The shape of a fair process

We tend to judge a decision by its result. Was it right? Did it reach the correct answer? This is the natural way to think about fairness, and it is not wrong, but it is only half of the thing. There is a second question, older and in some ways more demanding, that asks not whether the outcome was right but whether the process that produced it was fair. Did the person get notice that a decision was being made? Did they have a chance to respond before it was final? Was the decision made on the evidence and the applicable rules rather than on something extraneous? And could they appeal it if they believed it was wrong? These are questions about shape, not about result, and a decision can pass the test of result while failing every one of them.

That such a thing is possible — that an outcome can be substantively correct and procedurally unjust — is one of the harder ideas to hold onto, because it seems to say that being right is not enough. But it is exactly what it says. A decision that reaches the correct result by a process that gave the affected person no notice, no chance to respond, and no route to appeal has wronged that person even though it happened to land on the right answer. It wronged them by treating them as an object the decision was applied to rather than a party the decision was owed to. The correctness of the outcome is, in a sense, luck. The fairness of the process is the thing that was actually within the decider's control, and it is the thing they failed to provide.

The proceeding behind the verdict

A fair process has a shape, and the shape is familiar because we inherited it from centuries of learning what happens when it is absent. Notice comes first: the affected person is told that a decision concerning them is being made, before it is made, so that the decision does not simply arrive as a completed fact. Then a chance to respond: the person can offer their side, correct an error in the evidence, say the thing that the decider did not know. Then a decision made on the record — on the evidence that was actually presented and the rules that actually applied, not on a hunch or an extraneous fact or a consideration the person never had a chance to address. And finally a way to appeal, so that a wrong decision is not also a final one.

Automated decisions are structurally prone to skipping all of this. They are built to deliver verdicts, and a verdict is the easy part; it is the proceeding that is expensive. So the system produces the outcome — the denial, the flag, the score that closes a door — and there is no notice, because notice would slow things down; no chance to respond, because responding assumes a conversation the system was not built to have; no appeal, because appeal assumes a proceeding, and there was no proceeding, only a result. The person receives a verdict with nothing behind it. They were not given a bad hearing. They were given no hearing, and told the answer.

A verdict with no proceeding behind it is not a fast decision. It is the absence of a decision, wearing the result like a mask.

The record is the proof the process happened

Here is where the record becomes not a nicety but the whole hinge of the matter. A fair process is a thing that either happened or did not, and the only way to tell the difference afterward is by what it left behind. If the process was real — if notice was given, if a response was invited and considered, if the decision was made on the stated evidence and rules, if an appeal was possible — then each of those steps produced something, and the something is the record. If the process was skipped, there is nothing. And a process that leaves no trace cannot be shown to have occurred, which for every practical purpose means it did not occur, because a fairness no one can demonstrate is indistinguishable from a fairness that was never there.

This is the point people miss when they treat the record as mere documentation, an administrative burden bolted onto the real work of deciding. The record is not a description of the fair process. For an automated decision, it is very nearly the fair process itself, because the process consists precisely of the steps that leave traces — the notice sent, the response captured, the evidence and rules the decision actually rested on, the route to appeal held open. Take away the trace and you have not merely lost the evidence of fairness; you have removed the thing that made the process a process rather than a bare result. The proceeding and its record are, in the automated case, almost the same object seen from two sides.

The doctrine of admissible reality frames a fair process this way on purpose: not as a set of good intentions the system holds, but as a set of steps the system can demonstrate, each one leaving something a later party can inspect. Did the person get notice? Show me the notice. Were they able to respond? Show me the response and show me that it was considered. Was the decision made on the evidence and the rules? Show me which evidence, and which rules, as they stood at the time. Could they appeal? Show me the door, and show me it was open. A system that can answer all four has given someone a fair process. A system that can answer none has given them a verdict and called it justice — and the difference between the two is exactly the difference between a record and a silence.

So when an automated decision affects someone, the question to press is not only whether it got the answer right. It is whether there was a proceeding behind the answer, and whether that proceeding left the trace that would let anyone show it happened. If the record is there, the person was a party to a process. If it is not, they were merely the place where a result was applied — and no correctness of the result can make that fair.

— Dispatches · Summit Cognitive