The half-truth of the average

When an institution wants to reassure you that the system deciding your case is sound, it almost always reaches for an average. The model is accurate. The error rate is low. Outcomes have improved across the board. Each of these statements may be entirely true, and each of them may have nothing whatever to say about what happened to you. This is not a trick, exactly, though it functions like one. It is a category confusion so common we have stopped noticing it: the quiet substitution of a fact about a population for a claim about a person, offered in the moment when only the person's claim is at issue.

An average is a true sentence about a crowd. It describes the crowd accurately and describes no one in it. The system that is right ninety-something times out of a hundred is, by the same arithmetic, wrong some number of times, and the average is perfectly indifferent to whether you were one of them. It cannot be otherwise; that is what aggregation is for. To compute the average you must throw away exactly the information that distinguishes your case from every other — and then the institution hands you back the thing it built by discarding you and calls it an answer to your question.

The metric is silent precisely where you need it to speak

Consider what it would mean for the average to actually exonerate a particular decision. It would have to contain that decision — to know which case was yours, what went into it, and whether the process that produced your outcome was the same sound process that produced the good aggregate. But the average knows none of this. It is a summary statistic; it has forgotten the individuals on principle. So when it is offered as a defense of your case, it is being asked to do a job it was specifically designed to be incapable of doing. The aggregate measures the system's behavior over many decisions. Your grievance is about one. These are different objects, and no quantity of the first ever adds up to an account of the second.

This is why "the model is accurate" is not a reply to "the model was wrong about me." Both can be true at once, and their simultaneous truth is not a paradox to be resolved but the ordinary condition of any system that decides at scale. A high overall accuracy and a wrong individual outcome do not contradict each other; they coexist comfortably, and the more confident the aggregate, the more thoroughly it can bury the cases where it failed. The average does not just fail to defend the bad decision — it actively conceals it, folding the failure into a number that reads as success.

The average is the institution's strongest evidence and the individual's weakest standing. It speaks for the system and over the person — and the person is the one who was harmed.

Accountability lives where the average cannot reach

The deeper problem is not statistical but jurisdictional. The average and the individual decision belong to different courts, and only one of them is the court where harm is adjudicated. No one is ever harmed by an error rate. People are harmed by particular decisions — this denial, this flag, this score applied to this person on this day — and a particular decision can only be defended or refuted on its own terms. Was the right evidence in front of it. Were the active rules followed. Would the same inputs, run again, produce the same result, and was that result warranted. These are questions about one decision, and the average is structurally incapable of answering any of them, because it has already dissolved the one decision into the many.

What this means in practice is that population-level performance, however good, can never supply individual-level standing. The person who wants to contest the decision made about them does not need to know how the system performs in general; they need to know how it performed in their case, and that is a different record entirely. It is the record of the actual inputs, the actual rules in force, the actual path from those to this outcome — the things that make a single decision answerable without reference to any other. An institution that can produce a strong average but cannot produce that single account has armed itself to reassure the public and disarmed the individual at the same stroke. It can tell the world it is right while being unable to tell you why it was right about you.

So the next time a system's defenders lead with how often it is correct, notice what the number is being used to do. It is being used to change the subject — from your decision, which they may not be able to defend, to the population's outcomes, which they can. The honest position is not that averages are worthless; they are how you measure whether a system is fit to deploy at all. The honest position is that they settle nothing about any case inside them. Fitness for the crowd and justice for the person are separate questions, and the average answers only the first. Accountability begins exactly where the average ends: at the level of the one decision, where the person stands, and where no statistic about everyone else can be made to speak for them.

— Dispatches · Summit Cognitive