Human oversight is the arrangement by which a person can intervene meaningfully in what an AI system does, and it is the control everybody claims and few implement. It is also the one the EU AI Act names directly for high risk systems, so it is moving from good practice to obligation. The three arrangements go by the names human in the loop, human on the loop and human out of the loop in the older literature on automation, and the names matter less than the conditions under which each one works.
The sections below describe the three arrangements, say what makes a review real, set out how a product is designed for oversight, state the legal obligation and end with a measurement anyone can make in fifteen minutes.
The three arrangements
A human decides, the system advises. The model produces a suggestion and a person makes the call. Safest, most expensive, and the right arrangement where a wrong decision is costly and irreversible.
A human reviews, the system decides. The system acts and a person checks, either every case or a sample. Cheaper, and it only works while the reviewer retains the ability and the time to disagree.
A human can stop it. The system runs unsupervised and somebody is watching the aggregate, able to intervene or switch it off. Appropriate at high volume and low stakes, and it depends on having something worth watching, which is a monitoring problem before it is an oversight one.
The mistake is claiming the first, staffing the second and operating the third.
What makes review real
The second arrangement is the common one, and four conditions decide whether it is oversight or a signature.
The reviewer can judge the output. Somebody checking a legal summary needs to be able to read the contract. A reviewer without the expertise is a delay and never a control.
They have the time. Review capacity is the binding constraint, and it is what quietly turns oversight into rubber stamping. If the volume is a hundred an hour, the review is a glance.
Disagreeing is expected and cheap. If overriding the system is slower, requires justification or is treated as an error, people stop. What is measured decides this more than what is written in the policy.
They see what the system is unsure about. Confidence, the sources used, or the cases flagged as unusual. A bare answer gives a reviewer nothing to work with.
Designing for oversight
Those conditions are met by design, and rarely by staffing. The work is routed by risk, so that the low confidence, the unusual and the high consequence go to a person and the rest go through. That concentrates a fixed amount of attention where it changes outcomes.
The override is made the easy path in the interface. The reasoning, or at least the sources, is shown so that review is possible. And the override rate is measured, because a rate near zero usually means the reviewer has stopped reading, and only occasionally that the system is perfect.
The obligation under the EU AI Act
The EU AI Act requires that high risk systems be designed so a person can oversee them effectively, which explicitly includes understanding the system's limits, staying alert to automation bias, being able to disregard the output and being able to stop it. That is a design requirement, and it lands on the product team. The high risk duties were pushed back to December 2027 by the amending regulation of mid 2026, which changes the date and nothing about what will be asked.
Practise this
The exercise needs one feature where a person approves or edits model output, a stopwatch and fifteen minutes.
Watch ten consecutive items go through review, or run the review as the reviewer. For each item record the seconds spent on it, what the reviewer could actually see at the moment of deciding and whether they changed anything. Then estimate separately how long a genuine disagreement would take, counting the time to open the source, form a view and write whatever justification the process asks for.
Feature ..........................................
Who reviews ......................................
Items per shift ..................................
Item Seconds What was on screen Changed anything?
1 ______ ____________________ Y / N
2 ______ ____________________ Y / N
...
10 ______ ____________________ Y / N
Median seconds an item actually gets ______
Seconds a considered disagreement would take ______
Override rate across the ten ______ %
Read the last three numbers together. Where disagreeing costs more time than the median item is given, the reviewer cannot disagree at the pace the queue demands, and an override rate near zero is the arithmetic and no evidence that the system is right. Look just as hard at the middle column, because a reviewer shown a bare answer with no sources and no confidence has nothing to disagree with even when the time is there.
Fifteen minutes with a stopwatch says whether the oversight is a control or a signature, and it is the same measurement anybody auditing the team would make.