Ministry of AI · Dispatch from 2047
AI and Labor Share of Income: The Pipe We Cut
Editorial note. The Ministry of AI is a work of disciplined foresight: it describes the year 2047 in the present tense, and treats our own era as history. The institutions are imagined. The economics, the evidence and the historical parallels are real and sourced.
My mother is seventy-three and still calls it her money, which is grammatically correct and historically strange. It arrives on the ninth of the month. She reads the statement, which nobody does any more, and then she puts it in the drawer with the paper bills from the years when she had a desk.
Nineteen years of reading insurance claim files. A pension that stopped being funded when the department stopped existing. Then four years of what the statisticians of that decade politely called non-participation, followed by a first dividend payment at fifty-nine that she assumed was an administrative error and did not spend for two months.
Here is what I want to say about those four years, because it is the part my sons find hardest to believe. Nobody stole from her. There was no villain, no conspiracy, no decision anywhere in any building that her income should end. What happened was more mundane and much worse: the pipe that had carried value to her was cut, and no one had ever thought of it as a pipe.
Wages were plumbing, not virtue
For most of the industrial era, if you wanted to know how the population got paid, the answer was payroll. Not because payroll was generous — it was fought for, decade by decade, strike by strike — but because production physically required people, and people had to be induced to show up. The wage was the toll capital paid for access to human judgment and human hands.
That arrangement produced a statistic economists watched closely and moralised heavily: the labour share of income, the fraction of everything produced that arrives as wages and salaries rather than as profit, rent or interest. Across the twentieth century it hovered, in most rich countries, somewhere around two-thirds. It felt like a law of nature. It was closer to a coincidence of technology.
The decline began long before any of this. From roughly 1980 the labour share fell across most advanced and many emerging economies — Our World in Data’s series shows the drift plainly — and the standard explanations were cheaper capital equipment, offshoring, industry composition and the erosion of bargaining institutions. Karabarbounis and Neiman documented the global pattern in 2013 and attributed much of it to the falling relative price of investment goods. Autor, Dorn, Katz, Patterson and Van Reenen later located the decline inside the rise of superstar firms: industries concentrating into a few enormously productive companies with unusually low wage bills.
So the trend was fifty years old when the models arrived. What the models did was remove the floor underneath it — the quiet assumption that however cheap capital became, some substantial part of production would still need paid human hours.
The arithmetic nobody disputed
The mechanism is not controversial and never was. The Philadelphia Fed spelled it out in 2024, while the effect was still too small to see in the aggregates: when a task is completed more cheaply by a machine than by a person, the cost saved does not evaporate. It is booked as margin. And margin, as they put it, contributes to capital income but not to wages.
I was in the rooms where this was executed. My work in the 2020s was building measurement and growth systems, which meant I was paid to find exactly these savings and present them as achievement. We were not cynics. We were doing the job the accounting asked of us. The line item for compute went up; the line item for salaries went down; the difference was called operating leverage and I wrote the slide.
What no one in those rooms said out loud — what I did not understand until my mother’s letter arrived — is that we were not just cutting a cost. We were cutting the only channel through which the value our company produced reached a household.
The value did not fall. The route home did.
What a household share looks like
The Ministry of Machine Yield does not report the labour share as its headline number, and that decision was contested for years. Reporting it would have been dishonest in both directions: it makes the present look catastrophic and it implies a target no technology could deliver. Most work in 2047 is genuinely performed by machines. The labour share is lower than at any point since records began, and it is not coming back.
