human in the loop
A design in which a person reviews or approves an AI system’s outputs or actions at set points, or supplies the labels and corrections it learns from.
- A coding agent that asks before editing files or running commands
- A confirmation prompt before an agent calls an outside tool
- A weapon that selects targets but fires only on a human command
- A clinician assisting a model that has too few training examples
- Earliest AI use
- 1988
- Earliest found: NASA expert system to cut the need for a human-in-the-loop.
- Current meaning
- 2019
- EU experts name it as one of three ways to oversee any AI system.
- Opposite
- Out of the loop
- HRW, 2012: robots that select targets and use force with no human input.
In plain words
Human in the loop (HITL) is an AI design in which a person reviews, corrects or approves the system’s work at set points, such as before an agent acts.
A system with a human in the loop stops at set points and waits for a person. The person might approve an action an AI agent has proposed, refuse a tool call, or supply the labels a model learns from. IBM gives the purpose as accuracy, safety, accountability and ethical decision-making, and abbreviates the phrase HITL.
The phrase describes a degree of control. In a 2012 report on autonomous weapons, Human Rights Watch sorted robots into three kinds. A human-in-the-loop weapon uses force only on a person’s command. A human-on-the-loop weapon acts alone while a person watches and can override it. A human-out-of-the-loop weapon needs no person at all. The report warned that thin supervision can leave an on-the-loop weapon effectively out of the loop.1 In 2019 the European Commission’s expert group on AI applied in-the-loop and on-the-loop, plus human-in-command, to AI systems in general.
The idea now shapes law and software. The EU AI Act requires human oversight for high-risk AI systems.2 The Model Context Protocol, a standard for connecting AI applications to tools, says a person should always be able to deny a tool call.
A person in the loop is no guarantee of safety. Ben Green reviewed 41 policies that require human oversight of government algorithms. He found evidence that people cannot perform the oversight those policies ask of them.3 The term also differs from reinforcement learning from human feedback, where people rate a model’s answers to train it, not to approve its actions.
How it works
In an agent, the check sits between the model and its tools. The model asks to run a tool, a step called tool use, such as running a shell command or sending an email. The application pauses and shows a person the tool and its inputs. If the person approves, the call runs; if not, it does not. The MCP specification tells applications to ask for confirmation on sensitive operations and to show tool inputs before calling the server.4
Designers choose which actions need a stop, and too many stops carry their own risk. Anthropic warns that constant prompts can lead to approval fatigue, in which people stop paying attention to what they approve. Its Claude Code tool is read-only by default and asks before changes or commands. Running it in a sandbox, an environment walled off from the rest of the computer, cut prompts by 84% in internal use.5
In machine learning the loop is the training cycle. People label data or take on cases the model finds hard, and their answers feed back into training. In active learning, a common method, the system picks which examples people should label next. Wu and colleagues sort these methods by where people step in: preparing data, guiding training or designing the whole system.6
Origin and first known use
The loop is a control loop, the cycle in which a system senses, decides and acts. A person inside it is part of that cycle. A 1967 NASA contractor report used man in the loop for placing a person in the control loop of boosters. By 1988 NASA researchers used the gender-neutral human-in-the-loop in AI work, for an expert system meant to reduce the need for one. From 2012 the debate over autonomous weapons set humans in, on and out of the loop side by side. In 2019 EU experts applied the terms to AI in general. Machine learning uses it for people who label and correct training data.
- 1967 A NASA contractor report on space vehicle flight control devotes a volume, Man in the loop, to placing a person in the control loop of boosters.R. M. Smith, “Analysis and design of space vehicle flight control systems. Volume X - Man in the loop”, NASA CR-829, General Dynamics, July 1967 ntrs.nasa.gov →: source for 1967
- 1988 NASA Ames researchers describe a knowledge-based system that does some of a human operator’s control work, reducing the need for a human-in-the-loop. Earliest AI use found.B. J. Glass and C. M. Wong, “A knowledge-based approach to identification and adaptation in dynamical systems control”, NASA Ames Research Center, 1988, NTRS record 19890041203 ntrs.nasa.gov →: source for 1988
- 2012 Human Rights Watch and Harvard Law School’s human rights clinic sort robotic weapons into human-in-the-loop, human-on-the-loop and human-out-of-the-loop, by how far people control their use of force.Human Rights Watch and Harvard Law School’s International Human Rights Clinic, “Losing Humanity: The Case against Killer Robots”, 19 November 2012 hrw.org →: source for 2012
- 2019 The European Commission’s High-Level Expert Group on AI names human-in-the-loop, human-on-the-loop and human-in-command as ways to oversee any AI system, not only weapons.High-Level Expert Group on AI, “Ethics Guidelines for Trustworthy AI”, European Commission, 8 April 2019 digital-strategy.ec.europa.eu →: source for 2019
- 2021 Robert Monarch’s book Human-in-the-Loop Machine Learning covers annotation and active learning, noting that most deployed machine learning systems learn from human feedback.Robert (Munro) Monarch, Human-in-the-Loop Machine Learning: Active learning and annotation for human-centered AI, Manning, June 2021 manning.com →: source for 2021
