In an AI context

Keeping a human in the loop is the most common answer to AI risk and the most commonly hollow one. The phrase describes a position in a workflow. It does not say whether the person in that position can catch a failure. A reviewer who sees a confident recommendation, has thirty seconds to clear it and is measured on throughput will agree with the system almost every time. That is automation bias, and it turns oversight into a signature.

The governance question is therefore not whether a human is involved but whether the involvement is meaningful: whether the person has the competence to judge the output, the information to see when it is wrong, the time to look, and the authority to change the outcome.

In, on and over the loop

Three terms describe different degrees of involvement. Usage varies, so define them locally.

  • Human-in-the-loop: a person approves each decision or action before it takes effect.
  • Human-on-the-loop: the system acts on its own while a person monitors and can intervene.
  • Human-over-the-loop, sometimes called human-in-command: a person sets the parameters, reviews outcomes in aggregate and decides whether the system should operate at all.

The right choice depends on how reversible the action is. Where a decision cannot be undone, or harms someone before anyone sees it, oversight on the loop arrives too late.

What the law requires

Article 14 of the EU AI Act requires high-risk systems to be designed so that the people overseeing them can understand the system's capacities and limits, stay aware of automation bias, interpret its output, decide not to use it, and intervene or stop it. Article 26 requires deployers to assign that oversight to people with the necessary competence, training and authority.

In UK data protection law, a decision with no meaningful human involvement is solely automated, and since February 2026 Articles 22A to 22D of the UK GDPR attach safeguards to significant decisions of that kind; see automated decision-making. A rubber-stamp review does not take a decision outside those rules.

Oversight for agents

Per-decision review does not scale to an agentic AI system that takes dozens of actions per task. The workable pattern is to decide in advance which actions are irreversible or consequential (sending, paying, deleting, publishing) and require confirmation for those. Reviewers can also be attacked: OWASP's agentic list includes human-agent trust exploitation, which covers flooding approvers with requests until they stop reading them.

What good looks like

  • Defined checkpoints: the decisions or actions that need a person, chosen by reversibility and impact.
  • Competent, authorised reviewers, trained on the system's known failure modes and able to overrule it without escalation.
  • Designed for disagreement: the reviewer sees the evidence, not only the recommendation, and where possible forms a view first.
  • A workload that allows review: volumes and time per case that make real scrutiny possible.
  • Measured: override rates, review times and sampled agreements, watched for drift towards rubber-stamping.

The register entry for oversight that exists only on paper is inadequate human oversight design; the control is human-in-the-loop review.

Frequently asked questions

What does human-in-the-loop mean? A design in which a person reviews or approves an AI system's output or action before it takes effect.

What is the difference between human-in-the-loop and human-on-the-loop? In the loop, a person approves each decision before it takes effect. On the loop, the system acts and a person monitors and can intervene.

Does a human in the loop make an AI decision lawful? Not on its own. The involvement has to be meaningful, and a reviewer who rubber-stamps the output does not take a decision outside the UK GDPR's automated decision-making rules.

Does the EU AI Act require human oversight? For high-risk systems, yes. Article 14 requires it to be designed in, and Article 26 requires deployers to assign it to competent people with authority.