The burden of proof is moving

Inside every accountability regime there is a quiet allocation that almost no one states aloud: a default about who has to prove what, to whom, before anything happens. It is the most consequential rule in any dispute, and it is usually invisible precisely because it is a default — the setting that holds when no one has argued otherwise. Tell me who carries the burden, and I can tell you, before a single fact is examined, who is going to win most of the time. The facts matter at the margins. The default decides the rest.

For the better part of the automated era, that default has run in one direction. The system is presumed correct. The affected party — the person denied, flagged, downgraded, removed, declined — bears the burden of demonstrating that something went wrong. And they are asked to carry that burden under conditions that make it nearly impossible to meet: without access to the evidence the decision actually consulted, without the rules that were active at the moment it was made, without any way to reproduce the outcome and watch where it diverges. You are told a machine decided, that the decision was made fairly, and that if you believe otherwise, the demonstration is yours to construct. From the outside, with nothing in hand.

That arrangement was not a conspiracy. It was a reasonable accommodation to a world that no longer exists. When automation was narrow — a calculation here, a sort there, a threshold applied to a clean number — the stakes of any single decision were low, the logic was legible, and a motivated person could usually reconstruct what had happened well enough to contest it. The presumption of correctness was cheap because error was visible and rare. The burden sat on the affected party because, in that world, the affected party could actually lift it.

Why the old default is failing

None of those conditions hold anymore. Automated systems now make consequential calls at scale and at speed, across domains where the inputs are vast, the logic is layered, and the output arrives without anything resembling a legible trail. The decision that shapes a life is no longer a threshold on a clean number; it is the compressed result of a process the affected party cannot see, cannot query, and cannot rerun. The presumption of correctness has not changed. Everything that once justified it has.

So the burden now sits exactly where it can least be discharged. We ask the person with the least information, the least access, and the least standing to prove an error inside a system designed — sometimes deliberately, sometimes only by neglect — to be unprovable from the outside. This is not a high bar. It is an impossible one wearing the costume of a high bar. And a burden that cannot in principle be met is not a safeguard against false claims. It is a guarantee that real harms go uncontested, which is a different and worse thing.

A burden that cannot in principle be met is not a high standard of proof. It is the quiet retirement of proof altogether.

What makes this unstable is not only that it is unfair. It is that the unfairness has become legible to the people on the wrong side of it, and to the institutions watching from a distance. A default survives on consent — on the broad sense that it is roughly the right way to arrange things. That consent is draining. The more obviously a presumption protects the powerful from scrutiny rather than protecting the truth, the less weight it can bear, and the more pressure accumulates against it.

Which way it is moving

That pressure is converging on a new default, and it is arriving from several directions at once. Regulators are moving, in their idiom, toward duties to explain and to substantiate consequential automated decisions — shifting the obligation from the person affected to the party that deployed the system. Litigation is following the same logic: when one side controls all the relevant evidence and the other side controls none, courts have old instincts about where the burden ought to sit, and those instincts do not favor the party holding the black box. Procurement is moving too; the institutions that buy automated systems are beginning to ask, as a condition of purchase, whether the system can produce an account of itself on demand. And beneath all of it runs plain self-interest. An institution that cannot show why a decision was admissible is an institution carrying an unpriced liability, and liabilities like that do not stay unpriced forever.

I am describing a direction, not a statute. I will not pretend to a precise inventory of laws and rulings that does not yet exist in settled form, and you should distrust anyone who offers you one. But the vector is unmistakable, because all four forces point the same way. The party that deploys an automated decision should be able to produce, on demand, the account of why that decision was admissible — the evidence it actually consulted, the rules that were active when it was made, and enough state to let someone reproduce it. The burden is moving toward the only party that ever had the means to discharge it. That is not a radical proposition. It is the burden returning to where it can actually be lifted.

This connects to a distinction worth holding clearly. An institution can be entirely authorized to make a decision and still owe you an account of it. Authorization answers whether you were allowed to act. Legitimacy answers whether anyone should accept the result — and acceptance is exactly what is decaying. As the burden shifts, the question stops being were you permitted and becomes can you show your work. "Trust us" was always an answer about authorization. It was never an answer about legitimacy, and it is wearing out as a substitute for one.

The practical implication is unglamorous and, I think, decisive. The institutions that build for this now — that instrument their decisions so they can hand over a contestable record rather than a shrug — are the ones that will still be trusted when the default finishes turning over. A Decision Receipt that carries its evidence, its frozen rules, and its replayable state is not a compliance cost. It is the form an account takes once the burden has moved and "trust us" no longer counts as one. The institutions that wait will be the ones explaining, after the fact and under far worse conditions, why they cannot produce what everyone has come to expect. The burden is moving. The only choice left is whether you are ready to carry the part of it that is becoming yours.

— Dispatches · Summit Cognitive