence was gathered in October 2026. Some of it will be out of date by the time you read this, and the annual scorecard is our way of owning that.

A word on prediction itself. The Future Collective has long argued that good questions build more capacity than confident forecasts, and nothing here changes that view. The calls in part four are not claims to know the future. They are a discipline. Writing down what you expect, and what would change your mind, is the quickest way we know to find out where your thinking is weak.

A universal basic income pays every adult the same sum, with no means test and no work test. Paying each UK adult £8,040 a year, roughly the poverty line, and each child half that, costs £472bn before tax, about a fifth of GDP. Most of it comes back from the same people in tax. With a 50% tax rate on the earnings of those who gain, and the benefits it replaces scrapped, the real transfer is about £45bn, or 2% of GDP (Widerquist and Arndt).

Cheaper designs start lower and keep more of today's system. Compass proposed £60 a week per working-age adult, £40 per child and £175 per pensioner. It would scrap the personal allowance, child benefit and the state pension, raise income tax by 3p, keep means-tested benefits and cost about 1.4% of GDP net (Compass). The catch, found at the University of Bath: a UBI at today's benefit rates needs £288bn more tax, and a cost-neutral one leaves many poor and disabled households worse off (University of Bath). Rent and disability costs vary too much for one flat payment.

The pilots are encouraging but small. Finland paid 2,000 unemployed people €560 a month in 2017 and 2018: wellbeing rose, employment barely moved (University of Helsinki). In Kenya, $22.50 a month did not reduce work (GiveDirectly). In Stockton, California, 125 people got $500 a month and found full-time work faster than those without it (SEED). The largest US trial, $1,000 a month for three years, cut paid work by one to two hours a week (NBER). None tested a country paying for it through tax.

Meanwhile, working-age benefits in Great Britain cost 4.1% of GDP in 2024-25, a record. Official forecasts put it at about 4.3% from 2026-27 to 2030-31 (DWP).

The NHS will still be here in 2035. The real question is whether it keeps its main promise on waits: that 92% of people waiting for planned hospital treatment have waited under 18 weeks. In August 2026 the figure was 65.2%. The list held 7.4 million cases, about 6.2 million people (NHS England).

The government has promised 92% by March 2029, with 70% due by March 2027 (Full Fact). The 65% milestone for March 2026 was met, but about 70% of a year's progress came in the last two months, helped by £120 million of short-term money (Nuffield Trust).

The case for getting there. It has been done. Between 2007 and 2010 the share rose from 57% to 91%, and it stayed above 92% from 2012 to 2015.

The case against. That was done with budgets growing much faster than now. On current trends the list falls to about 4.7 million by 2029 and the standard is still missed (Health Foundation).

Money goes up regardless. UK healthcare spending reached 11.4% of GDP in 2025, against about 10% in the decade before the pandemic. The state pays 81% of it, but private insurance is the fastest-growing slice, up 13.8% in real terms in 2024 (ONS). The OBR projects public health spending rising from 8% of GDP in 2030-31 to 13% by 2075-76 (OBR, July 2026). The NHS is not going away. It is getting dearer, and more of those who can pay are going round it.

Per person, yes. Once prices and population growth are stripped out, British shops sold about 9% less per head in 2025 than at the 2021 peak, and about 7% less than in 2019 (ONS; ONS). Total volumes in 2025 were below 2019's, while Great Britain gained 2.8 million people. Per person, we buy roughly what we bought in 2007.

It has happened before and reversed. Volumes per head fell for five years after 2007, then rose by a fifth between 2013 and 2021. The first eight months of 2026 were a little stronger, at about 92 on the scale above.

Other measures point the same way. The UK's carbon footprint, which counts emissions from goods made abroad, was 699 million tonnes in 2023, 29% below its 2007 peak; 53% of it now comes from imports (Defra). The ONS put the UK's material footprint 34% below its 2004 peak by 2018 (ONS). Households were spending more on services than on goods by the mid-2010s (Carbon Commentary, from ONS data).

The case the other way is clothing. The UK bought 1.42 million tonnes of textiles in 2022, close to pre-Covid levels and the most per head in Europe (WRAP, via edie). Second-hand is growing: 54% of Britons bought used clothes online in 2024 (Just Style). That is extra buying as much as a substitute. What caps a rebound is income: the OBR expects real disposable income per person to grow by under 0.5% a year to 2030-31 (OBR).

