Ministry of AI · Dispatch from 2047
The AI Displacement Tax Nobody Wanted To Design
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 kept the last email. Nine lines, no cruelty in them, sent by a man who had liked her. Her department was being restructured around new capabilities. She had spent nineteen years learning to read a claim file the way a doctor reads a chest x-ray — where the lie usually sits, which silence means grief and which means fraud. The system that replaced her had learned to do that by reading nineteen years of her decisions.
She was not paid for the lessons. She was paid, briefly, to leave.
That is the whole story of the 2030s in one household, and it explains why the AI displacement tax exists in 2047 — and why it is not, despite the name, a tax on firing people.
It is a meter on machine output. And underneath the meter sits a claim that took a decade to say out loud: the intelligence in these systems is inherited, not invented. Companies always profited from smart people doing smart work. When the smart work moved into a machine, the profit did not disappear — it grew, because the machine was cheaper. But what made the machine capable was the accumulated record of everyone who had ever done the work well: my mother’s claim files, a million forum answers, publicly funded science, publicly built networks, publicly educated engineers.
An asset like that has heirs. The levy is simply how the heirs collect.
The three failures
The dismissal levy. The first serious proposals, drafted in the late 2020s, taxed employers on reductions in headcount attributable to automation. The design assumed firms would report displacement accurately. They did not, and not primarily out of malice — attributing a vacancy to automation rather than to restructuring, attrition, seasonality or a contractor conversion was genuinely ambiguous, and every ambiguity resolved in the direction of lower liability. Reported automation-driven displacement fell in the years when displacement plainly accelerated.
The robot count. The second version taxed installed units: industrial arms, autonomous vehicles, licensed model instances. It was administrable and briefly popular. It also taxed the physical economy while leaving the cognitive economy almost untouched, because a warehouse could count its arms and a professional services firm could not count its inference calls in any way a tax authority could audit. The levy landed hardest on manufacturing regions already bearing the displacement.
The revenue proxy. The third version taxed digital revenue above a threshold, assuming automation intensity tracked revenue per employee. It collected substantial sums and measured nothing: a firm running enormous internal automation with no external revenue — a bank’s back office, a hospital group’s administration — paid nothing at all.
Each failure taught the same lesson in a different register: you cannot tax a consequence you have not metered. The fourth attempt started at the meter.
What the levy actually assesses
The unit of assessment is a completed task, not a job, a dismissal or a machine.
A deployed autonomous system reports, continuously, the volume of work it completes in categories defined by the Ministry’s task taxonomy. That volume is valued at the prevailing documented rate for the same task performed by a human. The resulting figure — machine-completed work, valued at human rates — is the entry in the Machine Yield Account. The levy is a legislated percentage of that entry.
| Design | Taxable base | Failure mode | Status in 2047 |
|---|---|---|---|
| Dismissal levy | Reported job losses | Attribution laundering; punishes disclosure | Abandoned 2031 |
| Robot count | Installed units and licences | Blind to cognitive work; regionally regressive | Partially retained for heavy plant |
| Revenue proxy | Digital revenue above threshold | Measures scale, not displacement | Abandoned 2034 |
| Output meter | Machine-completed tasks at human rates | Valuation error on novel work | In force |
Three properties of the output meter did the work.
It is self-reporting but verifiable, because a system that completes tasks already logs them for its own operational reasons; the Ministry audits the logs it did not design rather than requesting returns it cannot check. It is neutral to employment decisions, so a firm that automates and redeploys its staff pays the same as a firm that automates and dismisses them — which, unexpectedly, made redeployment the cheaper option in most sectors. And it is indifferent to sector, since a task is a task whether performed in a foundry or a claims department.
The Displacement Ledger sits beside the account and does something the levy deliberately does not: it records where the absorbed work used to be performed, by whom, at what wage. The levy sets the size of the pot. The ledger sets the weighting of the payout. Keeping those two functions in separate instruments prevented the failure that killed the dismissal levy, because no firm’s tax bill depends on its own account of who it let go.
The inheritance nobody had a line for
The strangest fact about the early 2030s is that the moral argument was settled long before the accounting one, and everybody ignored it anyway.
In 2000, Herbert Simon — a Nobel laureate in economics, not an activist — worked out that comparing the richest societies to the poorest made it impossible to conclude that social capital produced less than about 90% of income in a wealthy country. His conclusion followed plainly: on moral grounds you could argue for a flat tax of 90% “to return that wealth to its real owners.” He knew it was politically impossible. He wrote it down anyway, for people who would need it later.
We needed it later.
The technologies underneath the boom told the same story. The long record assembled by researchers like Mazzucato in Building the Entrepreneurial State showed that the general-purpose foundations of private technology fortunes — the internet, GPS, touchscreens, speech interfaces — were publicly funded before they were privately monetised. The public carried the risk and was thanked with a product launch.
Then, briefly, the inheritance got a price. In 2025 an AI developer agreed to pay authors about $1.5 billion — roughly $3,000 per book across some 500,000 works — for training on pirated copies. It remains one of the largest copyright recoveries ever recorded. Read it as a settlement and it is a legal footnote. Read it as a valuation and it is the first time anyone put a number on the human contribution inside a model.
