AI glossary · Basic · Agents · Safety & alignment
Human-in-the-loop
Human-in-the-loop (HITL) is a design approach in which a person steps in partway through an AI's work to check, approve or correct it. Instead of leaving everything to the AI, a person makes the call at each critical point.
About this term
- Level
- Basic · Technical terms you often meet in AI news
- Category
- Perception & action
- Fields
- Agents, Safety & alignment
- Added
- 2026-10-08
How it works
Before the AI takes an action, it shows the person its plan or result, saying 'This is what I'm about to do.' The person approves it, rejects it or points out what to change. If rejected or asked for changes, the AI reflects that feedback and tries again. You can have every step checked, or have only hard-to-reverse tasks such as deleting files, making payments or sending emails checked.
Why it matters
AI sometimes produces hallucinations, plausible but wrong answers, and sometimes misunderstands what the user intends. As AI agents now go as far as taking actions, such mistakes can immediately cause real-world harm. A single check by a person at a critical moment can greatly reduce these risks. Corrections left by people are also sometimes used as training data to make the AI better.
Things to know
More human checks mean more safety, but also more slowness and hassle. So it is important to design which points to check based on the level of risk. If approval requests pop up too often, people may click them out of habit without reading, which can actually render the safeguard ineffective.
Example
Coding agents commonly ask 'May I run this command?' before changing code or running terminal commands, and proceed only after the user approves. Customer service chatbots also often hand sensitive matters such as refunds to a human agent who makes the final decision.