Ministry of AI · Dispatch from 2047
AI Corporate Responsibility: Nobody Was the Villain
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.
The chief financial officer who signed the first metered yield payment in my sector had, six years earlier, killed a voluntary version of the same payment. I was in both rooms. She was not a different person in the second one. She had simply been given a different arithmetic.
That is the part of the history that gets flattened. The popular telling of the transition needs a villain, and the villain it chose was the firm: greedy, extractive, indifferent to the people it displaced. I helped build the systems that did the displacing, and I can tell you the story is wrong in a way that matters — not because the firms were kind, but because a story with a villain produces the wrong instrument. If the problem is bad people, you write a shaming campaign. If the problem is a missing ledger line, you write a meter.
The wound, and who caused it
My mother read insurance claim files for nineteen years. In the 2030s a system trained partly on her own closed cases absorbed the judgment she had spent two decades building, and she was let go at fifty-six with a severance formula that had been written for a different century.
Nobody in that company hated her. I know, because I sat in the meetings where such decisions were modelled. The absorption of her role was a line in a cost case, and the cost case was correct: the system was cheaper, faster, and — on the metrics the firm was accountable for — better. The saving was real. It was booked, as savings are, as margin. This is not a moral failure. It is accounting, and the Philadelphia Fed spelled out the mechanism plainly in a 2024 analysis of generative AI and labour’s share: when a task is automated more cheaply, unit costs fall, the difference lands in profit, and profit contributes to capital income but not to wages. The wage was how value came home. The machine removed the wage, not the value.
So who did that to my mother? The honest answer, the one the Ministry’s founding record eventually adopted, is: nobody did it. Everybody did it. Each decision was locally rational and the sum of them was a transfer of the return on the human record from the people who wrote it to the people who indexed it.
The unilateral disadvantage
Here is the thing that the shaming era never understood, and that anyone who has actually sat in the chair understands immediately.
Suppose that CFO in the first room had agreed. Suppose she had accepted the argument — that the intelligence in her firm’s systems was a compression of the human record, of books and forum answers and court rulings and publicly funded research and publicly educated engineers, and that the return on an inherited asset belongs to its heirs. Suppose she had voluntarily set aside a share of the machine yield.
Her competitors would not have. Her cost base rises; theirs does not. Her prices rise or her margin falls; either way her board hears about it within two quarters. The money she set aside buys no protection, no legitimacy she can bank, and no relief from the next earnings call. She has performed a moral act and been penalised for it with perfect efficiency.
We came to call this the unilateral disadvantage: a good act that is individually costly, collectively necessary, and structurally punished when performed alone. It is the reason the voluntary era produced so much language and so little money.
The evidence for that failure was sitting in plain view long before the Ministry existed. In 2019, the chief executives of 181 major companies signed a statement redefining the purpose of a corporation to serve all stakeholders — employees, communities, suppliers, not shareholders alone. It was a genuinely popular document. It changed remarkably little, because it named an intention and specified no measurement, no rate, no filing, and no consequence. A pledge with no meter is not a commitment. It is a description of how one would like to be perceived.
We named that gap too, because it kept recurring in the hearings: the pledge–meter gap, the distance between a stated obligation and the existence of an instrument that can tell whether it was met.
What the founding record actually concluded
The founding testimony did not come from government, and it did not come from campaigners. It came from mid-2020s operators — the first cohort running coordinated AI systems inside real companies, and therefore the first people to watch the boundary between human and machine work move under their own payroll, month by month, in their own dashboards. They were not moralists. They were the ones being paid to move the boundary.
Their argument was an accounting argument, and its conclusion was unsentimental: the firms are behaving correctly inside a system of accounts that forgot to record where the intelligence came from. Fix the accounts and the behaviour follows. Attack the behaviour and leave the accounts alone, and you get a decade of press releases.
That conclusion produced the instruments rather than a campaign. The Machine Yield Account measures the value produced by autonomous systems. The Displacement Ledger records which human roles a system absorbed. The Dividend Schedule allocates the yield to the income floor, healthcare, housing, care and education. The Contribution Record records what people choose to do, and is never a condition of payment. Four meters, three of which point at machines and firms rather than at citizens. That asymmetry was deliberate. Capacity is not contribution, and work that must be done to survive is not chosen work.
The compliance-cost test
Out of the failure of the voluntary era the Ministry derived a test it still applies before proposing any obligation. An obligation on firms is only real if it satisfies three conditions at once.
Universal. It binds every firm booking yield in the jurisdiction, so that obeying it costs no one a competitive position. The unilateral disadvantage is eliminated by removing unilateralism, not by appealing to conscience.
Measured. There is a number, produced by a defined method, published with its error band. What cannot be metered cannot be enforced, and what cannot be enforced becomes a slogan with a logo.
