What is Human-in-the-Loop?

Every AI vendor now promises a human in the loop. Almost none can answer the only question that matters about that promise: what, exactly, can the human in your loop actually do? Ask it and the phrase splits into two very different things, one of which is a control, and the other a costume.

The honest definition

Human-in-the-loop, honestly meant, is this: a human makes or reviews the consequential call, with the authority and the context to disagree. Both halves carry weight. Authority means the human’s no is a real no: they can reject the recommendation without escalating three levels or breaking a process. Context means they can see what the machine saw: the evidence behind the recommendation and, critically, how good that evidence is. A human with authority but no context is guessing with power. A human with context but no authority is a witness.

The theatre version

Then there is the version that ships. A recommendation arrives as a score with no visible reasoning. The reviewer has dozens in the queue and no way to interrogate any of them. Approving takes one click; declining requires a justification the reviewer has no evidence to write, since the evidence was never shown. So they approve, the log records that a human approved, and everyone involved can say a human was in the loop. The human is there to absorb accountability, not to exercise judgement. A human clicking approve on things they cannot inspect is not a human in the loop. It is theatre with an audit trail.

The theatre version is worse than no loop at all, because it launders machine confidence through a human signature. When the call goes wrong, the record shows a person approved it, and the person had no realistic means of doing anything else.

Three tests of a real loop

• The human sees the evidence, and its quality. Every fact behind the recommendation carries a state on the evidence hierarchy: measured, modelled, inferred, stated or unmeasured. A recommendation resting on measured facts should look different, to the reviewer, from one resting on someone’s stated hopes. If the reviewer cannot tell the two apart, they are not reviewing.
• The human can challenge the recommendation. They can trace the conclusion back through its reasoning, ask why this option outranked that one, and reject the answer as a normal act of work rather than an act of courage. A recommendation that cannot be interrogated can only be obeyed or ignored, and neither is oversight.
• The override is recorded, and the outcome scored. When the human overrules the machine, the record keeps who, when, and against what evidence. When reality lands, the outcome is scored against both the recommendation and the choice. This is what a decision audit trail exists to hold, and it calibrates both parties: the machine learns where it is wrong, and the organisation learns when its people beat the model and when they do not.

This is how ONX builds the loop into Decision Rooms: options are ranked by the earliest date they clear every binding constraint, a human always makes the call, and overrides are recorded with who, when and against what evidence. The recommendation is machinery; the decision is a person.

One concrete example

Clearly illustrative, with no customer implied. A matching system ranks candidates for a role. In the theatre version, a recruiter sees five names with scores and approves the shortlist, because the interview slots are booked for Thursday and the scores offer nothing to argue with. In the real version, the reviewer sees why each candidate ranked where they did, with each contributing fact labelled by quality: which claims were measured against the record, which were inferred, which merely stated. She disagrees with one ranking, reorders it, and the system records her override and the grounds. Months later the hire’s outcome is scored against both the machine’s list and hers. One of these processes has a human in the loop. The other has a human near it.

Human oversight and the EU AI Act

The phrase has stopped being merely good practice. The EU AI Act requires high-risk AI systems to be designed for effective human oversight: people who can understand the system’s outputs, stay alert to automation bias, intervene in operation, and overrule the system. Employment-related uses, including candidate matching and screening, sit in the high-risk category under Annex III. The regulation’s direction of travel is precisely the distinction this entry draws: oversight means a human genuinely able to contest the machine, not a human present at the moment of output. ONX takes the strict reading of that category and treats candidate matching in Hiring as high-risk, with human review gates, and keeps its AI explainable and challengeable so the reviewer at the gate has something real to review. None of this is legal advice; it is an operating stance, and the fuller philosophy behind it sits within Enterprise Decision Intelligence.

The test for any system claiming a human in the loop fits in one sentence. Can the human see the evidence, challenge the recommendation, and be remembered for the call they made? If any answer is no, the loop is drawn around the human, not through them.