AI agents

Human in the loop

In short

Human in the loop (HITL) means a person can intervene in every decision cycle of an AI system. The 2019 Ethics Guidelines for Trustworthy AI call this one governance mechanism, and say it is in many cases neither possible nor desirable1.

Also called
HITL, Human-in-the-loop
Topic
AI agents
Level
Beginner

The High-Level Expert Group on AI, an independent expert group set up by the European Commission, describes three governance mechanisms for human oversight. Human-in-the-loop means a person can intervene in every decision cycle. Human-on-the-loop means a person can intervene during the design cycle and monitor the system while it runs. Human-in-command means a person oversees the system’s overall activity, including its economic, societal, legal and ethical impact, and can decide when and how to use the system, which can include overriding a decision1. The guidelines say intervention in every cycle is in many cases neither possible nor desirable. They also say that, other things being equal, the less oversight a person can exercise over an AI system, the more extensive testing and stricter governance is required1.

NIST’s AI Risk Management Framework does not use the name HITL. It says human roles and responsibilities in decision-making and oversight need to be clearly defined and differentiated, and that human-AI configurations can span from fully autonomous to fully manual2. A system might decide on its own, defer a decision to a human expert, or be used by a human decision-maker as an additional opinion2. Some systems may not require human oversight. NIST’s example is a model used to improve video compression. Other systems may specifically require it2.

NIST says human-AI interaction results vary. In some perceptual-based judgement tasks, the AI part can amplify human biases, so the combined decision is more biased than the person or the AI alone2. It says further study is still needed on how far people are empowered and incentivised to challenge an AI output, and that data on how often people overrule an output in deployed systems, and why, may be useful to collect and analyse2. Anthropic describes agents pausing for human feedback at checkpoints or when they meet a blocker3. The OpenAI Agents SDK pauses a run until a person approves or rejects a sensitive tool call, then resumes from the saved state4.

Example

The OpenAI Agents SDK’s documentation marks a cancel-order tool as always needing approval. When the model calls it, the run pauses until a person approves or rejects that call, and only then continues4.

Why it matters

HITL means a person can intervene in every decision cycle, and the 2019 guidelines say that is in many cases neither possible nor desirable1. NIST says human roles in oversight still need to be defined, and that some systems may not require human oversight2.

Questions people ask

What does HITL mean in AI?

HITL means human in the loop. In the 2019 ethics guidelines for trustworthy AI, it is the capability for a person to intervene in every decision cycle of an AI system. Those guidelines say that is, in many cases, neither possible nor desirable1.

Is human in the loop the same as human oversight?

No. The guidelines say human oversight helps ensure an AI system does not undermine human autonomy or cause other adverse effects. HITL is one governance mechanism for that oversight, alongside human-on-the-loop and human-in-command1.

Can a human in the loop prevent every AI mistake?

The guidelines do not say that. They say intervention in every decision cycle is in many cases neither possible nor desirable1. NIST says human-AI interaction results vary, and that in some perceptual-based judgement tasks the AI part can amplify human bias, leading to a more biased decision than the person or the AI alone2.

How do AI agents pause for a person?

Anthropic describes agents pausing for human feedback at checkpoints or when they meet a blocker3. In the OpenAI Agents SDK, a tool can require approval, and the run pauses until a person approves or rejects that call4.

Related terms

Tools that use it

Sources

  1. High-Level Expert Group on AI (2019), Ethics Guidelines for Trustworthy AI
  2. NIST (2023), Artificial Intelligence Risk Management Framework 1.0, Appendix C
  3. Anthropic (2024), Building effective agents
  4. OpenAI Agents SDK, Human-in-the-loop

Back to the glossary