Human in the Loop

Human in the loop describes an arrangement where a person sits between an AI system and its consequences. The system suggests, scores, drafts or ranks. A person decides.

It is one of the most claimed and least verified phrases in AI procurement, because it can mean anything from "every output is approved by a named individual who regularly overrules it" to "a dashboard exists that nobody has opened since March".

Why it matters legally

Where a decision produces legal or similarly significant effects and is made solely by automated means, UK GDPR gives individuals specific rights, including to obtain human intervention and to contest the outcome. Rejection from a job is generally treated as significant. So is a credit decision.

The word doing the work there is solely. A genuine human review changes the legal character of the decision. A nominal one does not, and a regulator or tribunal will look at whether the person actually had the information, the authority and the time to reach a different conclusion.

Under the EU AI Act, deployers of high-risk systems must assign human oversight to someone with the competence, training and authority to intervene.

What makes it real rather than nominal

A named person, not a role. Oversight assigned to a department is assigned to nobody.

Authority to overrule, actually used. If nobody has ever overturned the system, either it is perfect or the review is theatre. It is not perfect.

Enough information to decide. A reviewer shown a score and no reasoning cannot meaningfully review anything.

Enough time. Ninety decisions an hour is not review, it is clicking.

A record. Who reviewed, when, what they concluded. Without it you cannot demonstrate the review happened.

Where to put it

Not everywhere. Human review is expensive and applying it to everything means applying it carelessly to the things that matter.

Put it where a wrong answer has consequences for a person or for the business: hiring, credit, pricing for individuals, access to a service, anything leaving the building with your name on it, anything going to a regulator. Leave it off internal drafting, summarising and search, where the cost of an error is that somebody notices and fixes it.

The question to ask a supplier

"Which specific decisions require a human, and can that person actually override the system?"

Vagueness here is informative. So is a system where overriding requires more effort than accepting.