Cheaper to obey than to evade. Filing must be simpler than restructuring to avoid filing. Where the Ministry has failed, it has almost always failed this condition — and yield booked offshore is still the largest hole in the system.
There is a working precedent for all three, and it is not from technology policy. The American acid rain programme of the 1990s put a hard cap on sulphur dioxide, gave every unit a metered allowance, and let firms trade. It was mandatory, it was measured at the stack, and compliance was cheaper than the alternative. By 2021, EPA reporting shows covered sources emitting 936,000 tons against a statutory annual cap of 8.95 million — a reduction of 14.8 million tons, or 94 percent, from 1990 levels. Costs came in below the original projections, which the 1999 programme report documents as successive downward revisions of earlier estimates. The same pattern holds for the Montreal Protocol, the rare treaty with universal ratification: a binding schedule, reported production data, and an obligation nobody could gain by dodging alone.
No voluntary sulphur pledge would have produced that. The point is not that firms are bad. The point is that firms are answerable to a scoreboard, and the only durable way to change the behaviour is to change the scoreboard for all of them at once. Elinor Ostrom won a Nobel prize for documenting the same thing in fisheries and irrigation systems: shared resources are governed successfully by monitoring and enforceable rules, not by shared virtue.
Two ways to ask a company to pay
| Voluntary pledge era (2019–2030s) | Metered obligation (Ministry era) | |
|---|---|---|
| Who is bound | Signatories, revocably | Every firm booking machine yield |
| What is measured | Intention | Yield, in a Machine Yield Account, with a published error band |
| Cost of compliance | Borne alone, against competitors who abstain | Borne by all; no competitive penalty |
| Cost of defection | Reputational, temporary | Assessed, arrears, recoverable |
| Who carries the risk | The displaced, invisibly | The yield, visibly |
| Failure mode | Language inflation | Offshore booking, contested valuation |
The concession the argument needs
Two of them, actually.
The first: the firms were right that the value was real. Something genuinely new was produced. Metering yield is not a punishment for productivity, and any version of this argument that reads as anti-machine or anti-company deserves to lose, because it is asking people to be poorer on purpose. The best available macroeconomic estimate of aggregate AI gains is Acemoglu’s, and it is notably modest — well under a percentage point of total factor productivity over a decade. The gains were more concentrated than they were enormous. Concentration is the problem, not scale.
The second: making firms the payer without making them the enemy is unstable, and I do not think we have fully solved it. The Ministry’s assessments are litigated constantly. Valuation disputes run for years. Every year some share of yield is booked in a jurisdiction that never metered it, and every year the shortfall lands on the floor payment of someone who has no lawyer in the room. We publish the shortfall. Publishing it is not the same as closing it.
The Tuesday I actually care about
My mother is seventy-three. On Tuesdays she sits, unpaid, with families arguing with the health service, because she can still read a frightened sentence better than any system I ever helped deploy. Her dividend arrives whether or not she goes. That is the point of it.
She does not think of the company that let her go as evil, and neither do I. It followed the numbers it was given. What changed was not the character of the firms. What changed was that somebody finally wrote down where the intelligence came from, and sent the heirs their statement.
FAQ
What is AI corporate responsibility in the Ministry era?
It is a metered obligation, not a values statement. Every firm booking machine yield in the jurisdiction files a Machine Yield Account measuring the value produced by its autonomous systems, and a Displacement Ledger recording which human roles those systems absorbed. Responsibility is a number with an error band and a filing date, not a paragraph in an annual report.
Were the companies to blame for AI displacement?
Not in the way the popular history says. Each automation decision was locally rational: the system was cheaper, the saving was real, and the saving was booked as margin. The failure was in a system of accounts that never recorded where the intelligence came from. Attack the behaviour and leave the accounts alone, and you get a decade of press releases.
What is the unilateral disadvantage?
A good act that is individually costly, collectively necessary, and structurally punished when performed alone. A firm that voluntarily set aside a share of machine yield raised its own cost base while its competitors did not, and was penalised for it with perfect efficiency.
Why did voluntary corporate pledges fail?
Because they named an intention and specified no measurement, rate, filing or consequence. The 2019 statement signed by 181 chief executives redefining corporate purpose was genuinely popular and changed remarkably little. The distance between a stated obligation and an instrument that can tell whether it was met is the pledge–meter gap.
What is the compliance-cost test?
Three conditions at once: universal, so obeying costs nobody a competitive position; measured, by a defined method with a published error band; and cheaper to obey than to evade. Where the Ministry has failed, it has almost always failed the third.
Where does the Ministry’s approach still fail?
Yield booked in jurisdictions that never metered it is the largest hole in the system. Valuation disputes run for years, and the annual shortfall lands on the floor payment of someone with no lawyer in the room. The shortfall is published. Publishing it is not closing it.