Human in the loop

In plain English

Think of a bank teller who needs a supervisor's signature before releasing a large cash withdrawal. The teller prepares the transaction; nothing happens until the supervisor signs. In human in the loop design, the AI is the teller and a person is the supervisor.

Three features define it:

• Per decision. A person looks at each case, not a weekly summary.
• Before effect. The review happens before the outcome reaches anyone.
• Real authority. The reviewer can change or reject the AI's answer.

It is one of several oversight models. The others, human on the loop and human out of the loop, are compared in human in the loop vs human on the loop.

Example

Ashgrove Housing Authority, a fictional public body, uses AI to flag benefit applications for possible fraud. A caseworker decides every flagged case.

The design gives reviewers what they need:

• The AI shows its main factors and a confidence score, not just a verdict.
• For a weekly sample, caseworkers record their own view before seeing the AI's score.
• Caseworkers are measured on decision quality, not cases cleared per hour.
• Overrides are tracked, and nobody is penalized for disagreeing.

After three months, the override rate is 0.4%. Caseworkers also approved several planted test cases the AI had wrongly flagged. That is automation bias. Ashgrove gives reviewers more time per case, retrains them and repeats the test.

Why it matters

Origin. The EU expert group's 2019 Ethics Guidelines for Trustworthy AI describe human in the loop as the ability to intervene in every decision cycle. They note this is often neither possible nor desirable.

EU AI Act. Art. 14 does not mandate HITL. It requires oversight that lets people stay aware of automation bias (Art. 14(4)(b)) and override or reverse an output (Art. 14(4)(d)). Deployers must assign competent people with the authority to do so (Art. 26(2)). A human reviewer never changes a system's high-risk classification.

GDPR. Meaningful review of each case can take a decision outside "solely automated" under Art. 22. Token review cannot.

Other uses. OWASP recommends human-in-the-loop approval for high-impact actions by AI agents. Colorado's SB 26-189, from 1 Jan 2027, requires deployers to offer human review of adverse decisions, where commercially reasonable. That is review after the decision, not HITL.

How the AIGP exam tests it

Oversight design is taught in lesson III.A.4.

Common traps:

• After the fact. Reviewing cases only after complaints is remediation, not human in the loop.
• The double rubber stamp. A second reviewer who also sees the AI's answer first fixes nothing.
• Near-zero overrides. That is a warning sign of automation bias, not proof of a good model.
• HITL everywhere. Requiring approval for millions of low-stakes decisions causes delays and rubber-stamping.

The tell: if nothing takes effect until a person approves it, the design is human in the loop. Then check whether that person can genuinely say no. See also human oversight.