The Confidence You Don't Report

Show notes

In the Toulmin frame, every defensible claim carries a qualifier — the honest statement of how sure you are. It is the difference between "this is pneumonia" and "this is probably pneumonia, and here is what would change my mind." The qualifier is not hedging or weakness; it is the part of an answer that tells you how much weight it can bear. Strip it out and you are left with something that sounds more certain than it has any right to be — a conclusion wearing a confidence it never earned.

Medicine, almost uniquely, has institutionalized the qualifier. A differential diagnosis is a ranked list of competing explanations, kept open on purpose. A confidence interval reports not a point but a range, and the width of that range is the message. "We don't know yet" is a clinically legitimate finding, and "watchful waiting" is a real plan. The discipline trains people to say how sure they are, to name what they have not ruled out, and to leave room for the answer to be wrong. The uncertainty travels with the decision, on the record, where the next clinician — and the patient — can see it.

Machine-learning outputs rarely behave this way. A model hands back a single number, a label, a rank — clean, decisive, and silent about its own doubt. There are no error bars on the screen. There is no option to abstain, no "this case is unlike anything I was trained on," no signal for the distribution it has drifted away from. The confidence the model does not report does not disappear; it is simply transferred, uncalculated, to whoever acts on the output. The clinician inherits a certainty the system never actually had, and inherits the liability for it too.

So this episode asks what is lost when the qualifier is stripped out — and what it would take to put it back. Not a softer model, but an honest one: a system that reports not just an answer but the shape of its own uncertainty, that flags the inputs it has never seen, and that is permitted to decline to decide when declining is the right call. Abstention is a feature, not a failure. A decision that knows the edge of its own competence is more trustworthy than one that answers everything with the same flat confidence. This is the argument of the book Admissible Reality carried into the clinic, where the cost of false certainty is measured in patients.