The Human Terminus: What Admissibility Assumes, and What Survives When the Assumption Fails
METHODS · WORKING PAPER — Governance. Programme: Admissibility & Evidence. Analytical, anchored to dated literature and law as of July 2026.
AMENDMENT NOTICE. This paper amends the test set out in Assisting or Deciding? (Agentic Government, 2026) and revises the scope of The Decision That Was Never Made. The earlier formulation rested on an assumption this paper finds untenable. What follows corrects it in public, on the record.
Every method this Institute has built terminates in a human being. The delegation record must trace back to a person. The assisting-or-deciding test asks whether an official could comprehend the basis of a determination. The remedy for a ratified decision is a genuine re-decision by a human who actually decides. Admissibility asks who has standing to refuse. Each of these assumes that somewhere at the end of the chain stands a person who understands what happened and answers for it.
That assumption has never been examined. This paper examines it, and finds that it does not hold — not in the future, and not because agents will overtake us, but now, on the evidence, for reasons that are structural and already documented. What follows is not a concession that accountability is impossible. It is an argument that accountability was never located where we placed it. It does not live in a person’s comprehension at the moment of decision. It lives in whether the decision can be reconstructed, by someone with standing, and reversed. The terminus is not a mind. It is a record and an institution that answers for it.
What the method assumes
Stated plainly, so it can be attacked: every consequential AI-mediated decision must pass through a human who understands it well enough to refuse it. From that assumption everything else in the corpus follows. The record exists so the human can reconstruct. The stop exists so the human can refuse. The delegation chain must terminate in a person because a person is what answers. The assisting-or-deciding test asks whether the human really decided, because a decision no human really made is one the law did not authorise.
Remove the assumption and the architecture must be rebuilt. So it is worth knowing whether it is true.
The empirical challenge: the human cannot oversee
It is not true, and the finding is neither new nor speculative. Ben Green surveyed 41 policies prescribing human oversight of government algorithms and identified an uninterrogated assumption underneath all of them: that people are able to effectively oversee algorithmic decision-making. He concluded that the evidence indicates they cannot perform the oversight functions the policies require — and, crucially, that this failure is not neutral. Because the human cannot really oversee, oversight requirements legitimise government use of faulty and controversial algorithms without addressing anything fundamental about them. Rather than protecting against harm, they produce a false sense of security and let vendors and agencies shirk accountability.
This is the sharpest possible objection to the corpus, and it must be met head-on. A requirement that cannot be met does not fail harmlessly. It fails by becoming a ceremony that certifies the thing it was meant to constrain.
The moral consequence: the human absorbs the failure
The second half of the pincer was published earlier and names the aftermath. Madeleine Clare Elish’s concept of the moral crumple zone describes how responsibility for an action is misattributed to a human actor who had limited control over the behaviour of an automated system. The analogy is exact and merciless: a car’s crumple zone absorbs impact to protect the driver; the moral crumple zone absorbs blame to protect the integrity of the technological system, at the cost of the nearest human operator. Elish traced the pattern through Three Mile Island and Air France 447, described operators as liability sponges filling the accountability gaps that complex systems create, and noted that certification paradigms can themselves reify the conditions that produce these zones. The underlying insight belongs to a longer safety-science tradition — Dekker, Reason, Woods — reaching back to the 1980s.
Read Green and Elish together and the pincer closes. Before the failure, the human at the terminus legitimises the system. After the failure, the human at the terminus absorbs the blame. In neither moment does the human produce accountability. The placement of a person at the end of the chain does exactly two things, and neither of them is the thing it was placed there to do.
The Institute’s own contribution sits precisely between these two findings. Ceremonial oversight names the mechanism — a human positioned near the interface, not at the boundary, whose approval is collected after the system has determined the path. Elish’s moral crumple zone names its consequence — that same human, afterwards, holding the blame. The mechanism produces the consequence. That is a real contribution, and it is only legible when both authors are cited.
The trap our own test walks into
Now turn the instrument on ourselves.
Assisting or Deciding? holds that an official is genuinely deciding only if, among other conditions, they can reconstruct and comprehend the basis of the determination in the time available. Suppose that test is applied honestly, at scale, in a modern agentic administration. Green’s evidence predicts what happens: the comprehension condition fails almost everywhere. Then the remedy framework in The Decision That Was Never Made classifies the failure as structural, and the prescribed remedy is to withdraw the chain’s standing for that class of decision.
Applied faithfully, our own method suspends most AI-mediated public administration. That is not a governance framework; it is a demand that the state stop using tools it has already built its capacity around, justified by a condition we now have good reason to believe was never satisfiable. Our remedy paper warned against exactly this — that treating every lapse as structural would paralyse administration — and offered the isolated/structural distinction as the safeguard. But if comprehension fails by default rather than occasionally, the distinction collapses: everything is structural.
The alternative outcome is worse. The test is quietly softened, comprehension becomes a box, and the Institute has contributed one more instrument of legitimation — the precise failure Green documented. Either our test is unusable, or it is ceremonial. The assumption is what is wrong.
