What the world looks like after the machines can do the work

How to read this

Every number on this page comes from the machine-readable appendix each respondent was required to end with, so the tables compare like with like. The prose sections are this session's synthesis of all ten answers. Every full answer, unedited, is one click away at the bottom, and so is the exact question as sent.

What this is not

Ten language models are not ten independent experts. They share training data, and the prompt's structure shaped their answers. Where they agree, read it as "no model could find a reason to disagree". Where they split, the split is the finding. Confidence, in their own numbers, runs from 0.35 to 0.75.

All ten describe the same machine. Wages for any task a model can do fall toward the cost of compute; employment in those tasks falls three to eight years behind capability, gated by liability, regulation and the reluctance of managers to fire themselves. Prices of cognitive output fall by an order of magnitude, then physical goods follow the robots. What does not fall is what is scarce by law or by nature: land where people want to live, energy and grid capacity until the build-out catches up, minerals, licences, human attention, and positional goods. Every respondent lands on the same sentence about the consumption basket: goods get cheap, rent stays dear.

Profits rise for whoever owns compute, models, power and land. In every base case the machines are owned by a handful of American firms and the Chinese state, with broad-but-thin household participation through pension and index funds and control concentrated in a few boardrooms. Nobody forecasts broad ownership arriving on its own.

The demand problem is named by all ten and resolved the same way by nine: if wages collapse and the surplus pools with owners, nobody can buy the output, so the state redistributes, late, funded by taxes on capital, land, consumption and AI output, and by deficit monetisation where those lag. Falling prices do part of the work, which is the honest reason the transition is survivable. Kimi's phrase is the sharpest: capitalists need customers, and that argument wins after three to eight years of crisis politics, not before.

Interest rates: up during the compute and power build-out, then down hard on a savings glut. Government revenue is the binding constraint everyone identifies: about half of rich-country revenue rides on labour income, and that base shrinks by a third within a decade of AGI. The models that think hardest about this (Astra, the Fable agent) put numbers on it: a 30-point fall in the wage share removes roughly 7.5 percent of GDP in revenue, and a dividend worth 20 percent of GDP per head to 80 percent of people costs about 16 percent of GDP gross.

Cognitive jobs displaced, percent of 2025 positions

Physical jobs displaced

Peak non-employment, percent of working-age adults

Median household real income, 2040 versus 2025, percent

Probability of interstate conflict involving the region by 2040, percent

The ten answers describe one script with different dates. Retraining rhetoric first, triggered by nothing in particular. Wage subsidies and expanded unemployment cover when graduate non-employment passes about 15 percent. Working-time reduction and the first AI or capital tax fights when non-employment passes 20. A negative income tax or partial basic income after an election lost on the issue, placed between 2034 and 2038 by almost everyone. Then the fiscal reset: the tax base moved from labour to consumption, land, capital and AI output, plus state equity stakes or a sovereign fund, when the old base has visibly failed.

China is forecast to run the same script in reverse order and without waiting: state-directed deployment and SOE absorption from the start, a formal job guarantee by the early 2030s, dividends from state-owned AI by the mid 2030s, work reframed as civic duty and allocated, and surveillance to hold it together. Every respondent expects China to be materially better off and politically tighter in 2040; several expect it to be more stable than the democracies through the transition, at the price of agency.

On what works, the panel converges on the same three analogues. Sovereign funds and citizen dividends work where the rent is captured early (Norway, Alaska, the Gulf), at the cost of atrophy if there is nothing else. Transfers work but arrive late: Social Security passed six years after the crash. Retraining fails, and the China shock is the proof, a decade with no effective response and then populism. Enclosures are the worst case cited by eight of ten: productivity soared and it took a century of urban misery and political violence before the gains were shared, because displaced labour had neither ownership nor a vote.

The threshold at which redistribution becomes politically forced is put between 15 and 35 percent of working-age adults outside work, with most answers at 20 to 25. The consensus reason it arrives late is structural, not moral: the people who decide are the last to be displaced.

Roles that hold value for today's 40-year-old

What the next generation should build toward

Where the ten agree

• Accountability is the last moat. Roles where a named human must sign and carry liability (physician, surgeon, engineer of record, audit partner, director, judge) hold to about 2040 in every answer, at a fraction of today's headcount.
• Physical presence in unstructured places holds to the late 2030s. Electricians, plumbers, builders, on-site maintenance; every respondent dates the erosion to 2036 to 2040, when robots reach the messy world.
• Human preference for humans is real but pays badly. Care, early-years teaching, coaching, live performance, hospitality survive, mostly at low or winner-take-all pay.
• Ownership is the only role that holds indefinitely. Owner-operators who direct machines and hold the asset appear on eight of ten lists.
• Gone by 2032, by consensus: junior analysts, paralegals, copywriters, translators, developers below senior, customer service, back-office finance and accounting, radiology reading, most academic research production. "Gone" means hiring demand down by more than half, not zero humans.

