Measure · A metric I propose

Contestation Latency

How much time and context does a human actually have to disagree with the machine? I call that contestation latency. When it is three seconds, the oversight is a turnstile, not a judgment.

Most arguments about human oversight skip the most basic question: how long does the human actually get? A reviewer with three seconds per case, no access to the inputs, and a queue to the horizon is not overseeing anything. They are processing a line. Contestation latency is my name for the gap that decides this: the time and context a person realistically has to mount a disagreement before the decision moves on without them.

How much time does a human need to disagree?

More than the system usually gives. Disagreeing with a confident machine is not a reflex, it is work. You have to notice something is off, pull the relevant inputs, form an alternative, and be willing to defend it. Three seconds buys the noticing, on a good day. The honest version takes minutes, sometimes a conversation. When the time allotted sits far below the time the judgment needs, the human is set up to rubber-stamp, and then blamed when the stamp was wrong.

What is contestation latency, exactly?

A proposed measure, not a settled one, offered at v0.1. It pairs two things: the time a reviewer has per decision, and whether they have the context to use that time. A short latency is not automatically bad, a one-second veto on an obvious error is fine. It is bad when the latency is short and the decision is consequential and the inputs are hidden, which is exactly the combination most AI review pipelines produce. I offer the term because you cannot budget for what you refuse to name.

Why latency shrinks as the system scales

Throughput is the enemy of contestation. The entire economic case for automating decisions is volume, and volume means less time per item, by design. So the stronger the business case for the AI, the worse the contestation latency tends to get, until human review is a formality timed in seconds. If you want oversight to survive scale, you have to budget the latency on purpose, because the system will never grant it on its own.