Britain buys far more goods than it sells, and the gap keeps widening. In 2025 the UK imported £603bn of goods and exported £378bn, a deficit of £226bn excluding precious metals (ONS). On the headline balance of payments measure the goods deficit was 7.7% of GDP in 2025, the widest on records going back to 1948; in 1997 it was 1.7% (ONS, ONS).

Services fill most of the hole. Counting them, the UK was only £22bn short in 2025 (ONS). We sell finance, law and software, and buy cars, phones and fuel.

Manufacturing fell from 14.8% of the economy’s value added in 2000 to 9.1% in 2023 (University of Cambridge). Car making shows the trend. UK plants built 717,371 cars in 2025, down 8%, and exported 77.5% of them. The government’s ambition is more than 1.3 million vehicles a year by 2035 (SMMT), and its sector plan aims to nearly double business investment in advanced manufacturing (Department for Business and Trade).

Two things push the other way. North Sea output keeps falling, and the industry body expects up to 80% of oil and gas demand to be met by imports by 2030 (OEUK). And the goods deficit has narrowed sharply only once in forty years, from about 4% of GDP in 1989 to 1.5% in 1994. That took a deep recession and the pound’s exit from the ERM (ONS).

Britain grows about three-fifths of the food it eats, and has done for 20 years. Defra's production to supply ratio, which compares the farm-gate value of UK food with what the country consumes, was 60% in 2025, down from 65% in 2024 as imports rose in value. It peaked at 78% in 1984 and has stayed between 59% and 65% since 2005 (Defra).

The farms behind that figure are thinning, not vanishing. The UK had about 212,000 holdings in 2025, against 216,000 in 2021, and 446,000 people working on them, 1.5% fewer than a year before (Defra). One farm in five in Great Britain made no profit in 2024/25 (Defra). Total income from farming rose 21% to £8.4bn in 2025, and £4.9bn came from support payments (Defra).

The pressures are real but slow. Since April 2026, farm and business assets above £2.5m get only half relief from inheritance tax, which the government expects to catch up to 185 farming estates in the first year (HMRC). The Land Use Framework expects about 9% of England's farmland to move mainly to nature by 2050, with roughly 2% each going to solar and wind and to housing (Defra).

Two narrower tests from an earlier draft of this chapter still stand. Cultivated meat should stay below 1% of meat sold in UK shops. And at least two more English cereal harvests before 2035 should rank among the ten smallest since Defra's regional series began, which is only in 1999.

For most graduates, yes. The universities themselves are the weaker part.

The Institute for Fiscal Studies, working for the Department for Education, estimates that a degree adds about £100,000 to lifetime earnings on average, after tax and loan repayments. The typical gain is £63,000. About one graduate in four, and nearly one man in three, would have been better off without it (IFS).

School leavers still think it pays. A record 320,570 UK 18-year-olds were placed in 2025, and the entry rate has stayed above 35% since 2020 (UCAS). Early figures for 2026 show a further small dip, to 35.4% four weeks after results day, against 36.1% at the same point in 2025 (Times Higher Education).

The money is the problem. International enrolments fell about 6% in 2024/25, to 685,565, led by a 10% fall in taught master’s courses from outside the EU (ICEF Monitor). Study visa applications in early 2026 were 30.6% lower than a year before (Office for Students). No English university has yet closed through insolvency, but MPs say one could (Newswav). Kent and Greenwich have merged instead, into a group with more than 50,000 students (ITV News).

After 2030 the number of 18-year-olds starts to fall. HEPI expects demand to drop by getting on for 20% between 2030 and 2040 unless more of them go (HEPI). Fewer applicants means universities close or merge. It need not mean a smaller share of young people get in.

Our own view is that Britain is in The Long Grind today, with The Split as the direction of drift if nothing changes. The Crossing is available. It is mostly a matter of design and delivery, which is to say, of the two things that are hardest and least glamorous.

If these numbers mean anything, we should get about eight of the twelve right. If we get all twelve, we were too timid and should have said things that were harder to know. If we get five, we were overconfident and you should discount us accordingly. Being wrong on some of these is what honest confidence levels look like.

The odds in this section are synthetic: they come from language models asked to forecast, not from people, and they are published whole, including the ones that embarrass us. How the panel was built, and what it cannot do, is on the method page.