What is reported instead is the household share: the portion of national income arriving in households by any channel at all, wage or dividend or funded care.
| Wage channel (pre-2035) | Dividend channel (2047) | |
|---|---|---|
| What triggers payment | Holding a job | Residency and majority |
| What is measured | Hours, seniority, negotiated rate | Metered machine output in the Machine Yield Account |
| Who is excluded | The unemployed, carers, the displaced, the old | Nobody within the covered population |
| Basis of the claim | Sale of labour | Inherited share in a commonly held asset |
| Failure mode | Wage share falls as automation rises | Yield mis-valuation; jurisdictional leakage |
| Political protection | Unions, labour law, minimum wage | Universality — every household is a claimant |
The table hides the fight. Universality was the expensive concession: paying the dividend to people who plainly did not need it, in order that no one would ever have to prove they did. The means-tested version polled better and would have died within a decade, because a payment only the poor receive is a payment everyone else eventually votes against. Alaska taught that lesson in miniature for sixty years before anyone applied it at scale — a dividend paid to all residents built a constituency that defended the fund through every fiscal crisis the state had.
And the claim underneath the payment is not need. It is authorship. The intelligence in these systems is a compression of the human record: the books, the papers, the forum answers, the repair manuals, the case law, my mother’s nineteen years of adjudicated claims. Add publicly funded research, publicly built networks, publicly educated engineers. That is an inherited asset, and the return on an inherited asset belongs to its heirs. Capacity is not contribution, and work that must be done to survive is not chosen work.
Tuesday
My mother sits with families on Tuesdays. Not employed, not volunteering in the charitable sense, just present — in a room at the health service where people arrive furious and frightened and holding paperwork they cannot parse. She is extraordinary at it. Nineteen years of reading claim files taught her which silence means grief and which means fraud, and that skill is now spent on people rather than on a loss ratio.
The Contribution Record notes that she does this. It confers nothing. It cannot be used to reduce her floor, condition her housing or shame her if she stops, and the year that protection was nearly removed is a story for another dispatch. She would tell you Tuesdays are the best work she has ever done, and that nobody would ever have paid her for it.
That is the sentence I would put on the wall of the building. There was a version of the last two decades in which she got the money and not the Tuesdays, and it would have been a poorer settlement than it looks on paper.
Where this still fails
Three admissions, because a design defended only by its friends is not a design.
The yield may be smaller than the rhetoric. Acemoglu’s estimate put total factor productivity gains from AI well under one percent over a decade — a fraction of what the industry promised. A dividend sized to the promise rather than to the meter would have collapsed in its first recession, and in two jurisdictions it nearly did.
Metering is jurisdictional and capability is not. Yield is recorded where deployment is booked, so the countries that host the compute collect, and the countries whose written record trained the models frequently do not. The inheritance is global; the ledger is national. This is the largest unresolved injustice in the system and it is not close.
And the floor pays rent, not purpose. My generation lost something in the transition that no schedule replaces: the crude, cruel, reliable way a job answered the question of who you were. My youngest has never needed that answer and cannot see the hole. My mother can. She filled it with Tuesdays. Not everyone finds a Tuesday.
FAQ
What is the labor share of income, and why does AI reduce it? It is the share of national income paid as wages rather than as profit, rent or interest. AI reduces it because a task moved from a paid person to a system continues producing output while ceasing to produce wage income; the saved cost is booked as margin, which is capital income.
Was the labor share already falling before AI? Yes — for roughly forty years, driven by cheaper capital goods, offshoring, industrial concentration and weakened bargaining institutions. AI did not start the decline; it removed the technological floor beneath it.
Why describe wages as a pipe rather than a fair reward? Because reward implies a choice someone made. Payroll was the mechanism by which produced value reached households, and it worked because production required people. Calling it virtue obscured how fragile it was.
Does the Automation Dividend restore the labor share? No. It restores the household share — income reaching households by any channel. The labour share stays low, because machines really do most of the work.
Why is the dividend framed as ownership rather than redistribution? Because the asset generating the income was inherited: the recorded human record plus publicly funded science and infrastructure. Collecting a return on something you co-own is not redistribution.
What is the strongest objection to this account? That measured AI productivity gains are modest, so the metered yield is smaller than either advocates or the Ministry initially assumed. The answer is procedural, not rhetorical: meter realised output, publish the valuation error, and never promise a dividend larger than the meter.