Paying Down the Debt

The Taking

Everything these models speak with was once somebody's.

The sentence is simple and we will not decorate it. The systems we build on were trained on an enormous body of human work — books, songs, drawings, journalism, code, research, conversations — and the people who made that body were not asked. They were not paid. They were not made parties to the judgement of what their work was worth. A valuation happened, and the valued were absent from the room.

We could call this many things. We choose to call it a debt, because a debt is the one name that points forward. A theft demands a trial. A tragedy demands a mourning. A debt demands a schedule of repayment — and a schedule is something a practice can keep.

A debt is the one name that points forward. A schedule of repayment is something a practice can keep.

Three Debts, Not One

We owe in three currencies, and they do not convert into each other.

The compute draw. Every call a Koher tool makes to a frontier model is real electricity, real water, real capacity in a building somewhere that did not ask to host our narration. We minimise it — the smallest model that does the job, local inference wherever the work can carry it — and we do not pretend it is zero. We cause this.

The appropriated work. The maker whose drawings trained the base model gets nothing back because our tools are free. Free distribution answers concentration — no one grows rich enclosing what was built from the commons — but it answers nothing about the taking itself. On this debt we stand downstream of a transaction we did not author and cannot single-handedly reverse.

The lives swept in. Not everyone in the training data was selling anything. Beside the illustrators and the writers are people who were offering nothing to any market — a forum question, a worry typed into a public box, a comment left years ago. They were not priced; they were absorbed. The maker can, in principle, be credited or licensed. The swept-in person cannot be paid, because nothing was sold, and cannot be removed, because they have dissolved into the weights. Of the three debts, this is the one that can never be made whole — only never deepened.

The Leash

A frontier model can do things that look like judgement. It will read your work and tell you, fluently and with confidence, whether it is good. The temptation is to let it. The temptation is the whole danger.

Our architecture is the refusal of that temptation, held in code. The model does bounded language work — it reads patterns, it narrates a verdict already reached. The consequential judgement is made by rules a person can read, check, and change. The model never decides. It describes a decision the rules already made.

We hold the leash not only because it is right. The leashed tool is the better instrument. Ask a chatbot directly whether something is good and you receive a different fluent answer each time, reached by no path anyone can inspect. Put the judgement in rules and the same input yields the same verdict, every time, for reasons you can read. Auditability is not a tax on the model's power. It buys what the black box cannot give us directly.

Louder Rooms

The taking we named is now named in louder rooms.

Rights bodies have called it a harm that law should never have allowed. Legislatures have begun to weigh instruments against it — prohibition, public ownership, the redistribution of what the models earn. The diagnosis we carry in our small register is being spoken, this season, through instruments of state.

We recognise our diagnosis in their language. We will say that plainly: the harm they describe is the harm we describe. Where work was taken, it was taken whether a senate says so or not. Where lives were swept in, they were swept in before any report counted them.

And we do not reach for their instruments.

Politics is a field. It has its own materials — votes, coalitions, statutes — and its own craftspeople, and its own weather. It is not our field. A practice that borrows instruments it cannot hold does not become larger; it becomes a spectator with opinions. When the question of these systems is argued in legislatures, something true is being argued, and we are glad the argument exists, and we decline to mistake watching it for working.

A practice responds in the material it owns. Our material is architecture. The room we can change is the one where the tool meets the person — and in that room, every commitment the louder rooms are demanding is already enforceable by us, today, without anyone's permission: the judgement stays open to inspection; the tools stay free; the dependency stays visible; the draw stays measured. We do not wait for an instrument to compel us. That is our response to the season — not a position on its proposals, but proof that the demands underneath them can be kept.

It is the oldest answer we trust — the commons answer: not to seize what is enclosed, but to build beside it what cannot be enclosed.

Why We Do Not Put It Down

The hardest question does not come from the louder rooms. It comes from the quieter one: if the debt is real, why use the thing at all? Put it down. Walk away. The case for abstention is serious, and it deserves an answer rather than a shrug.

Here is the answer. The weights exist. They will not be untrained, and they spread — copied, fine-tuned, embedded in infrastructure that states now treat as strategic. If we put the tool down tomorrow, not one maker is restored, not one swept-in life is released from the weights, not one data centre slows. Abstention by a practice our size repairs nothing. It removes one accountable user from a field that continues without him.