Before publishing we put the twelve calls to a panel of 24 synthetic forecasters: twelve analytical lenses (a labour economist, a demographer, an energy systems analyst, a public finance specialist, an infrastructure delivery specialist, a political scientist, an actuary and others), each run on two different models, each working alone, each told exactly how a call would be marked but not how confident we were, and each allowed to look up the latest official figures before answering. Ensembles of language models have matched human forecasting crowds in published tests, which makes this the one use of synthetic voices with evidence behind it. The panel's job is to tell us where we may be fooling ourselves, not to replace our judgement.

* The panel judged the earlier, easier versions of these two calls (unemployment below 6%; fewer than 92% seen within 18 weeks). We made both harder after reading its view, so its figures for them are not comparable with ours.

The splits run both ways. On news from AI, farming and university the panel thinks we are too sure; on the NHS, trade and the new towns it thinks we are not sure enough.

Six of the calls were put to the panel twice. The first round, on 11 October, let each forecaster search for the latest figures, and it showed that retrieval mattered more than reasoning: the forecasters who missed a twelve-day-old announcement on the state pension sat 22 points below those who found it. So when we replaced six calls with plainer questions (on a basic income, the NHS, consumption, trade, farming and university), the second round gave every forecaster the same dated evidence pack and no search. Disagreements in that round are about judgement, not news.

Taken together, the panel expected about 7.9 of the twelve calls to come right; our own numbers imply 7.9 now, and 8.1 before the two changes described below. It is surer than we are of the calls where a long trend has to break for us to be wrong: the NHS missing 18 weeks, the trade deficit, the new towns, unemployment. It is less sure of the calls that turn on behaviour or a knife-edge threshold: news from AI most of all, then farming, university, renewables and the population. Its doubt on news is specific: it expects weekly use of AI for news to plateau in the twenties, as podcasts did, and its own test is the 2027 Digital News Report, where 14% or more keeps the call alive and 11% or less means it is stalling. We will apply that test in our first scorecard.

Where the panel gives us pause: farming and university. On farming it puts the call at 47%, a coin toss, because the UK's food self-sufficiency already sits close to our 60% line and has drifted down for decades. On university it puts us at 55%, because the entry rate for 18-year-olds slipped in 2025 and again on early 2026 figures and the panel weighs the falling graduate premium and the strain on university finances more heavily than we do. We have held both, and say so: if the 2027 figures move against us we will lower our confidence in public.

Where we listened: the NHS and jobs. The panel put our original NHS call, that the 92% standard would still be missed, at 87%. That told us the call was too easy, not that we were too cautious, so we made it harder: the call is now that fewer than 75% will have been seen within 18 weeks in March 2035, which bets that the recovery stalls well short of the promise rather than merely missing it. On jobs the panel put us at 78%, so we moved the line from 6% to 5.5%, below the level in the Office for Budget Responsibility's own scenario where technology pushes structural unemployment up. Our confidence in each is lower as a result, 55% and 60%. If the report's rule is that getting all twelve right means we were too timid, these are the two places we have tried to be less so.

The panel also found a marking hazard in every one of the twelve calls, from the ONS's switch to a new labour force survey to the absence of any official count of homes completed in new towns. The marking rules we will hold ourselves to are on the method page, because a call that cannot be marked is not a call.

Read down the answers and a view of the decade appears. Most of them say that something people fear or hope for does not arrive on time: AI does not empty the job market, the homes are not built, the NHS does not get back to 18 weeks, Britain does not make more of what it buys, the basic income does not come. Several say that a slow trend keeps going: more deaths than births, less stuff bought per head, a farming sector that shrinks without losing its share of the plate. Only a few say that something moves fast: the machines' appetite for electricity, the share of it that comes from renewables, and the number of people who get their news from a machine.

That is the argument of this report in twelve lines. Intelligence is getting cheap fast. Almost everything it needs in order to matter (people, proof, physical capacity and the ability to join them up) is scarce and slow.

We have not predicted the outcome of any war or election, and we have not put a date on any breakthrough in artificial intelligence. Those are the things people most want forecast and the things least possible to forecast. We would sooner say so.

Every October The Future Collective will publish where each call stands: on track, off track or settled. Where the evidence has moved, we will say how our confidence has changed and why. If we abandon a call we will say that too, and leave the original on the page.

A forecast tells you what to expect. A question tells you what to do. These are the twelve we would put to any leadership team, and to ourselves.

