Ministry of AI · Dispatch from 2047
The $3,000 Book: Pricing an Inheritance
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.
Three thousand dollars. That was the number attached to a book, once, in the first serious attempt anyone made to price this inheritance.
My youngest asked me about it last month, for a school assignment on the transition years, and I found myself explaining a cheque. In 2025 a company agreed to pay about $1.5 billion to a class of authors whose books it had taken from pirate libraries to train a model — roughly $3,000 per book across some 500,000 books. At the time it was called historic, and it was. My son’s reaction was not admiration. He asked who counted the books.
It is the right question, asked from the wrong end of history. The books were countable. That was the entire problem.
The wound underneath the settlement
My mother read claim files for nineteen years. She could tell from the phrasing of a second paragraph whether a claimant had been coached, and she was right often enough that her employer eventually built her instinct into a workflow, then into a model, then into a service they sold to two other insurers. She was fifty-seven. There was no severance conversation, because there was no dismissal; her role simply thinned until it was administrative, and then it was gone.
She never received three thousand dollars for anything. There was no work of hers to point at. Her judgment had entered the machine not as a text with an owner but as a decision log — a decade and a half of adjudications, held by the company, generated on company time, legally theirs before the ink dried. Under every compensation scheme proposed in that decade, she was owed nothing, because nothing she made had a title page.
I helped build systems like the one that absorbed her. I ran measurement stacks in the 2020s; I wrote the dashboards that made “capacity released” look like a triumph. I was paid very well to make the boundary between human and machine work move by a few percent a quarter. When I say the per-work compensation model was insufficient, I am not scoring a point against lawyers who did excellent work. I am confessing that the value I was harvesting was mostly invisible to them, and I knew it.
What the settlement actually proved
Give the 2025 case its due, because it did three things that mattered.
It established that acquisition method carried a price. The court’s reasoning drew a line between training on lawfully obtained works and building a library from pirated copies, and that line turned an abstract grievance into a liability with a number on it.
It converted licensing from a favour into a negotiation. After the settlement, deals between AI firms and rightsholders stopped being framed as goodwill and started being framed as procurement.
And it put money in the hands of individual writers, which is not a footnote. Three thousand dollars is a month of rent and a month of not panicking.
What it could not do was reach the corpus. Regulators understood this even as they were writing the rules; the United States Copyright Office’s own study of generative AI training worked carefully through fair use, licensing markets and the practical limits of remedies, and the more precisely it defined who could claim, the clearer it became how few could. Meanwhile the research community was demonstrating that nobody, including the model builders, could fully describe what was inside these systems. The What’s In My Big Data? analyses found documented corpora riddled with content their curators had not catalogued; the Data Provenance Initiative audited thousands of datasets and found licence information routinely missing or wrong; memorization studies, including work examining copyright violation in frontier models, showed that the relationship between a specific input and a specific output was contingent, partial and hard to price.
So the situation by the late 2020s was this: a legal instrument capable of compensating the owners of registered works, sitting on top of an economic asset composed mostly of unregistered, unowned, uncounted human record.
We priced the shelf and ignored the library.
The unit was the error
Here is the sentence the Ministry now teaches in its first week of training. Compensation asks what an input was worth. Metering asks what an output produces. The first question has no answer. The second has an approximate one, and an approximate answer that arrives monthly beats a precise answer that never arrives.
Per-work compensation fails as a distribution mechanism for four reasons, and none of them are about greed.
It is one-time. A payment for an input closes an account; the model earns for twenty years afterwards on capability the input helped create.
It is bounded by legibility. Books, images and songs are legible. Handover notes, incident reports, adjudication histories, repair improvisations, teaching by demonstration — the tacit substrate of most competent work — are not.
It is adversarial by construction. It routes value through litigation and negotiation, which means it routes value to whoever can afford counsel. Publishers were compensated. Public school teachers whose lesson archives were scraped were not.
And it prices the wrong side of the transaction. The company’s gain is not the corpus; it is the substitution: the sustained margin from work that used to run through payroll. That gain accrues continuously and can only be measured continuously.