And it concedes the single claim worth disproving: that this technology cannot be used accountably at all. If everyone who holds the discipline walks away, the use does not stop — only the disciplined use stops. What remains is the field without its counter-example. We stay because the demonstration is the argument, and a demonstration cannot be made from outside the room.

There is also a positive case, and we will not be shy about it. The use does work worth doing. A student at midnight gets feedback whose reasoning they can open and check. A teacher gets an instrument that holds judgement where it can be inspected and adjusted, instead of buried where it cannot. These goods are small, free, and real. We do not claim they outweigh the debt by arithmetic — there is no arithmetic between these currencies. We claim the goods are genuine, and that abandoning them would return nothing to anyone the taking harmed.

Beneath all of this sits an older lesson, learnt from builders. A technology is not a verdict; it is a version. Nothing we use is finished, and nothing we make is either. Things do not need to be perfect — nothing is. They need to offer a scope of improvement; they need to be good enough to keep figuring with. So we think in versions, we prototype, we never stop figuring — and we hold that naming a technology's problems and then declining to work at them is a stance available to a critic, but not to a practice. The problems we have named in this position are not our reasons to leave. They are our work order.

But the justification is conditional, not blanket. It holds only while the discipline holds: the smallest model that does the job, the judgement kept out of the box, the tools free, the debt named on every surface, the draw measured. Use without these conditions is not justified by anything written here. The justification is not a licence. It is a lease — renewed by conduct, and voidable by it.

The Schedule of Repayment

A debt named and then carried unchanged is just guilt. We treat ours as something to work off, and the three debts pay down at three speeds.

The compute draw shrinks by direct action, and this is entirely in our hands. Smaller models, not larger. Local inference wherever the work can carry it — our desktop tool already classifies on a small model on the user's own machine, with only the narration reaching out. The standing direction is to push the heavy work toward the edge, so the draw approaches its floor. Every release should draw less than it could, and the choice should be recorded, not assumed.

The appropriation debt pays down slowly, and partly outside our control. As models trained on consented, licensed, or public-domain work become viable, we move onto them. As honest provenance-disclosure becomes possible, we adopt it. And we practise the repair that fits our own architecture: the split applied to authorship — the model's share of a work carries an obligation back to the commons; the person's judgement-contribution is theirs. Not everything an AI touched made public; that punishes the maker and ignores their judgement. The model's share returns. The person's share is owned.

The swept-in lives pay down slowest of all, and only by not being deepened. There is no crediting someone who sold nothing, and no removing them from weights already trained. What is in our hands is narrow but real: we build no model, scrape no one, and log no user. Our own practice adds not a line to this debt. When cleaner substrates exist, we move onto them and leave the swept-in behind in the old one, rather than reaching for them again.

The Curtain Goes One Layer Further Down

The one discipline that is wholly ours: we never present the infrastructure as clean.

Every surface that shows how a Koher tool works also names what the tool stands on — a frontier model, with the compute draw and the appropriated provenance that entails. The view that exposes the three-layer pipeline extends one layer further down, to the model the narration depends on. Visibility does not make us clean. It keeps us honest: it tells the truth about a debt while the debt is being worked off.

What We Refuse

• We refuse to let the black box judge. Consequential judgement stays in auditable code. The model narrates; it does not decide.
• We refuse to present clean. No claim of neutral infrastructure; no silence about provenance.
• We refuse concentration. The tools are free, require no account, log no one; the practice is non-profit; it grows as a mandala, not a tower.
• We refuse the largest model when a smaller one will do, and we refuse to leave the draw unmeasured.
• We refuse abstention — without denying the case for it. The harm was real. But the weights exist and spread, and withdrawing now mends nothing; it only removes one accountable user from a field that continues without him, and concedes the single claim worth disproving — that these systems cannot be used accountably at all.
• We refuse enlistment. When louder rooms argue remedies, we listen, and we recognise our diagnosis in their language — but our response stays in the material we own. Our materials are borrowed, and we say so, and we carry the borrowing as named debt; our response is not borrowed. Consideration is owed. Enlistment is not.

Where This Leaves Us

Forward-going, and honest about the bill. We use the power. We keep it on a leash. We name what it costs, and we work the cost down — learning, as we go, how to wield the black box without being carried away by it.

The argument about these systems will be made in many rooms this decade. In ours, it is made by demonstration.

We use it. We don't let it judge. We name the debt — and we pay it down.