A checklist protects against forgetting, not against being wrong

The failure it is designed for

There is a particular kind of error in which somebody knows exactly what to do, has done it hundreds of times, and does not do it on this occasion. Nothing was misjudged. A step that was fully understood simply did not get executed, usually because attention was elsewhere and the sequence was interrupted.

That is the failure a checklist addresses, and it is worth being precise about it because the tool is frequently deployed against a different problem. A list cannot tell you that you have diagnosed the situation incorrectly, chosen the wrong procedure, or missed something that was never on anyone’s list. It handles omission of the known.

Why memory fails exactly when it is most needed

Working memory is small and interruption is expensive. Under time pressure, with several things in progress, the steps that get dropped are disproportionately the routine ones — precisely because they are routine and therefore attract no attention. Familiarity, which makes a task easy, also makes it invisible.

The pattern is worst for tasks with a stable sequence performed by people who are highly practised. A novice is checking constantly because nothing is automatic. The experienced operator is the one who can perform ten steps without noticing that they performed nine, and this is not carelessness but a straightforward property of automatic behaviour.

Two different objects that share a name

One kind of list is read and performed line by line, used when the sequence must be exact and the situation permits stopping. The other is used after the work is done, as a set of confirmations, and is meant to be short enough to complete without disrupting anything.

They fail differently and mixing them ruins both. A confirmation list that grows into a procedure becomes too long to use and is completed from memory, which defeats its purpose entirely. A procedure list used as a confirmation invites people to tick items they did not verify. Most unusable lists in circulation are the result of somebody adding to one type until it became the other.

The evidence is genuinely mixed and instructive

The best-known application is in surgery, where early studies reported striking reductions in complications and the idea spread rapidly on the strength of them. Later evaluations, including large population-level analyses of mandated implementations, found much smaller effects and in some cases none.

That divergence is the interesting part rather than an embarrassment. The leading explanations are that the early studies involved intensive support and engaged teams, that mandated rollouts often produced compliance with the paperwork rather than the practice, and that a substantial part of the original benefit may have come from the list forcing a team to speak to each other rather than from the items themselves. All of these are plausible and they are not fully resolved.

What follows from that for anybody using one

If a meaningful share of the effect comes from the pause and the conversation rather than the content, then the way a list is used matters as much as what is on it. A list completed silently by one person, retrospectively, is a different intervention from the same list read aloud at a point where the work stops.

It also means a list can be present and doing nothing. Ritual completion is the standard failure mode of any mandatory check, and it is close to undetectable from outside, since the record looks identical either way. Anyone relying on the existence of a list as evidence that the checking happens has confused an artefact with a behaviour.

What a good one looks like

Short, because length is what turns a check into a ritual. Restricted to items where omission is both plausible and consequential, since including obvious steps trains people to skim. Tied to a specific pause in the work rather than to a time of day. And phrased so that each item has a definite answer rather than requiring a judgement, because judgements are the thing this tool does not help with.

The final property is the least observed. A list should be revised when it is discovered that something not on it was missed, and pruned when items prove never to catch anything. A list that has only ever grown is a list nobody has evaluated.