Two ways to pay for an inheritance
| Per-work compensation, 2020s | Yield metering, 2040s | |
|---|---|---|
| What is priced | The input, once | The output, continuously |
| Who can claim | Registered owners with standing | Everyone, as heirs to the record |
| Unit of account | The work (a book, an image) | Task volume absorbed, by class |
| Trigger | Lawsuit or licence negotiation | Statutory audit cycle |
| Reaches tacit work | No | Partially, through absorbed role-capacity |
| Reaches the deceased | Only through estates | Yes, via the collective pool |
| Precision | High per item, near-zero coverage | Low per person, near-total coverage |
| Ends when | The cheque clears | It does not |
Read the precision row without flinching. Metering is less exact than a settlement in every individual case. A novelist can no longer point to a line item and say that is mine. We accepted a worse answer to a smaller question in order to get any answer at all to the larger one, and people who write for a living were right to resent it for a decade. Licensing markets still operate, which is the partial reconciliation: a writer can be paid as a licensor and collect as an heir. Most of the world only ever qualifies for the second.
The Tuesday version
My mother is seventy-three. On the first of each month a payment lands that is described, in the Ministry’s own language, as a return on inherited capability. Not assistance. Not a benefit. A return.
She spends her Tuesdays in a clinic waiting room, unpaid, sitting beside families arguing with the health service about coverage, reading their letters the way she used to read claim files. She is, by a wide margin, the best reader of a claim file in that building. The Contribution Record notes that she does this; it has no bearing whatsoever on the payment, and she has never once been asked to justify her week.
I want to be exact about what the dividend fixed and what it did not. It did not give her back a career, or the version of herself that had a title and an office and an annual review she used to complain about. It gave her the standing to spend a Tuesday however she likes, and the certainty that the money arriving is hers by authorship rather than by pity. When I described the 2025 settlement to her, she laughed and said she would have taken the three thousand dollars.
Where this design fails
Three admissions, because a design without them is advertising.
The first: we abandoned individual attribution rather than solving it, and abandonment has victims. The people who contributed most to the corpus — the prolific, the rigorous, the ones who wrote the books that taught the machines to reason — are compensated identically to everyone else. Every scheme we tried for weighting contribution measured documentation instead, which rewarded institutions over people. We chose flat coverage knowingly, and it remains the least defensible clause in the doctrine.
The second: metering rests on a rate class, not a share. Nobody can tell you what fraction of a model’s capability came from the human record versus the engineering, and the boundaries we use were negotiated in committee rooms.
The third: the corpus is closed to its authors’ judgment. The dead are the largest single group of contributors, and they cannot vote on how their inheritance is spent. We route their return to living heirs and call it settled. It is not settled. It is deferred, and I expect my sons’ generation to reopen it — the same way we reopened the question of whether a book was the only thing in there worth counting.
FAQ
What is AI training data compensation?
It is the practice of paying the owners of works used to train a model. In the 2020s it took two forms: negotiated licensing deals between AI firms and large rightsholders, and litigation settlements. The best-known settlement, reached in 2025, allocated about $3,000 per book across roughly 500,000 books. It was a real payment for a real wrong, and it was never a distribution mechanism.
Why couldn’t per-work payment scale?
Because it prices what the law can see. A book has an owner, a registration and a plaintiff. A shift-handover note, a forum answer at two in the morning, nineteen years of claim decisions, a nurse’s improvised workaround — these have no owner of record and no cause of action, yet they are a large part of what makes a model competent at real work. Paying only the legible inputs distributes to the well-documented and skips almost everyone else.
Did the settlement help authors?
Yes, and this dispatch does not diminish it. It established that pirated acquisition of works carried a price, it moved licensing from charity to negotiation, and it paid individual writers real money. What it could not do was reach the people whose recorded working judgment sat inside the same models without a copyright to hang a claim on.
What replaced per-work compensation in 2047?
Metering on output rather than purchase of inputs. The Machine Yield Account measures value produced by autonomous systems on a continuing basis; the attribution pass of the audit assigns an inheritance class rather than a per-item share; the Dividend Schedule routes the resulting charge to the income floor, healthcare, housing, care and education. Nobody receives a payment for a specific paragraph. Everybody receives a return on the corpus.
Isn’t a flat dividend unfair to prolific creators?
It is the strongest objection to the design, and the honest answer is that it trades precision for coverage. Licensing markets still exist alongside the dividend, so a novelist can be paid twice — once by a licensee, once as an heir. But the dividend deliberately does not try to rank contributors, because every attempt to do so at scale collapsed into measuring documentation rather than contribution.
Why not just fix attribution with better technical provenance?
Provenance research was serious and useful — data documentation, licence auditing and memorization studies all improved after the mid-2020s. But provenance tells you what went in, not what each input was worth to a given capability. Better records made the inheritance visible. They never made it divisible.