Publish your error rate

Take the number you already have and give it to the people who don't. You know how often your decision system is wrong — you measured it to build the thing, you watch it to keep the thing running, you cite it in the deck when you argue for the next quarter's budget. The overall accuracy, the false positives, the false negatives, the cases it fumbles, the groups it serves worse: those figures exist, in a dashboard or a validation report, on your side of the wall. On the other side is the person the system just decided about, who has none of them, and who has been handed a verdict with the texture of certainty. Close that gap on purpose. Publish the error rate, honestly and currently, to the people subject to the decision and to the people relying on its output — because the alternative is letting them mistake a fallible judgment for a settled fact, which is a thing you are doing to them whether or not you meant to.

The failure this corrects is quiet and nearly universal, because it is the path of least resistance. A decision system emits a clean output — approved, flagged, declined, high-risk — and the output travels without its uncertainty. Nothing in the interface says this is right about eleven cases in twelve. Nothing says and the twelfth is disproportionately someone like you. The number stays home; only the verdict ships. So the affected person, reasonably, fills the silence with the assumption a confident-looking system invites: that it is reliable, that a machine would not say declined unless declined were true. They cannot weigh what they were never shown. And the honest hard part is that you knew, and they didn't, and the not-telling was a choice you made by default.

The number you have and they don't

Start with who is misled, because it is more people than the obvious one. The affected person is misled, yes — left to treat a probabilistic guess as a determination. But so, very often, is the human placed in the loop to review the machine's output, the supposed safeguard. Give a reviewer a stream of confident verdicts with no error rate attached and they will calibrate to the confidence, not the accuracy. They rubber-stamp, because nothing in front of them signals which cases deserve a second look, and because a system that is right most of the time trains its reviewer to assume it is right this time. The undisclosed error rate does not just deceive the subject; it disarms the safeguard you installed to protect the subject. Both are trusting the system more than the numbers license, because you have the numbers and they have the verdict.

Name this plainly: withholding the error rate is misrepresentation by omission. You do not have to state a falsehood to mislead. Presenting a decision as a clean result, in a system you know to be fallible, without disclosing how fallible — that is a representation. It says, by everything it declines to say, treat this as certain. A verdict shipped without its error characteristics is not neutral silence; it is an implied claim of reliability that your own measurements do not support. The person is entitled to know that the thing deciding about them is sometimes wrong, and roughly how often, and in which direction, before they decide how much of their life to stake on believing it.

A system that hides its error rate is not being modest; it is letting you mistake its guesses for facts, which is the most useful lie it can tell.

Publish it so trust can be calibrated

The directive is the correction: disclose honest, current error characteristics to the people subject to the system and the people relying on its outputs, so that trust can be sized to reality instead of to presentation. This is the population-level companion to showing confidence on the individual decision. Per-decision confidence tells one person how sure the system is about their case; the published error rate tells everyone how often the system is right across all the cases like theirs. You need both, and the population number is the one that lets a person situate their own verdict — to know whether they landed inside a system that is nearly always right or one that is right just often enough to be dangerous.

Honest means the metrics that matter, not the flattering one. And current means maintained, because a system drifts. The world it was measured against moves, the inputs shift, the model gets retuned, and an error rate published once and never revisited becomes a stale reassurance — a true statement about a system that no longer exists. So the disclosure carries a date and gets refreshed as the system changes, the way you would refresh any standing claim whose premises decay. A published error rate is not a plaque you mount and forget. It is a live figure that tracks the thing it describes, or it is a decoration that happens to have once been accurate.

Crucially, disaggregate. A single overall accuracy number can be excellent and still conceal that the system fails two or three times as often for one group as another — the aggregate averages the harm away, which is exactly what makes the headline number so comfortable and so misleading. If performance varies across populations, and it almost always does, then the honest disclosure reports where it varies. The person deciding whether to trust a verdict about themselves needs the error rate for people like them, not the one flattered by everyone else. An aggregate that hides an uneven distribution of failure is not transparency; it is a more sophisticated way to keep the number home.

The honest hard parts

An error rate can be gamed, and you should assume yours will be tempted toward it. There is always a flattering metric to feature and a favorable slice to average over — accuracy where recall would embarrass, the aggregate where the disaggregation would indict. The discipline is to publish the metrics that actually bear on the harm the system can do, and the breakdown that reveals where the harm concentrates, rather than the single reassuring figure. If a number can only be defended by choosing which number to show, it was cherry-picked, and cherry-picking an error rate is just misrepresentation by omission wearing a lab coat.

There are legitimate limits, and they are narrow. Some disclosure genuinely helps adversaries game the system — publishing the exact operating point of a fraud filter tells fraudsters where the line is — and some detail is properly constrained by security. But these are specific exceptions, not a blanket license to keep the subject in the dark. The test is honest: are you withholding this figure because disclosing it would let a bad actor defeat the system, or because disclosing it would let an affected person question the verdict? The first is a real limit. The second is the very thing you are obligated to overcome, dressed up as the first.

And a published error rate is not absolution. Disclosing that you are frequently wrong does not make being wrong acceptable where the stakes are high; a transparent injustice is still an injustice. Publishing the number does not license the harm — it lets the harm be seen, contested, and priced, which is the precondition for deciding whether the system should be deployed at that error rate at all. Transparency about a failure rate and responsibility for the failures are two obligations, not one. So publish the number, keep it honest, keep it current, break it out where the harm is uneven — and then answer, separately, for whether a system that wrong belongs anywhere near a decision that consequential.

— Dispatches · Summit Cognitive