When you place a human "in the loop" of a high-velocity algorithmic process, you are not giving them control. You are giving them liability.
The sponge takes many shapes: the junior reviewer clicking approve on 400 cases a day; the resettlement officer countersigning outputs they have no capacity to audit; the content moderator whose override rate is a KPI. In every case, the architecture of the system manufactures a scapegoat whose job is to absorb blame when the system fails.
The diagnostic question: If this system harms someone, whose name ends up in the post-mortem? If the answer is a human sitting on the end of an algorithmic firehose, you are looking at a Liability Sponge, not a human in the loop.