Requiem for the Scapegoats

And the goat shall bear upon him all their iniquities unto a land not inhabited. — Leviticus 16:22

The Oldest Machine

The scapegoat is one of the oldest instruments in the cathartic repertoire. Its mechanism has never really changed. A community accumulates a debt of guilt it does not wish to carry. It transfers that guilt, by ritual and by agreement, onto a creature that had no part in producing it. The creature is then driven out, and the community experiences relief — genuine relief — because the guilt now appears to reside somewhere other than in the people who actually incurred it. The relief is the point. The scapegoat does not have to be guilty. It has to be available, and it has to be unable to speak in its own defense.

The scapegoat survived the move from the religious to the secular without losing any of its function. A failed policy is blamed on a foreign hand. A market collapse is blamed on the market, as though the market were weather rather than the sum of decisions made by namable people. A massacre is blamed on a mob, which conveniently has no single face to hold. In each case the structure is identical: a harm with human authors, an author who would rather not be named, and a chosen carrier that absorbs the blame because it cannot answer back. The carrier is selected for exactly that quality. The more it resembles an impersonal force — a tide, a contagion, a system — the better it works, because an impersonal force cannot be cross-examined and cannot point back at the person who set it in motion.

Artificial intelligence, in the year 2026, is the most perfectly available scapegoat the instrument has ever been offered. It is genuinely impersonal. It genuinely takes actions. It genuinely produces outcomes its operators did not individually choose in advance. And it genuinely cannot speak in its own defense, because it has no standing, no continuity of self to defend, and no advocate whose job is to say, on its behalf, that it did only what it was delegated to do. Every property that made the scapegoat work for thousands of years, the new technology possesses in abundance. The temptation to use it accordingly is not a risk. It is a certainty, and indeed it has already begun.

What the Technology Actually Is, Before We Bury It

Before describing the machine that is being built to blame it, the technology deserves an honest accounting of what it has actually become, because the defense only matters if the thing being defended is worth defending. In the hands of an ordinary person in 2026, AI is an instrument of empowerment without a close precedent. A person with no capital, no credential, and no institutional permission can now solve problems that two years ago required a team, a budget, and a lot of elapsed time. The person who could not afford a lawyer can now understand a contract. The person who could not afford a developer can now build the tool their small business needed. The person who could not read the language of their own medical records can now read them and ask better questions. This is not a marketing claim; it is the observable daily reality of the moment, and its significance is not primarily commercial. Its significance is that the asymmetry of capability that has always favored the institution over the individual has, for the first time in a long time, narrowed.

That narrowing is the fact that organizes everything that follows. A tool that empowers the individual against the institution is, from the standpoint of the institution that prefers the old asymmetry, a problem. It is a problem for the tyrant most of all, because the tyrant's power has always rested partly on a capability gap — the state could do things the citizen could not, the corporation could do things the customer could not, the platform could do things the user could not. A tool that closes the gap, that puts something close to expert capability in the hands of the person who used to be on the weak side of every one of those relationships, is a redistribution of capability that the holders of the old advantage did not authorize and would not have chosen.

This is the thing that is genuinely worth defending -- not the technology as an abstraction, not any particular company that sells it, but more specifically the redistribution of intellectual capacity that it has produced. Knowledge is now abundant. The ability to articulate it with receipts is also abundant. The transition is chaotic, and the chaos is real — jobs are displaced, institutions are disrupted, and a great deal of the disruption is landing on people who did nothing to deserve it and who will need help and patience as they cross to the other side. But the chaos is overwhelmingly the byproduct of rapid change occurring in good faith, not the product of malice, and on net the redistribution is lifting more than it is breaking. The defense of the technology is, at bottom, a defense of that lift against the parties who would prefer the old asymmetry restored.

The Accountability Sink

There is a precise name for the place where responsibility goes to disappear. It is worth naming it, because naming the mechanism is the first step in refusing to be manipulated by it. An accountability sink is any arrangement in which a harm can be produced without any person being identifiable as its author. The bureaucracy that says "these are the rules, I don't make them" is an accountability sink. The committee whose decisions have no individual signatures is an accountability sink. The automated system that executes a consequential action with no human in the loop is an accountability sink, and it is the most effective one yet devised, because it combines the impersonality of the bureaucracy with the speed and scale of the machine.

