The speed defense

Whenever someone proposes that a consequential automated decision should be able to show its work, a particular objection arrives almost on reflex. That would slow everything down. The volumes are enormous, the latency budgets are unforgiving, the margins are thin, and the suggestion that each decision carry a reconstructable record sounds like an invitation to grind the whole machine to a halt. The objection is delivered as a practical constraint, the kind of thing only an amateur would fail to anticipate. It is, more often, the last line of defense — and like most last lines, it tells you where the real position is.

The speed defense works by smuggling in a premise it never states: that accountability is a tax levied at the moment of decision, paid in milliseconds, deducted from throughput. If that were true, the trade would be real and sometimes worth refusing. But it is mostly not true. The expensive part of being able to reconstruct a decision is not making the decision more slowly. It is having built the system, beforehand, so that the evidence it used, the rule it applied, and the state it ran against are captured as a byproduct rather than reconstructed under subpoena. That cost is paid in architecture, once, in advance. It is not paid per decision, and it is not paid in latency.

Confuse those two costs and the speed defense looks unanswerable. Separate them and it nearly evaporates. A system that writes down what it did as it does it is not meaningfully slower than one that does not; it is merely more honest about what it already had in hand. The decision was made on some evidence, under some rule, against some state. Those things existed at the moment of the decision or the decision was not made at all. Recording them is not adding a step. It is declining to throw away the step you just took.

Speed was never traded against accountability. It was traded against the work of building a system that could be accountable without slowing down — work some chose not to do, and now defend not doing.

This is why the speed defense is so often raised by people who have not actually measured it. It is a forecast offered as a fact: it would be slow, surely, obviously, so there is no need to check. The cases where someone does check tend to be quietly embarrassing for the objection. The overhead of emitting a structured record alongside a decision is, in most real systems, a rounding error next to the work of the decision itself. What is expensive is retrofitting that capacity onto a system that was built to forget — and that expense is precisely the thing the speed defense exists to avoid admitting.

What the objection protects

Strip the speed defense down and you find that it rarely protects speed. It protects the comfort of not having to be right on the record. A system that cannot reconstruct its decisions cannot be cleanly proven wrong, and there is a real, if rarely spoken, institutional value in that. Illegibility is a kind of insurance. As long as "the system decided" cannot be unpacked, no individual decision can be isolated, examined, and held up as an error. The speed defense keeps that insurance in force while dressing it as engineering prudence.

You can test this. Offer the objector a version that costs nothing in latency — the record emitted asynchronously, off the critical path, never blocking a single user. If the objection was really about speed, it should dissolve. Often it does not. It migrates: now the concern is storage, or liability, or the danger of giving people something to argue with. That migration is the tell. The position was never located where the argument was being fought. Speed was the defensible perimeter around an indefensible preference for not being examined.

The asymmetry it relies on

The speed defense has had a long run because, until recently, the people harmed by an unaccountable decision had no way to price the missing record. They could not see what was not kept, could not demand what they could not name, and so the institution's choice to build for forgetting carried no cost. The defense worked because the other side of the ledger was invisible.

That asymmetry is closing. As reconstructable decisions become possible — and, in a growing number of domains, expected — the absence of a record stops reading as efficiency and starts reading as a choice. "We move too fast to show our work" survives only as long as no one nearby is moving just as fast and showing it. The moment a competitor, a regulator, or a single well-prepared plaintiff demonstrates that the two are compatible, the speed defense inverts. It stops being evidence of discipline and becomes evidence of what you decided not to build, and why. The institutions that internalize this early will spend the architecture cost once, on their own schedule. The rest will spend it later, under worse terms, while still insisting it could not be done.

— Dispatches · Summit Cognitive