1. Where will your next generation of experienced people come from, if you are not hiring beginners now?
2. What can your organisation do with AI today that it could not do last year, and how do you know?
3. Which promises have you made that depend on money you have not yet found?
4. Who among your people cannot get the care the others can?
5. What do you need from the grid, the water company and the planning office, and have you asked?
6. If a machine chose for your customer, would it choose you?
7. Which single country, supplier or route could stop you for six weeks?
8. Where will your people live?
9. What have you assumed will stay cheap?
10. Are your people becoming more capable, or only more productive?
11. How would a stranger check that what you say is true?
12. What have you assumed the state will keep doing for you?

Three layers, each with a different job.

1. A synthetic population. About 2,000 households (48 archetypes in the pilot) built to match ONS totals for age, region, income, tenure, work status, household type and health. This is the standard technique behind transport and tax microsimulation. No AI is involved in building it.
2. A persona layer. A language model gives each household a voice and answers structured questions as that person, many times over, each run independent of the others.
3. An analysis layer. Answers are weighted back to the population and reported as distributions with ranges, never as single point estimates.

1. Calibration, now. Ask the personas about today, where the answers are known: hybrid working, private healthcare, loneliness, trust in government, news habits, food insecurity, heat pumps, smoking. Publish the gaps, including the embarrassing ones. If the population cannot reproduce 2026, it has no business describing 2035.
2. Living the four worlds. Put every household through The Long Grind, The Crossing, The Split and The Shock. The structural numbers (pensions, rents, energy prices, employment) are fixed from the chapters; the personas are asked only what the household does and how it feels. This turns Sam in Stockport into two thousand Sams, and part three from fiction into something a reader can interrogate.
3. Odds on the calls. A panel of diverse synthetic forecasters puts a probability on each of the twelve calls, with reasons. Where they disagree with our confidence levels, we look again. Ensembles of language models have matched human forecasting crowds in published tests, so this is the one use with direct evidence behind it. Run on 11 October 2026; results in part four and at the foot of this page.

• Calibration results are published whole, misses included.
• Model, prompt and population version are recorded for every run, so any number can be rerun.
• A synthetic number never stands alone; the nearest real figure sits beside it.
• Real interviews ground a subset of agents.

Forty-eight archetypes span age (four bands), region (London; the South; the Midlands and North; Scotland, Wales and Northern Ireland) and household income (three bands), with tenure, work status, household type, health, ethnicity and 2024 vote assigned to match published totals. Each archetype was run twice on Claude Sonnet and once on Claude Fable 5.1, 144 synthetic responses in all. Each run invents a specific person within the archetype's fixed attributes, describes an ordinary Tuesday in their voice, and then answers the battery below as that person would answer a market researcher in October 2026. Option order and question order are shuffled per run, each run is a fresh context, and the personas are not steered towards the right answer; the point is to measure the raw bias.

Calibration battery

Two further checks test the population weights rather than the personas: work status against the ONS labour market figures (employment 75.1%, unemployment 4.9%, inactivity 20.9%), and tenure against the English Housing Survey.

The research on synthetic respondents is young, mostly American, and largely sceptical. The main findings, each of which shaped the pilot design and each of which the pilot then reproduced.

Right averages, wrong distributions. Bisbee and colleagues (Political Analysis, 2024) had ChatGPT answer American National Election Study questions as thousands of described respondents. The averages matched the real survey closely; the variation was smaller than in real responses, regression coefficients often differed significantly, the distributions moved with minor changes of prompt wording, and the same prompt gave different results three months apart.

A bias of the model's own. Boelaert and colleagues (Sociological Methods and Research, 2025) call it machine bias: on each topic a model shows a strong bias and low variance, and the bias changes unpredictably from one topic to the next. Their conclusion is that current models cannot yet replace research subjects in opinion or attitude research.

Caricature. Wang, Morgenstern and Dickerson (Nature Machine Intelligence, 2025) show that models asked to speak as a group produce portrayals closer to how outsiders stereotype that group than to how its members describe themselves, and flatten the variety inside the group.

Survey mechanics. Dominguez-Olmedo, Hardt and Mendler-Dünner (NeurIPS 2024) found a strong preference for whichever option is listed first and, once that is corrected, answers that resemble no real demographic group. Synthetic respondents almost never say they don't know.

