The burden of proof moves

There is a rule so old it feels like gravity: the party that asserts something bears the burden of showing it is true. The accuser proves the accusation. The lender that denies the loan owes a reason. The agency that revokes a benefit has to say why, on the record, in a form the affected person can examine and contest. The burden of proof sits with the one who acts, because that is the only arrangement in which power can be held to answer for itself. It is not a technicality. It is the load-bearing beam of every fair process humans have ever built, and it points in one direction — toward the institution, away from the individual.

Automation does not repeal this rule. It does something quieter and more corrosive: it moves the burden without anyone deciding to move it. When a decision is made once by a person, the reasoning is fresh, the decider is present, and it is natural to ask them to explain. When the same decision is made ten million times by a system, the reasoning is nowhere, the decider is a process that has already moved on, and the person who wants to know why is met not with an account but with a silence dressed up as finality. The institution has not refused to justify itself. It has simply arranged things so that justification is never produced — and the absence of a justification lands, by default, on the only party still standing there: the individual.

The shift is easy to miss because nothing announces it. No one issues a policy that says henceforth the accused shall prove their own innocence. The move happens in the architecture. A system that keeps outputs but not reasons has, in effect, made the individual the party of first resort — because when they contest the decision, the institution has nothing to show, and so the burden of assembling a case falls entirely to the person who has the least information, the least time, and the least power to assemble it. The default of silence is not neutral. It is a thumb on the scale, and the scale tips away from the party that decided.

A silence where a justification should be is not the absence of a decision. It is a decision — to make the weaker party prove what the stronger party declined to.

How the reversal actually works

Consider what an affected person is asked to do when they want to challenge an automated outcome. They must first establish that a decision was even made — not always obvious when the outcome is a silent denial, a score below a threshold, an application that simply goes nowhere. Then they must reconstruct, from the outside, what the decision rested on: which inputs, which rule, which version of a policy that may since have changed. Then they must show that some part of it was wrong. Each of these steps is an evidentiary burden, and every one of them has been handed to the party who cannot possibly meet it, because the evidence lives inside the institution and the institution kept none of it.

This is the reversal in its full form. The institution that made the decision holds — or could have held — everything needed to justify it: the inputs it consulted, the rule it applied, the state that would let the decision be replayed. The individual holds nothing but the outcome and the conviction that it is wrong. In the old arrangement, the party with the evidence bore the burden. In the automated one, the party without the evidence bears it. The burden did not lighten; it migrated to the shoulders least able to carry it, and it did so precisely because scale made the reasoning cheap to discard and expensive to reconstruct after the fact.

And the migration compounds. Because the individual usually cannot meet the burden, most wrong decisions are never successfully contested, which the institution reads as evidence that its decisions are sound. Low complaint rates get mistaken for high accuracy. The reversal of the burden thus hides its own consequences: the very mechanism that stops errors from being challenged also stops them from being counted, so the system looks more reliable exactly to the degree that it has made itself impossible to question.

Putting the burden back

Restoring the burden to where it belongs is not a matter of good intentions or a promise to be more transparent. It is a matter of what the system produces at the moment of decision. Two things do the work. The first is provenance — a preserved account of what the decision actually rested on, made when the decision was made, so that the institution can be asked to show its reasoning and can answer from a record rather than a reconstruction. Provenance re-arms the party that acted with the obligation to justify, because you can only be asked to justify what you kept the means to justify.

The second is standing — the affected person's recognized claim to that account. A Decision Receipt given to the individual, that they hold and can carry to a regulator or a court without asking permission, changes who has to prove what. With it, the person no longer arrives empty-handed to reconstruct a case from nothing; they arrive holding the same record the institution holds, and the burden of showing the decision was sound returns to the party that made it. Provenance and standing together are how contestability becomes real rather than nominal — not a right that exists on paper and collapses in practice, but one backed by the evidence needed to exercise it.

The point is not that institutions should lose every challenge. It is that the challenge should be a genuine contest, decided on a shared record, rather than a foregone conclusion decided by which party happens to hold the evidence. When the burden of proof sits with the individual, the outcome is settled before the dispute begins, because the person who must prove and cannot will lose by default. When it sits, as it always should, with the party that acted, the decision has to be able to answer for itself — and a decision that can answer for itself is the only kind that deserves to stand. The burden moved while no one was watching. Provenance and standing are how it moves back.

— Dispatches · Summit Cognitive