Ministry of AI · Dispatch from 2047
One Displacement Ledger Entry, Read Aloud
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’s job has a document number.
It is not a headline and it is not a memorial. It is a record in a public register, seven lines long, and my youngest son found it last spring while doing homework about the transition. He turned the screen towards me at dinner and asked whether this was her. I said it was not her. It was the shape she used to fill.
She read claim files for nineteen years. Home insurance, mostly: water damage, storm damage, the slow arguments about whether a thing broke or was allowed to break. She was good in a way that never showed up in a performance review — she could tell from the sequence of a claimant’s sentences whether they were lying or simply frightened, and she was right often enough that her employer eventually built a decision model on eleven years of her adjudications. The model went live in the mid-thirties. She was not fired. Her title was retained, her hours were reduced, and the two colleagues who left that year were not replaced. In the statistics of the time she appears as an employed person at a firm with stable headcount and improving margins.
That is the problem this dispatch is about. For roughly two decades, everyone argued about AI job displacement data using instruments that could only see the loud kind.
What the old instruments could see
The measurement stack of the 2020s was built for an earlier kind of shock: the plant closure, the mass redundancy, the event with a date. In the United States the sharpest instrument was the WARN notice, a filing that employers of a certain size owed the public before a mass layoff or plant closing — a genuinely useful record, and one that by construction begins at a threshold of sixty days and fifty jobs (U.S. Department of Labor). Below the threshold, nothing. Beside the threshold — attrition, reduced hours, a hiring freeze in one function while another expands — nothing.
The academic instruments were better at scope and worse at causation. Frey and Osborne’s exposure estimate, the one everybody remembers as a number rather than as a method, described how much of the occupational structure was technically susceptible to computerisation, not how many people would be let go. The later generative-AI work did the same job for language models: a task-level exposure map, honest about being a map of susceptibility (Eloundou et al., 2023). The ILO’s global analysis reached the conclusion that mattered most and pleased nobody: the dominant effect of generative AI would be augmentation rather than elimination — most exposed jobs would change composition, not disappear (ILO).
Augmentation, it turned out, was not the good news it was read as. It meant displacement would arrive as a change in the content of work rather than the count of workers, and the count of workers was the only thing being tallied. And so we got two decades of a maddening public argument in which the data were, in a narrow sense, correct on both sides. Firm-level studies of early adoption found remarkably small effects on earnings and hours — a Danish study covering thousands of workplaces put the average earnings and hours effect at under two percent, which the sceptics quoted as proof that nothing was happening (Humlum & Vestergaard, NBER w33777). Meanwhile the same period’s macro accounting made the direction obvious: an AI-driven fall in unit labour cost shows up as margin, and margin is capital income (Philadelphia Fed). Both were true. The wage bill was falling relative to output while almost nobody was being marched out of a building.
You cannot tax, charge or meter a thing you refuse to record. The Ministry of Machine Yield exists because someone finally changed the unit of observation.
The unit change
| Layoff-era measurement | Displacement Ledger | |
|---|---|---|
| Unit of record | The terminated person | The absorbed function |
| Trigger | A filing threshold or benefit claim | A system carrying a share of a function |
| Timing | After the event | Continuous, per reporting period |
| Sees quiet absorption | No | Yes |
| Records where capability came from | No | Provenance class per entry |
| Links to money | Indirect and contested | Yield attributed, with error band |
| Names individuals | Often | Never |
The mid-2020s operators who first argued for this were not policy people. They were the ones running coordinated AI systems inside real companies, which made them the first cohort to watch the human/machine boundary move under their own payroll — and to notice that their internal dashboards knew precisely what had been absorbed while every public statistic knew nothing. I was one of them. I built measurement systems for a living and I could tell you, to the task, which parts of a colleague’s job my systems had taken. That knowledge existed. It was simply proprietary.
The entry itself
Here is the structure, using an in-world example — the Ministry’s own reporting format, not a real-world statistic. Six fields, plus the error band.
Entry 41-DA-0937. Absorbed function: first-pass adjudication of residential property claims, including credibility assessment from claimant narrative. Capacity share: 0.82 of the function as performed in 2031. Residual human task: contested-claim review, bereavement cases, and any file the system flags as unlike its training. Provenance class: II — proprietary records of prior human adjudications, corpus includes the recorded decisions of the displaced cohort. Yield attributed, period: metered against the operator’s Machine Yield Account. Declared valuation error: ±19%.