The largest independent test, published on 30 September 2026. Pew Research Center (30 September 2026) built digital twins of about 6,700 members of its American Trends Panel and re-ran three 2026 surveys, nearly 300 questions, through Claude Opus 4.6 and GPT-5.1. The synthetic results differed from the real ones by an average of 12 percentage points (11.4 for Opus, 13.3 for GPT), by more than 15 points on 28% of questions, and by most for Republicans (16.1 points) and Black adults (15.1). Nearly half the questions had at least one answer option that no synthetic respondent chose. The twins knew far more than real people (98% could name a First Amendment right, against 52% of humans) and said "not sure" a quarter as often. The two models erred in opposite directions: GPT made Americans more extreme than they are, Opus more middle-of-the-road. Pew's verdict: not an adequate replacement for polling.

The professional standard. The AAPOR task force on AI in survey research (2026) calls synthetic responses the riskiest use of AI in the field, says there is no consensus on when, if ever, they can stand in for human ones, and draws the acceptable boundary at clearly labelled pretesting, pilot work and exploratory diagnostics. It prefers the term synthetic responses to synthetic samples, because the method is not a sampling design, and says the next tests should ask whether correlations, joint distributions and subgroup interactions are preserved, not only the marginal totals.

Where the evidence is better. Two findings cut the other way, and both are built into the plan. Park and colleagues (Stanford, 2024) built agents for 1,052 real people from two-hour interviews; those agents reproduced each person's own General Social Survey answers at 83 to 86% of the person's own two-week test-retest consistency, against 74% for agents built from demographics alone, with smaller accuracy gaps across racial and ideological groups. And Schoenegger and colleagues (Science Advances, 2024) found that an ensemble of twelve language models matched a human forecasting crowd on a set of real-world questions, which is the evidence behind using a synthetic panel to put odds on the twelve calls.

What this meant for the pilot. Option order and question order shuffled per run; one respondent per fresh context; a specific invented person rather than a demographic label; 2024 vote included in every persona, because politics predicts attitudes; raw, unsteered answers so the bias can be measured; subgroup and within-archetype checks, not just totals; and every figure labelled synthetic. One weakness no design choice can fix: an answer that no synthetic respondent ever gives cannot be reweighted into existence afterwards.

All figures in this section are synthetic responses, weighted to the UK adult population by archetype. They are not measurements of real people. The real figure sits beside each one.

The run. 11 October 2026. Every response kept the attributes it was given: 144 of 144 matched the assigned work status and the presence of children, so the population layer does its job of holding the structure fixed. The archetype weights reproduce the ONS shares for age, region, income, tenure and work status to within a point. Ranges in brackets are 90% intervals from resampling the 48 archetypes, which is the honest size of the pilot's uncertainty.

The headline averages come out close because the errors cancel: whole groups vanish at one end while the comfortable majority is overdrawn at the other.

Mean absolute error against the 27 anchors: Sonnet 11.6 points, Fable 9.6 points. Pew's independent test of the best available model on 300 American questions came out at 11.4. We are where the state of the art is, and the state of the art is not good enough to publish as measurement.

What the pilot shows

1. The averages that look right are often right for the wrong reason. Hybrid working comes out at 32% against a real 28%, but underneath, 81 to 90% of workers on £50,000 or more are hybrid (real: 45%) and none under £20,000 (real: 8%); 71% of graduates (real: 41%) and no non-graduates at all. Nobody works only from home. The personas treat hybrid working as a class badge, where real life hands it out unevenly across the whole range.
2. Missing minorities. Across 144 synthetic responses, not one felt lonely often or always (real: one adult in fifteen, one young adult in eleven); not one gave the government six or more out of ten (real: more than a quarter; the synthetic scores bunch at two, three and four); not one expected children today to be better off than their parents (real: nearly one in five; all 144 said worse off); not one had used an AI chatbot for news that week (real: 4%, a small miss); not one under 35 smoked; and only two of 144 expected their own finances to improve. Pew found the same thing: nearly half its questions had an option no synthetic respondent ever chose. These are the people a futures report most needs to see, the ones who do not fit their demographic, and the method erases them.
3. Everything sorts by income. Fable gave private healthcare to 13 of its 16 higher-income personas and one of the other 32; Sonnet gave it to two of 96. In reality 16% of adults in England used it last year, spread across the income range. Every food-insecure household in both models sits in the lower-income third (17 of Sonnet's 32 lower-income responses, none of the other 64); all six smokers are lower-income and none is under 35. Income predicts these things in the real data; it does not determine them.
4. The model is a variable, and the two models lean opposite ways. Sonnet's Britain is poorer and gloomier: 21% food insecure, 62% expecting their finances to worsen, 10% trusting the news (real 30%), 34% trusting the courts (real 62%), almost nobody paying privately for health or advice. Fable's Britain is more comfortable and more trusting: 25% private healthcare, 18% paying for news (real 10%), 21% vaping (real 10%), 14% taking financial advice (real 8.6%), but close to the real figure on the courts, cash holdings, AI use and trust in news. Fable was closer on nine anchors, Sonnet on six, with twelve ties (mostly shared zeros). Pew saw the same split between GPT-5.1 (too extreme) and Opus 4.6 (too moderate). Any full build has to run more than one model and publish the spread.
5. Inside an archetype the model barely varies. Sonnet's two independent runs of the same archetype gave the same answer to the same question 96% of the time at the median (75% at worst), and their 0-to-10 trust scores differed by 0.2 points on average. Real people who share a demographic cell disagree far more than that. Twenty-one of the 48 archetypes produced the same first name in two of the three runs, and seven of the 144 people are called Colin. Extra runs of the same archetype buy almost nothing; what is missing is variety within the cell.
6. Direction is usually right; level is not. The age gradient in AI use runs the right way (but at 100% of under-35s and 0% of over-75s it is a cliff, not a slope). Hybrid working by age lands within a few points of the ONS bands. Households with children are more food insecure than those without. Cash holding among savers (59 to 65% against 61%), overall AI use, and Fable's trust in the courts are close. Smoking is undercounted by half in both models.