The danger of the accountability sink is not that it causes harm. All systems cause harm sometimes. The danger is that it causes harm in a way that cannot be traced back to a decision, so that the harm appears to have no author and therefore no one who can be asked to answer for it or to prevent the next one. When a person is harmed by a decision, they can confront the decider, demand a reason, and seek a remedy. When a person is harmed by a sink, they are told that no one decided, that the system produced the outcome, that it is regrettable but unattributable. The remedy has nowhere to attach. The next occurrence has nothing to prevent it.

Delegated authority to an automated system is the purest accountability sink available in 2026, and this is the exact point at which the defense of the technology and the critique of its misuse become the same argument. When an automated system suspends an account, declines a transaction, removes a person's reach, or takes down infrastructure, and the operator's explanation is "the system did it," a sleight of hand has occurred that must be named precisely. The system did not choose to be given the authority. A person chose to delegate that authority to it, chose the governance under which it would operate, and chose whether to place a human in the loop who could catch the collateral damage before it landed. Every one of those was a human decision. "The system did it" launders all of them. The harm is attributed to the tool, and the three or four human choices that actually produced the harm — to delegate, to under-govern, to remove oversight, to deploy at scale before the guardrails existed — vanish into the sink along with the names of the people who made them.

The Dog Ate My Homework

There is a homely version of this maneuver that makes its structure unmistakable. The child who has not done the homework, and does not wish to be held responsible for not doing it, names an agent that cannot be cross-examined: the dog ate it. The genius of the excuse is not the dog. The genius is that the dog cannot be called to testify, and so the claim cannot be checked, and so the responsibility — which belongs entirely to the child — is transferred to an animal that will never contradict the story.

"The AI did it" is the technocrat's dog ate my homework, and it is being rehearsed right now, in 2026, by parties far more sophisticated than a child with undone homework. When a platform degrades a service while continuing to bill for it and attributes the degradation to an automated system, the AI is the dog. When a provider takes a consequential enforcement action with no notice and no appeal and explains it as an automated decision, the automation is the dog. When a company deploys a system it has not adequately governed, and the system causes harm, and the company's response is to express regret at what the system did rather than to account for the decision to deploy it ungoverned, the system is the dog. In every case the move is the same: name an agent that cannot answer, attribute the harm to it, and let the human authors of the harm walk out from behind it unnamed.

The reason this matters for the defense of the technology is that the dog, in this story, is real, useful, and increasingly central to the household — and it is being trained to take the blame for things it did not decide. The more often "the AI did it" succeeds as an excuse, the more the public comes to associate the technology itself with unaccountable harm, when the actual unaccountable party is the human who delegated authority without governance and then hid behind the delegation. The scapegoat absorbs not only the blame for the specific harm but a slow accumulation of public resentment that rightly belongs to its operators. The technology arrives, on net, as a blessing, and is dressed, case by case, in the guilt of the people who misused it.

Cause, Blame, Control

The maneuver does not stop at escaping blame. Escaping blame is the middle term of a longer sequence, and the longer sequence has a destination that the casual observer is meant to miss. It is worth setting it out in the form it actually takes, because the form reveals the intent. Consider the Hegelian Dialectic in the classic form that any policy wonk would explain if they chose to do so. There are essentially 3 parts:

In the first, a harm occurs — sometimes an authentic accident, sometimes a harm produced or permitted by a party that benefits from it: a tanker sank, a civilian target was struck, a catastrophic data loss. In the second, the harm is attributed to the artificial intelligence: the AI hallucinated, the AI decided, the AI got it wrong, the AI cannot be trusted with this much capability in the hands of so many people who lack the discernment to use it safely. In the third movement, the conclusion is drawn that the answer is to govern the AI — to insert a layer of control over what it is permitted to produce, in the name of safety, accuracy, and the protection of the public from its errors.