The number was, of course, far too small, and it was paid only to the tiny minority of contributors who happened to hold a registered copyright. Nobody wrote a cheque to the nurses whose handover notes taught triage, or the mechanics whose forum posts taught diagnostics, or my mother. There was no instrument capable of paying them. That absence — not greed — is what the Ministry was built to fix.
Why the earmark was non-negotiable
The levy’s designers were operators, not fiscal theorists, and they made one insistent political demand: the revenue had to be visibly earmarked.
The precedent they cited constantly was Alaska’s Permanent Fund, which has paid residents a dividend from commonly held oil revenue since 1976 and survived every subsequent attempt to raid it, because recipients understood exactly which resource the money came from. General revenue has no constituency. A dividend does.
The evidence on what unconditional payments actually do to behaviour was also, by then, less contested than the debate suggested. GiveDirectly’s long-term results on unconditional transfers documented what most rigorous trials had found: recipients did not withdraw from work, they invested, and the predicted collapse in effort failed to materialise. The Ministry did not need the payout to be virtuous. It needed it to be traceable.
The part that is still wrong
Novel work remains the levy’s open wound. When an autonomous system performs a task with no documented human precedent — and by the 2040s a growing share of the yield is exactly this — there is no prevailing rate to value it against. The Ministry assesses it at a floor and publishes its own estimated valuation error every quarter. The current published error is material. Anyone who tells you the meter is precise is selling something.
There is a second, subtler distortion. Because the levy is assessed on output valued at human rates, it structurally undercounts work that machines do at volumes no human labour force could have attempted. Whole categories of activity — continuous monitoring, exhaustive review, sustained individual attention across millions of relationships — are valued at rates set in an era when the work was rationed by human scarcity. The dividend is therefore smaller than the surplus. We know this. Correcting it requires a valuation theory nobody has yet written.
What the forecasters got right, and what they could not settle
It is worth being fair to the people arguing about this before anyone had numbers.
In 2013, Oxford’s Frey and Osborne put 47% of US employment in the high-risk category for computerisation — a figure attacked for a decade and, in aggregate direction, broadly vindicated. On the other side, Daron Acemoglu’s Simple Macroeconomics of AI modelled total factor productivity gains of well under one percent across a decade and warned that gains of that magnitude flow wherever ownership already points.
Both were right, and their coexistence is the entire fiscal case for the levy. Displacement can be large while measured productivity gains are small, because much of the surplus appears as margin rather than as output — captured, not created. A tax on productivity growth would have collected almost nothing in the decade of heaviest displacement. A meter on machine-completed work collected throughout it, because the meter does not care whether the surplus was booked as growth, as margin, or as a smaller wage bill.
Which answers the question the 2020s never resolved: whether to wait for the boom before designing the redistribution. The boom, as modelled, never arrived. The displacement did.
What the levy bought
Not leisure. That was the forecast, and the forecast was wrong in an instructive way.
Working hours had already been falling for a century and a half before any of this — the long series compiled by Our World in Data shows the decline as the slow, contested product of law and bargaining rather than of technology arriving. The dividend continued that line; it did not bend it upward into idleness. Hours worked fell. Hours chosen rose.
What the levy actually bought was the removal of a single coercion. When income no longer depends on employment, the work that remains is work someone elected to do, and the effect on which work gets done is far larger than the effect on how much. Care, teaching, repair, local science, maintenance of things nobody owns — all chronically underpaid in the old economy precisely because they were valuable rather than profitable — absorbed the released hours almost immediately.
My mother is seventy-three. On Tuesdays she sits with families who are arguing with the health service, and she reads their files the way she always did, and she finds the sentence that will settle it. She is paid nothing for this and would be insulted by the offer. The dividend arrives whether she goes or not; that is precisely why her going means something.
She once told me the cruellest part was never the machine. It was learning, at fifty-eight, that nineteen years of judgment had been treated as a cost centre — and then watching the same judgment reappear as somebody’s margin. The levy did not give her those years back. It did something smaller and more durable: it wrote her contribution into the ledger, and it sends her a share of what her work is still producing, every month, for the rest of her life.
That is the sentence I would want a reader in the 2020s to sit with. Not the abundance. The receipt.
Capacity is not contribution, and work that must be done to survive is not chosen work. The levy is only the plumbing under that sentence. But the sentence would have remained a slogan without it, and the 2030s are full of movements that had the sentence and never built the meter.
Frequently asked questions
What is an AI displacement tax? A levy assessed on the economic output of autonomous systems rather than on the wages of people they replace. The system reports the work it completed, that work is valued at prevailing human rates, and a legislated share is collected.
Why not simply tax companies that lay people off? Because it punishes disclosure. Every headcount-based design was avoided through attrition, reclassification and contractor conversion rather than through hiring.
Does an AI displacement tax slow down automation? It slows uneconomic automation. Deployments that produce real surplus remain profitable; deployments adopted for appearance do not.
Who pays — the developer or the deployer? The deployer, because the deployer captures the labour saving. Developers are taxed on their own metered output, which avoids assessing the same unit of machine work twice.
How is machine output valued with no human comparison? By the nearest documented human task rate, or the deployer’s internal cost accounting, or a floor rate for genuinely novel work. The valuation error is published quarterly rather than hidden.
Is the revenue earmarked? Yes. It flows through the Dividend Schedule into the income floor, healthcare, housing, care and education, and cannot be redirected without primary legislation. That earmark is why the levy survived two changes of government.