Human-in-the-Loop
A system design pattern where humans participate directly in AI decision-making processes, typically by reviewing, approving, or correcting AI outputs before they take effect.
Variants
Human-in-the-Loop: Human approval required for each decision. High control, low throughput.
Human-on-the-Loop: Human monitors system operation and can intervene, but doesn’t approve each action. Balanced control and efficiency.
Human-out-of-the-Loop: Fully autonomous operation. Efficient but highest risk.
Why It Matters
HITL is the default assumption for “safe” AI deployment: if something goes wrong, a human catches it. But this assumes:
- Humans can detect AI errors (often false, automation bias)
- Humans have capacity to review volume (often overwhelmed)
- Review is meaningful, not rubber-stamping (often performative)
Design Considerations
What triggers human review? All outputs, or only uncertain ones? What information does the human see? Just the output, or confidence, alternatives, reasoning? What can the human do? Approve/reject, edit, or escalate? How is disagreement handled? Does human override trump AI?
The Alert Fatigue Problem
Too many low-value reviews train humans to approve without thinking. Effective HITL requires calibration: humans should see cases where their judgment genuinely adds value.
Related: 05-molecule—dynamic-trust-calibration, 07-molecule—ui-as-ultimate-guardrail, 01-molecule—appropriate-reliance-framework