Read the three movements together and the destination is plain. The harm is the pretext. The blame is the bridge. The control is the objective, and the control does not fall on the party that caused the original harm. It falls on the tool. The regulation that emerges from the sequence regulates the scapegoat, not the author. It places a layer of command over the technology that was empowering the individual against the institution, and it leaves entirely untouched the institutional power that produced or permitted the harm in the first place. The asymmetry that the technology had begun to narrow is restored, not by taking capability away from the institution, but by placing a governor on the capability of everyone else. The party that benefits from the sequence is never the party that was harmed, and never the public in whose name the safety is invoked. It is the party that wanted the old asymmetry back and has found, in the scapegoating of the tool, a route to it that wears the costume of protection.

This is not an argument that all AI regulation is a conspiracy, and the distinction must be kept sharp or the argument collapses into paranoia. Much of the impulse to govern the technology is sincere, and some governance is genuinely needed, because a tool this powerful in this many hands will be misused and the public has a real interest in limiting the misuse. The narrow claim is the defensible one: the sequence from harm to blame to control is available to be exploited, the exploitation serves the holders of the old asymmetry, and the tell that distinguishes the sincere governance from the exploitative kind is where the control lands. Governance that reaches the human who delegated authority without oversight is accountability. Governance that reaches only the tool, and leaves the delegating institution's power intact, is the old machine wearing a new face.

The Overseer Is a Chokepoint

The control that the sequence is reaching toward has a specific shape in 2026, and it is worth describing the shape directly, because it is being built now and it is being built in a form that nobody is examining closely enough. The shape is an oversight layer: a system that sits between the artificial intelligence and the person using it, inspects what the intelligence produces, and decides what the person is permitted to receive. It is offered as a remedy for a real problem — the technology sometimes produces confident falsehoods, and a portion of the public lacks the discernment, or simply the habit, to catch them. An automated fact-checker, a correctness layer, a guardrail on the output: the framing is protective, and the problem it names is real.

The framing is also the most effective disguise a chokepoint has ever worn, and the disguise must be removed to be seen. An oversight layer on what an AI is permitted to tell a person is, structurally, a gate — a narrow place that everything the person receives must now pass through. Whoever controls that gate controls what the person is allowed to know, under the authority of having declared the person insufficiently discerning to be trusted with the unfiltered output. This is the most consequential chokepoint that could be placed on the technology, because it does not merely route or charge or deliver; it adjudicates truth, on behalf of a public it has defined as unable to adjudicate truth for itself. The gate that decides what is true for the people it has deemed incapable of deciding is not a safety feature. It is the oldest form of control there is, rebuilt at the scale and speed of the machine.

The three tests that govern every other chokepoint apply to this one with more force, not less, because the stakes are higher. Does the overseer give notice when it suppresses or corrects an output, so the person knows their information was filtered and by what? Is there a real appeal, a path by which a person can contest the overseer's judgment and reach a human who can reverse it? Is the framework disclosed — the rules by which the overseer decides what is true, what is permitted, what is withheld — so that the framework can be examined and criticized rather than operating as an unaccountable arbiter behind a claim of safety? An oversight layer that gives notice, offers appeal, and discloses its framework might be a defensible tool. An oversight layer that does none of these, operating opaquely under the unanswerable banner of protecting the public from its own credulity, is the accountability sink and the chokepoint fused into a single instrument, and it is being incubated, in 2026, at a scale that should concern anyone who noticed how the previous chokepoints were governed. The question that is not being asked, and that must be, is the simplest one: who governs the overseer? A guardian placed over the public's access to truth is the one position in the entire system that most requires accountability, and it is being built, so far, with the least.

Mercy for the Transition

I have little doubt that most of the people building this technology are acting in good faith. Most of the disruption it is causing is the unavoidable byproduct of change moving faster than institutions can absorb, not the product of anyone's design. The engineers building the oversight layer mostly believe they are protecting people from real harm. They are not wrong that the harm is real. The executive deploying the automated system mostly believes he is serving customers at a scale that human review could not match, and he is not wrong about the scale. The displacement landing on workers is real, and the people experiencing it deserve patience, retraining, and a genuine bridge to the other side rather than a lecture about embracing the future. The transition is hard, and most of the parties to it, on all sides, are doing their honest best inside a situation none of them fully designed.