For children now 9 to 14

• "Career" is the wrong frame in all ten answers. The replacement is a portfolio: one physical competence at trade level, one human-facing capability, ownership of capital and of something small, a dense community, and fluency at directing and checking machines.
• Learn: mathematics and statistics to the point of auditing a claim; writing and speaking that persuades a room; history, law and how institutions decide; a trade or craft with the hands; a sport kept for life; how to run a small enterprise.
• Credentials that still hold at entry: medicine and allied health, engineering with a licence, trades licences, law only for future signatories.
• Do not bother: coding as a vocation, law as volume practice, accounting, translation, generic business and commerce degrees, any credential whose value is the credential. Gemini goes further and drops medicine and engineering too; nobody else does.

Characteristics that rise

Characteristics that fall

The ranking is remarkably stable across ten independent answers. Agency, the habit of starting without being told, is in the top three on eight of ten lists, for the same reason each time: when execution is free, the scarce input is wanting something and being responsible for it. Trustworthiness is next, because in a world of fluent fakes a verifiable reputation is capital. Then judgment and taste, choosing well among infinite cheap options; physical vitality, the one asset nobody can rent; social warmth and the ability to lead a group; and the capacity to hold attention and find meaning without paid work, which several respondents name as the trait most likely to decide who thrives.

What depreciates is the profile the last century's professional class was selected for: processing speed, memory, credentialism, verbal fluency, diligence at routine tasks, narrow specialisation, obedience. Astra's line stands for all of them: intelligence does not become worthless, its market scarcity declines. On what can be built in a child, the answers converge on method as well as content: real responsibility with real consequences, sport played seriously, promises held binding at home, unstructured time with other people, devices rationed before the habit forms, and adults who model the stance.

What to stop, by consensus

• Stop optimising the children for ATAR, selective schools and university prestige as the goal; eight of ten say it in those words.
• Stop steering them toward law, accounting, IT or generic commerce because those rewarded intelligence in your generation.
• Stop paying for content tutoring and for credentials whose value is the credential.
• Stop assuming your own role is safe because it feels expert; research, analysis and other cognitive production are on every "gone by 2032" list, and universities are forecast at a third to a half of their present size by 2040.
• Stop doom-talking about the future in front of them, and stop promising it will be fine. Say instead that useful lives remain possible and preparation must be broader than a job title.

Fable 5.1, this session: the ten in order

1. This year: make AI a tool the children watch you direct and verify, out loud, every week.
2. This year: strength, sport and outdoor competence as non-negotiable as school.
3. This year: real responsibility with real consequences for each child.
4. Five years: one trade-level physical skill each, from a real practitioner.
5. Five years: each child owns capital in their own name and learns what it is.
6. Five years: a dense local community, held.
7. Five years: your own career from producer of cognitive output to verifier, bridge and institutional judge, plus one income stream that does not need your hours.
8. Fifteen years: geographic optionality for the children, and one more language.
9. Fifteen years: teach, explicitly, that worth is not a job.
10. Stop: the safe-profession funnel, credential premiums, scheduled hours, doom-talk.

Self-reported confidence in the numerical forecast, 0 to 1. Two respondents distinguished confidence in the direction (high) from confidence in the trajectory (low); the number shown is the one they put in the appendix.

The falsifiers they share

• Capability stalls. Named by every respondent: if frontier systems still cannot run multi-week or quarter-long real-world work unsupervised by 2030, every date shifts five to ten years and the slow-diffusion branch becomes the base case.
• Energy caps compute. If grids cannot supply the data centres, deployment halts regardless of capability.
• A democracy acts early. A citizen equity fund, universal dividend or AI-funded transfer above 1 percent of GDP legislated before 2030 raises the shared-abundance branch by 10 to 20 points in every answer that models it.
• Taiwan. A blockade or war before 2030 doubles the disorder branch and pulls forward capital controls.
• Robots arrive early. Reliable humanoids under about USD 30,000 by 2030 bring physical displacement forward five years and take the trades' decade away.

Where they are least sure is the same place: the political response. The economics of the comfortable-dependency and shared-abundance branches are nearly identical; which one a society lands in is decided by elections, elites and timing that no respondent claims to forecast.