The conflation
The error is identifiable and, once seen, elementary. The method used the word human to mean two different things and assumed they were the same person.
The comprehender: someone who, at the moment of decision, understands the basis on which it is being made.
The terminus of answerability: some entity to which the decision can be traced, which can be made to explain it, reverse it, and bear consequences for it.
Comprehension is a cognitive state of an individual at a moment. It fatigues, queues, and degrades with volume and complexity — which is why Green finds it failing and why the decision-to-time ratio is such a reliable signal. Answerability is a structural property of an institution across time. It does not fatigue. It does not have two hundred cases before lunch.
We assumed the terminus had to be the comprehender because, historically, they were the same official. Agentic systems break the identity, and the method mistook a historical coincidence for a logical requirement.
What the corpus already knew and did not follow through
Here is the uncomfortable part: the answer was already written, in the Institute’s own foundational rule, and we did not read it.
No record, no standing does not require that anyone comprehend anything in real time. It requires that a decision be reconstructable — that an independent party could rebuild it well enough to challenge it. Reconstructability is a property of a record, not a state of a mind. A record can be read tomorrow, next year, by someone who was not present, by a court, by an auditor, by an institution. It does not need the official to have understood at 4:47 p.m. with sixty cases still queued.
The Institute’s evidentiary rule was never a comprehension standard. It was a reconstruction standard. Every part of the method that assumed a comprehending human was an unforced addition, imported from an administrative world where the two happened to coincide.
Reconstructive accountability
So the condition of a lawful AI-mediated decision is not that a human comprehended it. It is that five things hold:
1. Reconstructability. The decision can be rebuilt from a record — including the delegation record, and the legible substance of the exchanges that produced it — well enough to be challenged.
2. Institutional answerability. A named institution, not a nominated individual, is answerable for the decision, and cannot discharge that answerability by producing a signature.
3. Standing to trigger. Someone — including, centrally, the person the decision was about — can compel the reconstruction, without needing to first prove the harm the reconstruction would reveal.
4. Reversibility. Where reconstruction shows the decision was wrong, it can actually be reversed, before or after the fact, with a route that reaches the person.
5. Independent verification. The reconstruction is performed or checkable by a party that does not answer to the operator.
These are the conditions. Notice what has left: the requirement that a specific official held the whole basis in their head at the moment of signing. Notice what has arrived: a record that must actually exist, an institution that cannot hide behind a person, a person who can compel it to answer, and an outsider who can check.
This is a tightening, not a relaxation
The obvious objection is that we have abandoned the human safeguard. The opposite is true, and this is the paper’s central claim.
The human signature was the cheap condition. It can be produced by an exhausted official in four seconds, and it is unfalsifiable from outside: a signature looks identical whether or not a decision occurred. That is precisely why it proliferated, why it legitimises, and why the failure lands on the signatory. Reconstructive accountability cannot be produced in four seconds. A reconstructable record either exists or it does not; independence either holds or it does not; reversal either happens or it does not. Every one of the five conditions is observable, falsifiable, and expensive.
Moving the terminus from a person to a record and an institution does not lower the bar. It moves the bar somewhere it can bear weight. The signature never could.
Administrative law reached this conclusion centuries ago — and then forgot it
There is a conservative version of this argument, and it is stronger than the radical one.
Administrative and public law, in many traditions, has long secured accountability reconstructively, not cognitively. The duty to give reasons does not ask whether the decision-maker understood; it asks whether the reasons stand on the record. Judicial review examines the record, not the official’s state of mind. The appellate structure exists precisely because comprehension by the first decision-maker was never assumed to be reliable.
Those traditions do also contain a comprehension-like requirement — the decision-maker must exercise their own judgment, must not fetter their discretion, must not act under dictation. And it is worth noticing what that ancient ground of review describes: an official who ratifies the determination of another. The ceremonial human has an ancestor in administrative law, and its name is acting under dictation. But that doctrine developed where the dictating party was another official — someone who had comprehended. It was never designed for a dictating party that is a chain of agents no one can read.
So the current wave of “meaningful human oversight” requirements is not, as it presents itself, a modern safeguard. It is a regression to a comprehension standard that public law had already learned not to rely on, imported into the one setting where comprehension is least available. The reconstructive machinery — reasons, record, review, independence — is the older and better answer, and it is the one that scales.
Amendment to the assisting-or-deciding test
The test as published asks whether the official could reach a different outcome, could reconstruct and comprehend the basis in the time available, and recorded genuine engagement. That test is not withdrawn. It is relocated, and its status changes.
As a diagnostic of ceremonial oversight, the test stands, unamended. Override rate, decision-to-time ratio, and recorded engagement remain the three best observable signals of whether a human is deciding or ratifying. They measure the ceremony accurately, and the ceremony is worth measuring.
