Not full automation and not full manual work — a designed seam where a person still gets to look before the system acts, or corrects it after.
Human-in-the-loop (HITL) is a specific term of art in interactive machine learning, not just a description of “people using AI.” It refers to systems deliberately architected so a human’s feedback, correction, or approval is part of the operating loop — training the model, gating its outputs, or both — rather than the system running fully autonomously end to end.
Saleema Amershi and colleagues’ widely-cited 2014 survey frames this as a tight coupling between system and user: end-users don’t just consume outputs, they shape the model’s behaviour through ongoing interaction, catching errors and steering the system in ways a fully automated pipeline would simply propagate.
Whether “human-in-the-loop” survives as a real design constraint or becomes a hollow phrase covering rubber-stamp approval is a live question — see Agency for why who is actually supervising whom is the more fundamental question underneath the technical pattern.
Source: Amershi, S., Cakmak, M., Knox, W.B., & Kulesza, T. (2014), “Power to the People: The Role of Humans in Interactive Machine Learning,” AI Magazine, 35(4), 105–120.