- 2024 The EU AI Act enters into force on 1 August. It requires appropriate human oversight measures for high-risk AI systems.European Commission, “AI Act”, Shaping Europe’s digital future, page updated 3 August 2026 digital-strategy.ec.europa.eu →: source for 2024
- 2025 The Model Context Protocol specification, a standard for linking AI applications to tools, says a human in the loop should always be able to deny tool calls.Model Context Protocol specification, revision 2025-06-18, “Tools” modelcontextprotocol.io →: source for 2025
In the wild
-
“Human-in-the-Loop Weapons: Robots that can select targets and deliver force only with a human command”
Human Rights Watch and Harvard Law School’s International Human Rights Clinic, report hrw.org →: Human Rights Watch and Harvard Law School’s International Human Rights Clinic, report, 19 Nov 2012Archived →: copy of Human Rights Watch and Harvard Law School’s International Human Rights Clinic, report, 19 Nov 2012 -
“This “human-in-the-loop” can be beneficial in solving computationally hard problems, e.g., subspace clustering, protein folding, or k-anonymization of health data…”
Andreas Holzinger, Brain Informatics link.springer.com →: Andreas Holzinger, Brain Informatics, 2 Mar 2016Archived →: copy of Andreas Holzinger, Brain Informatics, 2 Mar 2016 -
“…proper oversight mechanisms need to be ensured, which can be achieved through human-in-the-loop, human-on-the-loop, and human-in-command approaches”
High-Level Expert Group on AI, European Commission digital-strategy.ec.europa.eu →: High-Level Expert Group on AI, European Commission, 8 Apr 2019Archived →: copy of High-Level Expert Group on AI, European Commission, 8 Apr 2019 -
“Human-in-the-loop aims to train an accurate prediction model with minimum cost by integrating human knowledge and experience.”
Wu et al., Future Generation Computer Systems arXiv 2108.00941 →: Wu et al., Future Generation Computer Systems, 2 Aug 2021Archived →: copy of Wu et al., Future Generation Computer Systems, 2 Aug 2021 -
“For trust & safety and security, there SHOULD always be a human in the loop with the ability to deny tool invocations.”
MCP specification, revision 2025-06-18 modelcontextprotocol.io →: MCP specification, revision 2025-06-18, 18 Jun 2025 -
“…HITL means that humans are involved at some point in the AI workflow to ensure accuracy, safety, accountability or ethical decision-making.”
Not to be confused with
-
Human on the loop
A person supervises and can override, but the system acts without waiting for approval. In the loop, each action needs the person first. Human Rights Watch drew this line for weapons in 2012.
-
Reinforcement learning from human feedback
RLHF uses people’s ratings of a model’s answers to train its general behavior. Human in the loop, in its oversight sense, puts a person in the path of particular outputs or actions while the system runs.
-
Human oversight
The broader duty, in law such as the EU AI Act, that people can oversee an AI system. A human in the loop is one way to meet it; EU experts also list on-the-loop and in-command oversight.
-
Human-in-the-loop simulation
In engineering, a simulation in which a real person operates part of the system, as in a flight simulator. It tests how people and equipment work together; it does not mean a person reviewing an AI system’s decisions.
Questions
What does human in the loop mean in AI?
It describes an AI system that includes a person at set points in its work. The person might approve an agent’s action before it runs, deny a tool call, or supply training data the model learns from. IBM gives the purpose as accuracy, safety, accountability and ethical decision-making, and abbreviates the phrase HITL. The EU AI Act requires human oversight for high-risk systems.
What is the difference between human in the loop and human on the loop?
In the loop, the system waits for a person before it acts. On the loop, the system acts by itself while a person watches and can override it. Human Rights Watch drew this line for autonomous weapons in 2012, with a third kind, out of the loop, that needs no person at all. It warned that thin supervision can leave an on-the-loop weapon effectively out of the loop.
Does the EU AI Act require a human in the loop?
It requires human oversight, a broader duty. The European Commission lists appropriate human oversight measures among the obligations for high-risk AI systems. The Act entered into force on 1 August 2024, and rules for high-risk systems in sensitive areas apply from 2 December 2027. In 2019 the Commission’s expert group named human-in-the-loop as one way to oversee AI, beside human-on-the-loop and human-in-command.
Why do AI agents ask for permission before acting?
An agent acts through tools such as a shell or a file editor, so its mistakes can change real files and systems. The Model Context Protocol says a person should always be able to deny a tool call, and that applications should show tool inputs first. Anthropic’s Claude Code asks before it changes files or runs commands. Anthropic also warns that too many prompts can cause approval fatigue, in which people stop checking what they approve.
Does a human in the loop make AI safe?
Not on its own. Ben Green reviewed 41 policies that require human oversight of government algorithms. Evidence suggests people cannot perform that oversight, he found, and such rules can give a false sense of security. Software shows a similar problem: Anthropic reports that constant approval prompts can lead people to stop paying attention to what they approve. Human Rights Watch raised the same concern about thinly supervised weapons.
Watch
Cite this entry
“Human in the loop.” SUPERR_dictionary, Superrduperr, 24 Sept. 2026, superrduperr.com/
Superrduperr. (2026, September 24). Human in the loop. In SUPERR_dictionary. https://superrduperr.com/
Superrduperr. “Human in the loop.” SUPERR_dictionary. September 24, 2026. https://superrduperr.com/