As the condition of a decision’s lawfulness, the test is withdrawn, except in a reserved tier. In its place:
Tier 1 — reserved decisions. Where the law reserves the decision to a human being — deprivation of liberty being the clearest case, and any decision whose value lies partly in a person having considered another person — comprehension by the deciding official remains a necessary condition, and a ratified decision is unlawful. This tier should be defined narrowly and defended explicitly, on grounds of dignity rather than accuracy.
Tier 2 — everything else. The condition is reconstructive accountability: the five conditions above. A human signature is neither necessary nor sufficient. A decision with a signature and no reconstructable record fails. A decision with no signature, a reconstructable record, an answerable institution, standing, reversibility, and independent verification passes.
This bullet has to be bitten openly: under this amendment, a ratified decision with an intact record and a live route to reconstruction and reversal is not, for that reason alone, unlawful. The Decision That Was Never Made said such decisions never simply stand. That was wrong in Tier 2, and the reason it was wrong is that it treated the signature as if it had been doing work. It was not. Where the record fails, the earlier paper’s remedy framework applies with full force and, indeed, greater force — because now the record is the whole of the safeguard, not one of two.
What is lost
Something real is lost, and an institute that pretended otherwise would be doing the thing it accuses others of.
The demand that a person consider your case is not only an accuracy claim. It is a dignity claim: that a human fate should pass before human attention. Green observes that policymakers invoke human rights and dignity to justify human oversight while rarely offering evidence that it delivers them. That observation cuts the empirical justification, not the moral one. It is coherent to hold that being considered by a person is owed to you regardless of whether it improves your outcome — and if that is right, reconstructive accountability may deliver better decisions and worse treatment.
This paper does not resolve that. It only insists on the honesty: a comprehension requirement that is empirically unmeetable does not deliver dignity either. It delivers a signature, a legitimated system, and, when things fail, an operator holding the blame. If dignity is what we want, it must be secured deliberately, in a narrow reserved tier, and paid for — not assumed to arrive as a side effect of a box that someone ticked.
The regress, and the limit of this paper
The deepest objection is the one we cannot answer, and it is worth stating in its strongest form.
If no human comprehends a decision, and the record that makes it reconstructable is itself read and checked by machines, in what sense is anyone accountable? Machines auditing machines, producing renderings no person verifies against anything, is not accountability. It is a hall of mirrors with an audit trail.
The partial answer is that accountability was never a cognitive state. It is a liability relation. A corporation is accountable without comprehending; so is a state. What is required is an entity that can be compelled to explain, made to reverse, and forced to bear consequences — and that entity can answer without any single person having held the whole decision in mind. This is why condition 2 names an institution rather than an individual, and why condition 5 demands independence rather than understanding.
But the answer is only partial, and here the paper stops honestly. An institution that cannot understand what it did cannot reliably fix it. A person facing a state that produces a perfect record no human ever read is facing something for which we do not yet have a word, and possibly not a remedy. We do not know how far accountability can travel from comprehension before it becomes a shape rather than a fact. We claim only that it travels further than the corpus assumed, that the human signature marked the wrong boundary, and that the record marks a better one.
What this paper does not settle
Four questions, held open. First, the boundary of Tier 1 — which decisions the law should reserve to human consideration on dignity grounds — is a political and constitutional question this Institute has no standing to answer alone, and should not pretend to. Second, whether machine-performed reconstruction of a machine-produced decision constitutes independent verification, or merely relocates the problem, is unresolved and may be the central question of the next decade. Third, Green’s own remedy is not ours: he proposes shifting from human oversight to institutional oversight with democratic pre-approval before an agency may adopt an algorithm at all. That is admissibility with a democratic gate, it is more demanding than anything proposed here, and the Institute has not yet taken a position on it. Fourth, whether the ceremonial human should ever be retained deliberately — as a moral crumple zone the state accepts because someone must be answerable in a way an institution cannot be — is a question we find repugnant and cannot dismiss.
FAQ
Is this paper saying we should remove humans from AI decisions?
No. It says the human signature was never the safeguard it was taken for: on the evidence, officials cannot perform the oversight these policies require, and requiring it legitimises the systems it was meant to constrain. Accountability should be relocated to where it can bear weight — a reconstructable record, an answerable institution, standing to compel reconstruction, reversibility, and independent verification.
Doesn’t that lower the standard?
It raises it. A signature can be produced in seconds and looks identical whether or not a decision occurred. A reconstructable record, an independent verifier, and a working route to reversal cannot be faked into existence. The human signature was the cheap condition, which is why it spread.
Does the Institute now accept ratified decisions?
In a reserved tier — where the law reserves the decision to a human, most clearly deprivation of liberty — no: a ratified decision remains unlawful. Everywhere else, a decision stands or falls on reconstructive accountability, not on whether an official comprehended it. This amends our earlier position, which treated the signature as if it were doing work it was not.
Isn’t “beyond the human terminus” a speculative claim?
It is the opposite. The empirical finding that human oversight fails is dated and published; administrative law secured accountability reconstructively — through reasons, records, and review — long before AI existed. The speculative position is the one that assumes a comprehending official is available at the end of every agentic chain.