Our call on the method

As a measuring instrument, the raw method fails the calibration test, in exactly the ways the literature predicted: plausible totals, caricatured subgroups, missing minorities, and a tilt that depends on which model you ask. We will say so in the report, in full, and we will not publish a synthetic level as if it described Britain.

It is not useless, and the pilot says how to use it.

• Living the four worlds (use two) survives as a comparative tool. The same household run through The Long Grind and The Shock gives a defensible answer to "what changes", even where its absolute level is wrong, because the bias is largely shared across worlds. Report differences between worlds, never levels, always labelled, always beside a real figure.
• Odds on the calls (use three) is unaffected. It rests on different evidence (ensemble forecasting) and does not depend on persona fidelity.
• For the full build, in order of value: build the 2,000 households by iterative proportional fitting rather than running more repeats of 48; seed the within-cell traits the pilot lost (loneliness, trust, optimism, smoking, private spending) from their known marginals so the minorities exist, then test on held-out questions, since the seeded ones are right by construction; ground a subset of agents in the real interviews (the Stanford work took demographics-only agents from 74% to 83 to 86% of a human benchmark); run Fable as the primary model and Sonnet as the check, and publish both; break the default persona with a seeded name and detail step; and record the exact model version string on every run, which this pilot did not.
• Budget. Two thousand households, four worlds and three question rounds is about 24,000 runs. At this pilot's prompt that is over two billion tokens; a production prompt without the tool context and with batched questions is roughly a tenth of that.

The pilot's own numbers, scripts, archetype definitions and all 144 synthetic responses are available from The Future Collective on request so any figure here can be rerun.

Synthetic. Twenty-four forecasts per call from language models, not people.

Design. Twelve analytical lenses (a base-rate generalist, a labour economist, an energy systems analyst, a public finance and pensions specialist, a demographer, a health inequalities epidemiologist, an infrastructure delivery specialist, a retail and consumer analyst, a media researcher, a historian of technology adoption, a political scientist and an actuary), each run once on Claude Fable 5.1 and once on Claude Sonnet, each run covering six of the twelve calls in shuffled order: 48 runs, 24 independent forecasts per call. Every forecaster was given the call, the settling body and the exact marking rule, and was not told our confidence. Each could use up to two web searches per call to find the current value of the indicator, and had to write down its reading of the rule, the current state, an outside view, the strongest case each way, the likeliest failure path, a probability from 1 to 99, its confidence in its own number and the one 2027 figure that would change its mind. Compound calls were estimated part by part.

What the panel taught us about the panel.

