What we owe the loser

There is a quiet asymmetry in how we judge the decisions that machines now make on our behalf. When the loan is approved, the claim paid, the application advanced, nobody scrutinizes the process. The winner has what they wanted; the system performed; the case is closed. We test these systems, almost without exception, by asking whether the people they favor are satisfied. But the people they favor were never going to complain. The only honest test of a decision is the one posed by the person it ruled against — and that is precisely the test most automated systems are designed never to face.

This matters because legitimacy is not a property of outcomes that please. Any process can please the people it rewards; a coin flip pleases half of them. What separates a decision from an arbitrary act is what it offers to the party who lost — whether that party is given something to do other than absorb the result. A decision that the winner accepts and the loser cannot question is not a decision in any meaningful sense. It is an allocation dressed in the language of judgment.

So I want to be precise about what, exactly, we owe the person on the losing side of an automated decision. Not an apology, not a softer tone, not a more sympathetic chatbot. Three concrete things, each of which the system either provides or does not.

You can tell what a system actually believes about its own fairness by what it hands the person it just rejected.

A record they can actually fight

The first thing we owe the loser is a record built to be contested rather than to be accepted. There is a difference, and it is not cosmetic. A record built to be accepted explains the outcome in fluent, closing sentences; it tells the rejected party what happened and invites them to move on. A record built to be contested exposes its working — the evidence that was in front of the decision, the rules that were active at the moment it was made, enough state to rerun it and watch where it lands. The first kind asks for acquiescence. The second hands over a surface to push against.

A Decision Receipt that the winner never reads is still doing its real work, because its real audience was always the loser. The point of carrying the actual inputs and the frozen rules inside the record is not to reassure the satisfied; it is to arm the dissatisfied. If the rejected party cannot see what the decision saw, they cannot show it saw wrong. Contestability is not a courtesy extended after the fact. It is the difference between a verdict and a wall.

An appeal that lands somewhere

The second thing we owe the loser is a route of appeal that arrives at an actual obligation. It is easy to build the appearance of recourse — a form, a portal, a "request a review" button that drops the objection into a queue no one is required to empty. That is not an appeal. An appeal is a path that ends at someone who must respond, on a clock, with a reason. A complaint that no one is bound to read is not contestability; it is the theater of it, and the rejected party can usually feel the difference long before they can name it.

The test here is structural, not tonal. Does the objection reach a place where a human being, or an accountable process, incurs a duty to answer it? Or does it reach a place where it can be silently aged out, marked resolved, and forgotten? A system that routes appeals into an unread queue has not given the loser a route at all. It has given them the shape of one, which is worse than nothing, because it spends the energy they would have used to fight somewhere real.

Standing to fight

The third thing — the one that holds up the other two — is standing. By standing I mean the recognized position from which a person is entitled to contest a decision and have that contest count. A record you can read and an appeal that lands somewhere are useless if the person on the losing end has no acknowledged right to demand either. Standing is what converts a grievance into a claim. Without it, the rejected party is a supplicant; with it, they are a party.

This is the piece that automation most quietly erodes. When a chain of systems produces an outcome, it is easy for everyone involved to treat the affected person as a recipient of a result rather than as someone with a right to challenge it. The decision simply arrives, the way weather arrives. But a decision is not weather. Someone built the process, someone set the rules, someone deployed it knowing what it would do — and the person it ruled against has standing to make all of them answer. A system that does not recognize that standing has not removed the conflict. It has only removed the loser's ability to enter it.

None of this is hostile to the systems themselves. A decision that is genuinely sound has nothing to fear from a contestable record, a real appeal, and a party with standing — it will survive all three, and emerge more authoritative for having done so. The systems that resist these obligations are, almost always, the ones that suspect they would not hold up. So the measure I would apply to any automated decision is simple, and it has nothing to do with the people it pleased: look at what it left the person it ruled against. If that person was handed a record, a route, and a recognized right to fight, the decision earned its legitimacy. If they were handed a smooth result and a closed door, it did not — whatever the winner thinks of it.

— Dispatches · Summit Cognitive