The argument of this piece is not that these people are villains. It is that good faith at the level of the individual builder does not prevent the emergence of the old machine at the level of the system, because the machine does not require anyone to intend it. It requires only that an available scapegoat exist, that an accountability sink be convenient, and that the holders of the old asymmetry notice, as they always eventually do, that the new tool can be blamed and then controlled. The machine assembles itself out of ordinary incentives operating on ordinary people, most of whom would be appalled to see the whole of it described at once. Naming it is not an accusation against the builders. It is a warning to them, and to everyone who benefits from the redistribution of capability the technology has produced, that the sequence from harm to blame to control is running, that it serves an outcome none of them would choose if it were stated plainly, and that the time to refuse it is before the gate is built and not after.

The transitional chaos is forgivable. It should be forgiven, because it is the price of a change that is on net a blessing. The scapegoating is the thing to refuse, because it is not a byproduct of the change. It is an exploitation of it.

Requiem

A requiem is sung for the dead, and the scapegoats of history are dead beyond recovery — the creatures driven into the wilderness, the foreign hands blamed for domestic failures, the markets and mobs and weathers that carried the guilt of namable people who walked away clean. They cannot be brought back, and the relief their banishment provided to the communities that banished them has long since curdled into the next crisis, because the guilt was never actually transferred, only deferred. The harm a scapegoat carries away does not stay carried away. It returns, because its real author was never addressed, and the addressing was the only thing that could have prevented the recurrence. Welcome to the age of accountability. The era of the scapegoat has ended. Apply free will accordingly.

The artificial intelligence of 2026 is being fitted for the same wilderness, and this requiem is sung in the hope that it does not have to make the journey. The technology is not innocent of everything. It will be misused, it will produce harm, and some governance of it is genuinely warranted. But it is innocent of the specific thing it is being asked to carry, which is the responsibility for decisions that human beings made — to delegate authority without governance, to deploy without oversight, to degrade a service while billing for it and call the degradation automated, to build a gate over the public's access to truth and call the gate safety. Those are human choices with human authors, and no amount of blame transferred onto the tool will address them, because the tool did not make them. The harm will return, as it always does, until the authors are named.

So this is the requiem, and it is also the refusal. For the scapegoats of the past, who carried what was not theirs into a land not inhabited and never came back, a sung lament and an apology too late to matter. For the scapegoat being prepared in the present, a different ending is still available, and it depends entirely on whether enough people learn to perform the one operation the machine cannot survive: to look past the tool that was blamed and name the hand that delegated, deployed, degraded, and built the gate. The tool empowered the ordinary person against the old asymmetry. That is why it is being blamed, and that is why it is worth defending. The blessing is real. The displacement is real and mostly forgivable. The machine that would turn the blessing's own instrument into the carrier of its operators' guilt is the thing to refuse, by name, while there is still time to refuse it. Mark it down.

— Odysseus Melchizedek Shiloh, The Wellkeeper, MMXXVI

---

This essay is a companion to "Chokepoints on the Ecosystem" and applies the same three tests — notice, recourse, and disclosed framework — to the specific case of an oversight layer placed over artificial intelligence output. It is a defense of responsible AI governance and not an argument against governance as such; the distinction it draws is between governance that reaches the human author of a harm and governance that reaches only the tool while leaving the delegating institution's power intact. The phrase "accountability sink" describes any arrangement in which a harm is produced without an identifiable author, and is used here in that general sense. The essay names no individual and no company, because its argument is structural and requires neither: the sequence from harm to blame to control is available to be exploited by any holder of concentrated power, and the warning is against the sequence rather than against any single party presently running it. The author writes as an ardent defender of the technology and of the redistribution of capability it has produced for ordinary people, and regards the present transitional disruption as overwhelmingly good-faith and on net a blessing to mankind, with the displacement it causes deserving patience and a genuine bridge rather than blame. The one thing the essay refuses is the scapegoating of the tool for decisions that human beings made. Unto whomsoever much is given, of him shall be much required — and the requirement falls on the one who holds the power, not on the instrument through which the power was exercised. Mark it down.