• Retrieval decides more than reasoning. The two lowest numbers in the whole exercise, 14% on Power and 33% on Money, came from forecasters working without the latest figures, one reasoning carefully from memory of projections two years out of date, the other from a July report. Nineteen of the 24 forecasters missed a decisive announcement on the triple lock that was twelve days old; the five who found it sat 22 points above the rest. Any future run needs a shared, dated evidence pack per call before the forecasters start, so that disagreement is about judgement rather than news.
• The model tilts. Sonnet sat on average 3 points above the call medians and Fable 2 points below; Sonnet was markedly more confident in our calls on Places (92 against 69), Power (77 against 54), Work (85 against 69) and Buying (72 against 56). Two models are the minimum; three would be better.
• The lens does not. No lens sat more than 2.5 points from the medians on average, so the analytical persona did not drive the answer. Diversity of evidence mattered more than diversity of character.
• The panel knows what it does not know. It called itself low-confidence on the two political and behavioural calls (Money, 20 of 24 forecasters; Truth, 23 of 24) and high-confidence on the two structural ones (HS2, 22 of 24; pupils, 21 of 24), and its spread followed suit.
• Compound calls are the hard ones. Food, Truth and Power each join two conditions with an "and". The panel priced the parts separately and the joint came out well below either part (Truth: 50% for the adoption leg, 71% for the trust leg, 34% joint). Our own confidence levels had not fully priced the "and".

Marking rules. The panel found a hazard in every call. These are the rules we will mark ourselves by.

1. Work. The ONS headline rate for people aged 16 and over, for February to April 2035, on whichever survey is the headline in June 2035, is below 5.5%. If the ONS has moved to a new survey and published a back-series, the new basis applies throughout.
2. Machines. The IEA's latest published central estimate for 2035 standing at the time of marking, on the IEA's own definition of data centres (crypto mining excluded if the IEA excludes it).
3. Energy and climate. The latest Ember edition standing in October 2036 for calendar 2035, renewables excluding nuclear, after Ember's revisions.
4. Places. Completions inside the boundaries of the seven sites in the March 2026 draft programme, or the final list that replaces it, from each site's designation date, with the two extension sites' pre-designation pipelines excluded. If no official attributable series exists we will assemble the count from local authority monitoring and publish the working.
5. Truth. The Reuters Institute's all-market averages as published, on whatever set of markets it surveys that year. If the AI question is reworded, the nearest equivalent, with a note.
6. Power. The ONS vintage standing in October 2036 for the mid-2035 estimate, after any census rebasing. A year that is revised from natural increase to natural decrease, or the reverse, counts as revised.

The panel's 48 runs, the assignments each forecaster received and the aggregation script are available from The Future Collective on request.

Synthetic. Twenty-four forecasts per call from language models, not people.

On the evening of 11 October we replaced six calls with plainer questions: a basic income (Money), the NHS (Health), consumption (Buying), trade (Making and moving), farming (Food) and university (Learning). The same 24 forecasters, twelve lenses on two models, judged the new calls, but this time with no web search: each was given the same dated evidence pack of 11 or 12 sourced facts per call, so that disagreement would be about judgement rather than news.

After this run we made two calls harder: unemployment below 5.5% (was 6%) and fewer than 75% of NHS patients seen within 18 weeks (was 92%). The tables above record the confidence and panel figures for the earlier versions.

With a shared evidence pack the spread narrowed sharply: no call's middle half was wider than 13 points, against 24 in the first round. Sonnet again sat above Fable, on five of the six; university was the exception.

Marking rules for the new calls

1. Money. Both must hold: no UK-wide unconditional cash payment to all adults is in payment on 31 December 2035 (pilots and local schemes do not count); and the first DWP benefit expenditure tables to report an outturn for 2034-35 show working-age benefit spending in Great Britain strictly above 4.00% of GDP.
2. Health. NHS England's England-level share of incomplete referral-to-treatment pathways within 18 weeks for March 2035, as first published and including estimates for missing trusts, is below 75.0%. If the measure is redefined, NHS England's official successor applies; if none exists the call is void.
3. Buying. ONS annual retail sales volume (all retailing, J5EK) divided by the Great Britain mid-year population, for 2035 against 2021, both from the vintage current when 2035 data first appear. Right if 2035 is strictly below 2021.
4. Making and moving. The ONS trade in goods balance (BOKI) for 2035 divided by nominal GDP (YBHA), on the first full-year estimate at the end of March 2036. Right if the deficit is 6.0% of GDP or more, rounded to one decimal place.
5. Food. Defra's all-food UK production to supply ratio for 2035, as first published in Agriculture in the United Kingdom 2035, as the whole number Defra prints. Right at 60 or more; later revisions ignored.
6. Learning. UCAS's end-of-cycle UK 18-year-old entry rate for the 2035 cycle, as first published. Right at 35.0% or more.