Human in the Loop, or Perceived Oversight? The Psychological Inference That Drives AI Credibility

Abstract

Industry practice and AI governance frameworks increasingly treat human oversight of AI systems as a core trust-building mechanism. Yet the specific psychological inference through which human-in-the-loop (HITL) systems build credibility remains unidentified. We address this gap across two experiments in AI-generated investment research. In Study 1, we manipulate the disclosed source of an identical report (human, black-box AI, human-supervised AI, or glass-box AI) and find that source labels exert only small effects on credibility and no effect on investment behavior. Critically, human-supervised AI produces the lowest credibility of any condition, contradicting the assumption that adding a human is always trust-positive. In Study 2, we decompose HITL into five theoretically grounded candidate signals through which perceived source could shape credibility. Perceived oversight adequacy, the inference that “someone competent reviewed this work before I saw it,” emerges as the dominant mediator of the source-to-credibility relationship. The findings reframe the HITL question from whether to add a human to how to signal that the analytical process was sound.