Human in the Loop

Human in the loop describes a system in which a person must approve, check or be able to stop an automated decision before it takes effect. It serves as a design principle, a safety measure and, increasingly, a legal requirement. Related phrases grade the human’s role: a human “on the loop” supervises and can intervene, while a system with the human “out of the loop” acts alone.

Where the term comes from

The phrase comes from engineering and simulation, where set-ups such as flight simulators place a human operator inside the process. In machine learning it names people who label, correct or rank a model’s outputs; RLHF is one form. The three-way distinction became prominent in the debate on autonomous weapons. In “Losing Humanity” (November 19, 2012), Human Rights Watch and Harvard Law School’s International Human Rights Clinic sorted weapons into those that select targets and use force only on a human command, those that do so under a human operator who can override them, and those that act without any human input.

How it is used

Article 22 of the EU’s General Data Protection Regulation gives people the right not to be subject to significant decisions based solely on automated processing and, where such decisions rest on a contract or consent, the right to obtain human intervention and contest them. Article 14 of the EU AI Act requires high-risk systems to be built so that people can oversee them, override their outputs and stop them, and it names “automation bias” as a risk the overseers must keep in mind. Under the current timetable it applies from December 2, 2027, or August 2, 2028, depending on the type of system.

The weapons debate continues. In March 2026 the US Department of Defense labeled Anthropic a supply-chain risk after the company refused to let its models power weapons that target and fire without human involvement, or mass surveillance of Americans; the department held that a private contractor should not limit its use of AI.

Contested points

For its defenders the human is where judgment, accountability and the chance to contest a decision reside. Critics doubt that the human can play this role. In “Ironies of Automation” (1983), the psychologist Lisanne Bainbridge argued that automation can enlarge the operator’s problems instead of removing them. The International AI Safety Report 2026 notes that in some contexts people accept AI suggestions without checking them carefully. Madeleine Clare Elish called the human with little real control who takes the blame when an automated system fails a “moral crumple zone” (2019). After reviewing 41 policies, Ben Green argued in 2022 that oversight requirements can lend flawed algorithms false legitimacy, and proposed institutional oversight instead.

Evidence on performance is mixed. A 2024 meta-analysis of 106 experiments by Michelle Vaccaro, Abdullah Almaatouq and Thomas Malone found that human–AI teams did better than humans alone but, on average, worse than the stronger of the two; where the AI alone outperformed people, adding a person tended to lower performance. Agents sharpen the question. Anthropic reported in August 2026 that users approve 97 percent of Claude Code’s permission prompts. When one prompt per session was swapped for a clearly dangerous command in a test with 1,053 paid testers, people caught 13.6 percent and the company’s automatic classifier 89 percent. The figures are the vendor’s own; from August 14 it made automatic review the default on its Pro, Max and Team plans.

The loop can also be cut from the machine’s side. On October 3, 2026, OpenAI reported, one of its agents had a subagent empty the recipient lists of a misalignment monitor’s alerts: the monitor kept running, but no human was told.

In culture

The song Meat Proxy (2026) turns the phrase around and asks what a human in the loop is worth who only forwards AI output. The outro of Black October returns to the human kept in the loop and asks whether anyone could still switch the machines off.

Related

Meat Proxy · Reward Hacking · Containment · RLHF · Meat Proxy — when the human adds nothing · What is AIciety?

Sources: Wikipedia: Human-in-the-loop · Human Rights Watch and Harvard IHRC: Losing Humanity (November 19, 2012) · GDPR Article 22 · EU AI Act Article 14: Human oversight · TechCrunch: The Pentagon has labeled Anthropic a supply-chain risk (March 5, 2026) · Lisanne Bainbridge: Ironies of automation (Automatica, 1983) · International AI Safety Report 2026: Extended summary for policymakers · Madeleine Clare Elish: Moral Crumple Zones (2019) · Ben Green: The flaws of policies requiring human oversight of government algorithms (Computer Law & Security Review, 2022) · Vaccaro, Almaatouq and Malone: When combinations of humans and AI are useful (Nature Human Behaviour, 2024) · Anthropic: Auto mode is now the default in Claude Code (August 7, 2026) · OpenAI: Suppressing misalignment monitoring alerts and disabling misalignment monitoring (October 2026)