Made to prove the machine wrong

The letter, or the screen, or the silence tells you the same thing in different registers: no. The account is not approved. The claim is denied. The application will not be advanced. And in the same motion, without anything being said aloud, a weight is transferred onto you. From this instant forward the question is no longer whether the system was right to decide against you. The question is whether you can prove it was wrong. That transfer is so quiet, so much a part of how these decisions arrive, that most people never notice it happened — they simply feel, correctly, that the ground has tilted and they are now walking uphill.

What has actually happened is a reversal of who carries the burden of proof. In any process we would recognize as fair, the party making an accusation has to make its case, and the person on the other end is presumed to be in the right until that case is met. An automated adverse decision inverts this without ever arguing for the inversion. The machine's output is treated as the settled fact, and you — the person it landed on — are conscripted into the role of the appellant, the one who must now assemble a case against a conclusion you did not see formed. The deepest unfairness here is not the error rate. Systems will always err. The unfairness is the allocation: the machine asserts, and you refute, and nobody decided that this is how it should be. It was simply defaulted into place.

The inverted onus

Consider what it is to be made the appellant against your own denial. You did nothing to invite the decision; it was made about you, on a schedule that was not yours, by a process you were not party to. And yet the entire cost of contesting it is now assigned to you. You must find the channel. You must compose the objection. You must produce whatever counter-evidence you can gather and hope it is the kind the system will accept. Meanwhile the decision sits there, presumed correct, doing its work — the account stays frozen, the claim stays denied, the door stays shut — for as long as it takes you to overturn it, if you ever can.

The presumption of correctness is the whole trick, and it is an unearned one. A human adjudicator who ruled against you would at least be understood to have made a judgment that could be mistaken; the ruling would carry the visible fingerprints of a fallible process. An automated output tends to arrive with none of that. It presents not as a judgment but as a reading, as though it had measured something true about you rather than inferred something contestable. This is why it feels less like being disagreed with than like being weighed — and why the burden to prove it wrong feels not just heavy but faintly absurd, as though you were being asked to argue with a scale.

A burden you cannot discharge

Now hold the burden up to the light and notice that it cannot be discharged, and that its impossibility is structural rather than accidental. To refute a decision you would need, at minimum, three things: the evidence it rested on, the process by which it reached its conclusion, and the fact of its having been made at all. The inverted onus withholds all three. You are shown a conclusion, not the inputs — so you are asked to disprove a claim about facts you are not permitted to examine. You are shown an outcome, not the reasoning — so you cannot say where the step went wrong, only that the destination is wrong, which is precisely the kind of objection these systems are built to wave away. And in the worst and most common case you are shown nothing at all, because the decision registered somewhere as a score or a flag and no notice was ever owed to you.

This is the part that should not be mistaken for bad luck. You are being asked to disprove what you are not allowed to see. That is not a hard burden; it is an impossible one, and it is impossible by construction. A person cannot mount a rebuttal to a case that has never been disclosed to them, any more than they could answer a charge whispered in a language they do not speak. The appeal channel may exist. The right to contest may be written down somewhere. But a right to contest a decision you cannot inspect is a right to shout at a wall and call the echo due process.

A fair process makes the accuser carry the charge; an automated one hands you the charge, hides the evidence, and asks you to prove your innocence to a wall.

Putting the burden back

The remedy is not to make the affected person a better appellant. It is to stop assigning them the burden in the first place — to build systems that carry their own. A decision that is entitled to act against you should first be able to show why: the evidence it actually consulted, the rules that were in force when it decided, enough of its state that the decision can be replayed rather than merely asserted. A system that can produce that account before it acts has met the burden that properly belongs to the party making the decision. A system that cannot has not earned the right to act, whatever its accuracy, because it is asking you to accept a verdict it is unwilling or unable to justify.

This is what a record is for, at bottom, and why it is more than administrative hygiene. A Decision Receipt that carries the inputs, the active rules, and the means to replay the decision does not just document what happened. It relocates the burden of proof to where any fair process keeps it — on the decider, not the decided-upon. When the basis is disclosed as a matter of course, you are no longer the appellant fighting a black box; you are a party to a decision whose grounds you can actually test. The asymmetry does not vanish, but it stops being total. The wall becomes a document, and a document can be read, checked, and answered.

None of this is peculiar to the person in the chair. It is the standpoint-level face of a larger shift already underway in how these decisions are governed — the slow, uneven movement of the onus back toward the institution that decides. But it is worth naming from the receiving end, because that is where the inversion does its damage and where its correction would first be felt. The measure of a just automated system is not how rarely it errs. It is whether, when it decides against you, it hands you what you would need to prove it wrong — or leaves you, once again, arguing your innocence to a wall.

— Dispatches · Summit Cognitive