Human in the loop (HITL) is a way of building AI and machine learning systems where a person checks, corrects or approves the AI's work at set points. The AI handles what it is sure about. People handle what it is not. Their corrections then teach the AI to do better next time.
You have probably been a human in the loop without knowing it.
You have already done this job
Around 2008, the internet asked millions of people to type two wobbly words into a box to prove they were not robots. One word was the test. The other often came from an old book or newspaper that scanning software could not read.
Every time you typed it, you fixed a machine's mistake. The Carnegie Mellon team behind the system, called reCAPTCHA, reported in Science that it transcribed text with over 99 percent word accuracy. At the time, it handled more than 4 million hard words a day.
That is human in the loop at its simplest. The machine gets stuck, a person steps in, and the answer goes back into the system.
What human in the loop means
It is a loop with three steps:
- The AI makes a call.
- If it is unsure, or the stakes are high, a person checks it.
- The person's answer goes back into the training data, so the AI is less likely to get stuck there again.
Step three is the part people forget. Someone who only fixes mistakes is doing cleanup. Someone whose fixes flow back into the model is teaching it.
The chatbot that learned manners from people
ChatGPT did not learn to be helpful just by reading the internet. People taught it.
In a 2022 OpenAI study, human reviewers wrote example answers and ranked the model's replies from best to worst. Those rankings trained the model.
People preferred answers from a 1.3 billion parameter model trained this way over the original 175 billion parameter GPT-3.
The smaller model won because people had shown it what a good answer looks like. The method is called reinforcement learning from human feedback, or RLHF, and it is the kind of work our human-in-the-loop teams do for AI companies every day.
Even self-driving cars ask for help
Waymo's robotaxis drive themselves. But when one meets something confusing on the road, it can send a question to a person on its fleet response team. That person does not take the wheel. They answer the question, and the car decides what to do with it.
Most good AI systems work this way. The AI does the driving, and people supply judgment when the AI runs out of it.
When does AI need a human in the loop?
Ask four questions. If the answer to any of them is yes, put a person in the loop.
- Is a mistake costly or hard to undo? A bad song suggestion costs nothing. A wrong loan decision or a missed tumor costs a lot.
- Is the AI unsure? Most models report a confidence score. Send the low scores to people and let the rest run.
- Is the world changing? New slang, new products and new scams appear every week. People spot them before a model does.
- Does the law require it? Under Article 14 of the EU AI Act, high-risk AI must be built so people can oversee it, and override or stop it.
If every answer is no, the AI can usually run alone. Even then, check a sample every so often. Models drift, and a team that never looks will not see it happen.
People can join the loop before launch, by annotating training data (new to that term? start here), or after it, by reviewing hard cases and ranking outputs. Either way, the reviewers have to agree with each other. That is harder than it sounds, as we explain in Your agreement score cannot see half your errors.
It is also why our own reviewers never work from gut feel. At Impact Outsourcing, every review task ships with a written rubric drafted with the client, and we measure how often reviewers agree with each other every week.
The best AI systems still use people. They put them where judgment matters most: the costly calls, the unsure cases and the problems the model has never seen.
Human in the loop FAQs
What is an example of human in the loop?
A content filter that sends posts it is unsure about to a human moderator, whose decision then trains the filter. ChatGPT's training, where people ranked its answers, is another.
Is human in the loop the same as RLHF?
No. RLHF is one kind of human in the loop, used mainly to train chatbots. Human in the loop also covers people reviewing images, labeling text, checking flagged transactions or approving medical findings.
What is the difference between human in the loop and human on the loop?
In the loop, a person checks or approves decisions before they count. On the loop, the AI acts on its own while a person watches and can step in when something looks wrong.
Does human in the loop make AI slower?
Only for the cases people review. If the AI is confident on most items, people see a small share of the work and the system stays fast.