The honest interface

Somewhere a model produces a number, and the number is hedged, conditional, drenched in the uncertainty that any honest estimate carries. Then the number travels to a screen, and on the screen a person reads it and acts. Everything that matters about whether the system is honest happens in that last short journey — between the estimate and the eye. The model can be as careful as you like about what it knows and does not know, but the interface is where that care is either preserved or thrown away, and the interface is the only part the human ever sees.

This is rarely treated as a serious question. We think of the interface as packaging — the cosmetic layer that wraps the real work of the model in something presentable. But the interface is not packaging. It is the point of contact between a machine's estimate and a human's action, and it is the last place where the truth about that estimate can be told before someone relies on it. What the screen chooses to show, and what it chooses to omit, decides whether the person acts on a fact or on a fact's confident impersonation.

There are, broadly, two ways to build that screen, and they look almost identical until you ask what they are doing to the human. One shows the output as a bare result: a score, a verdict, a clean recommendation, presented with the flat authority of something already settled. The other shows the output as a claim that rests on something — surfacing the uncertainty the model actually carried, the inputs the result was built from, the basis on which it was reached. The first hides the seams. The second shows them. And the difference between them is not a matter of taste.

What the bare output conceals

A model's output is, at bottom, an estimate under uncertainty. It was produced from particular inputs, under particular conditions, with a particular and usually wide margin of doubt. An interface that shows only the result strips all of that away. It presents the estimate as if it were a measurement — as if a guess, however well-founded, were the same kind of thing as a fact read off an instrument. The hedging that lived inside the model does not survive the trip to the screen, and the human, seeing only the clean number, treats it as cleaner than it ever was.

This is what I mean by laundering. The bare interface takes something provisional and returns it scrubbed of its provisionality. The doubt does not disappear; it is simply not shown, which means the human absorbs none of it. A person looking at a confident figure with no indication of its uncertainty has no way to weigh it as the model itself would have weighed it. The interface has made the estimate more authoritative than the model that produced it — and it has done this silently, as a side effect of showing less.

An interface that shows only the answer has not simplified the decision. It has impersonated certainty the model never had, and handed the impersonation to the one person obliged to act.

An honest interface keeps the decision legible

The honest interface does the harder thing. It shows the output and what the output rests on, in the same glance, so that the human acting on the decision can see not just the answer but the answer's footing. It surfaces the uncertainty the model carried, rather than absorbing it. It exposes the inputs the result was built from, so that a result resting on thin or stale evidence announces itself as such. It points to the basis — the rule, the threshold, the reasoning — so that the verdict can be read as a claim with a foundation rather than a pronouncement from nowhere.

This is not about clutter, and it is not about distrust for its own sake. It is about keeping the decision legible at the exact moment it is being relied upon. The person at the screen is the last human in the chain, the one who turns an estimate into an action with consequences. They are precisely the person who most needs to see the footing, because they are the one who will be answerable for the step. An interface that hides the basis from them does not make their job easier; it makes their judgment impossible, by handing them a conclusion with no way to weigh it.

And this is why the choice is ethical, not aesthetic. To design the bare interface is to decide that the human should act on a result without seeing what it rests on — to decide, in effect, that the convenience of a clean screen outweighs the person's capacity to judge. That is a moral choice wearing the costume of a layout decision. The honest interface makes the opposite choice: it accepts a busier screen so that the human keeps the standing to disagree. It treats the moment of contact as a moment of accountability rather than a moment of delivery.

The clean screen will always be tempting, because certainty is easier to look at than doubt, and a single confident number is easier to ship than a number with its footing attached. But the screen is where the system meets the person, and a system that tells the truth everywhere except the one place a human reads it has not told the truth at all. The honest interface is the discipline of letting the doubt through — of refusing to launder the guess into a verdict on the last screen before someone has to act.

— Dispatches · Summit Cognitive