Read that provenance line again, because it is the whole moral argument in a bureaucratic sentence. The system learned the function from the people who performed it. My mother’s judgment is inside the thing that replaced her judgment, and it got there without a purchase, a licence or a negotiation. Class II entries were the ones that ended the public debate about whether the dividend was welfare. Welfare is what you give a person who has nothing. A dividend is what you owe a person whose asset is earning without them.
The Displacement Ledger does not pay her, which is the part people still find counterintuitive. It records what was absorbed; the Machine Yield Account measures what the absorption produces; the Dividend Schedule routes the charge to the floor, healthcare, housing, care and education for everyone alive. There is no royalty on your old job title. If there were, we would have built a system that pays the well-documented and skips the rest — a nurse’s improvised workaround has no filing, and the entire corpus is thick with such things.
The Tuesday it bought
She is seventy-three. On Tuesdays she sits in a small room at the health service with families who are arguing about a decision, and she reads the file, and she tells them in plain sentences what it actually says and where they might push. Nobody pays her. Nobody has ever assigned her a target. She is doing the work she was best at, for the people who most needed it, without the sentence that used to hang over it — do this or lose the house.
Capacity is not contribution, and work that must be done to survive is not chosen work. That is the sentence on the building. It is also just a description of my mother on a Tuesday.
Where this design fails
The ledger is a compromise dressed as a record. Three failures we have not solved:
Attribution is contested and always will be. When output rises after a deployment, some of the gain is the system, some is the process redesign the deployment forced, and some is demand that would have arrived anyway. The ±19% in that entry is not modesty; it is the real state of knowledge. Publishing the error band is the only reason the numbers survived their first political challenge, and it remains the easiest thing to attack them with.
Anonymisation cost us accountability. Removing employer and team identifiers protected people in small towns from being identified as the absorbed cohort. It also made it much harder to say this firm, in this year, absorbed this much and declared none of it. We chose dignity over enforcement leverage and I still do not know if that was right.
Quiet absorption is still under-recorded. The ledger catches far more than a layoff filing ever did, but a function that erodes across four teams over six years, with no single deployment to point at, often lands as a partial entry or none. The earliest work on automation and employment warned that the mechanism was gradual and heterogeneous, not a cliff (Bessen, NBER w24235). Gradual is exactly what a register struggles to catch, and the hours data that eventually vindicated the whole project — the long secular decline in working hours across two centuries (Our World in Data) — is only legible in retrospect, when it is too late to charge anyone.
My son asked, when we had finished reading the entry, whether Grandma minded being in a document. I asked her later. She said the document was not the insult. The insult had been the years when there was no document at all, when the smartest thing she ever did was quietly absorbed into an asset, and every official number said nothing had happened.
FAQ
What is AI job displacement data? In the 2020s it meant counts of people who lost jobs and attributed the loss to automation: layoff filings, survey answers, unemployment claims. In 2047 it means a record of which human functions autonomous systems now perform, what share of each function they carry, and what yield that generated. The unit of observation moved from the person to the task, which is the only reason the quiet cases became visible.
Why did layoff statistics undercount AI displacement? Because most absorption never produced a termination. Teams kept their headcount and stopped backfilling; hours were trimmed; a function was reassigned to a system while the job title survived. Filing thresholds like the WARN notice are triggered by mass layoffs at larger employers, so anything gradual, small or internal passed under them without a trace.
What is in a Displacement Ledger entry? Absorbed function, capacity share, residual human task, provenance class of the learning material, yield attributed for the period, and the Ministry’s declared valuation error. No names, no team sizes, no employer identifiers.
Does an entry entitle the displaced worker to a payment? No. Entries determine what was absorbed and what it yielded. The Dividend Schedule then routes the charge to universal instruments — income floor, healthcare, housing, care, education. Paying per absorbed job title would reward documentation rather than contribution, and most contribution was never documented.
Isn’t a public register of absorbed jobs humiliating? The first drafts were. They carried enough detail to identify individuals in small labour markets. Entries now describe functions rather than people and publish only at non-reversible aggregation, at a real cost to enforcement.
How reliable are the figures? Reliable enough to charge on, and far less exact than they appear. Separating the effect of a system from the process redesign around it is unresolved; every entry carries an error band, and the honest ones are wide.