Human-in-the-loop
- Standard term
Human-in-the-loop is a design where a specific AI action is held until a person approves, edits, or rejects it.
The pause is useful when a person can supply missing information or authority while the action can still be stopped. The system should prepare a concrete proposal, so the person can judge the exact amount, wording, or files involved before deciding.
Where you’ll meet it
Your assistant app gathers household bills from email and pays them from your bank account, holding each one until you tap Approve. This month's request looks routine, and you almost tap before opening the bill and seeing that its amount is five times last month's.
What to do
Make the amount and original bill visible before approving, alongside last month's amount for comparison. If your bank supports an enforceable payment limit, use it as an additional control; an approval button alone does not catch a wrong amount.
Easy to mix up
Human-on-the-loop systems run at their own speed while someone watches, ready to stop them. Human-in-the-loop systems hold each action for the person, so their replies set the pace.
Check yourself
Your coding agent asks you to approve about forty file deletions a day. All affected files are disposable project files committed to version control, and restoration has been checked. You've started approving without reading. What fits this limited setup better?
Go deeperA checkpoint helps most where the person knows what the system can't
The phrase predates generative AI and appears in control engineering and simulation, where an operator participates in a feedback loop. OWASP, a security nonprofit, names human approval of high-impact actions as a defense against excessive agency in language-model applications. The EU AI Act also has human-oversight requirements for certain high-risk systems; the applicable duties depend on the system and legal timetable.
Someone who knows that a supplier recently changed banks may spot a payment problem the system cannot resolve. A deterministic check can compare an amount with an invoice consistently, but that does not establish whether the invoice itself is legitimate. The checkpoint should identify what judgment or authority the person adds.
Sources
- LLM06:2025 Excessive Agency, OWASP GenAI Security Project
- Human-in-the-Loop Modeling and Simulation, sigsim.acm.org
- Article 14: Human oversight, ai-act-service-desk.ec.europa.eu