Reasonable, on what record?

The most durable standard the law has for a decision is also the least flattering to it. We do not, for most consequential choices, ask whether the decider was correct. We ask whether the decision was reasonable — whether a person, knowing what this person knew, weighing what they had in front of them, could have arrived here without abandoning their judgment. It is a forgiving test on purpose. It lets defensible decisions stand even when they turn out wrong, and it refuses to bless arbitrary ones even when they happen to land right. Reasonableness is how the law makes peace with the fact that good process and good outcome are not the same thing.

The test has lasted because it travels well. It governs the administrator denying a permit, the doctor choosing a course, the officer making a stop, the trustee allocating a fund. In each case it asks the same thing and it asks it in the same way: not what the right answer was in hindsight, but whether the path to this answer holds up when laid out and examined. Reasonableness is a claim about a process, and a claim about a process is only as good as your ability to see the process. That last point is doing more work than it appears to. It is, in fact, the whole essay.

Because here is what happens when the decider is a system rather than a person. The standard does not change — the law still wants to know if the decision was reasonable. But the thing the standard reaches for is no longer there. There is no remembered deliberation to recount, no contemporaneous file, no person who can be put under oath and asked what they were thinking. There is an output, and there is whatever the system happened to retain on its way to producing it. If that is nothing, then reasonableness has not been satisfied and it has not been failed. It has simply been rendered unanswerable. You cannot weigh the reasonableness of a process you cannot reconstruct.

Reasonableness was never a verdict on the answer. It was a verdict on the path — and a path you cannot retrace is not a path you can defend.

This is the quiet collapse, and it is worse than a loss. A decision found unreasonable is a decision the law has examined and rejected; the standard worked. A decision that cannot be assessed for reasonableness has slipped out from under the standard entirely. It is not defensible and it is not indefensible. It is unreviewable, which in a system built on review is a category that should not exist for consequential choices, and yet is exactly the category an unrecorded automated decision occupies by default.

The record is not the paperwork

It is tempting to read all of this as a documentation problem — keep better logs, save the inputs, and the standard has something to chew on again. That is half right and dangerously incomplete. Reasonableness does not ask what data went in. It asks what the decider made of it: which evidence was treated as load-bearing, which rule was actually applied, what state of the world the decision ran against, and whether the move from that evidence under that rule to this conclusion is one a sound process could make. A log of inputs and an output, with the middle missing, does not answer that. It is the courtroom equivalent of a witness who will tell you what they saw and what they concluded but goes silent on how one became the other.

So the record reasonableness requires is not a transcript of activity. It is a reconstruction of judgment: enough of the evidence, the rule, and the prevailing state that the decision can be run again and seen to follow — or seen not to. The difference between those two records is the difference between proving you were busy and proving you were reasonable. Only one of them was ever the standard.

The standard does not bend for the machine

There is a hope, often unspoken, that the law will simply make allowances. That automated decisions are too numerous, too fast, too statistical to hold to a test designed for a single human mind, and that the standard will soften to accommodate them. I would not build on that hope. The pressure runs the other way. When a decision is automated, it is also typically scaled, which means a single unreasonable pattern is not one error but a class of them, applied uniformly, at volume, to people who never met the decider. That is the profile that attracts a standard, not the profile that escapes one. The reasonableness test has bent to accommodate new kinds of deciders before, but it has bent by demanding they become legible, never by agreeing to stop looking.

Which leaves the builders of these systems with a choice they can make now or have made for them later. Reasonableness is coming for automated decisions exactly as it came for every prior species of authority, and it will ask its one unchanging question. The only variable is whether, when it asks, there is a record capable of answering — a reconstruction of the path, not a receipt for the activity. A system built to keep that record can be found reasonable or unreasonable, and either is survivable. A system built to forget can only be found unreviewable, and that is the one verdict an institution should fear, because it is the one the law has historically refused to let stand.

— Dispatches · Summit Cognitive