Why do experts miss AI’s errors? Evidence from a randomized labeling experiment

As organizations increasingly rely on algorithmic decision aids, human oversight is vital to prevent automated errors from spreading. But what makes experts correct an algorithm or let it stand? In a preregistered randomized experiment, education experts reviewed identical student work paired with an intentionally inaccurate score labeled as either human- or AI-generated. We independently varied whether the score was too harsh or too lenient. The outcome-the grading fairness gap-measures the distance between the expert's revised mark and the objective rating. Under a harsh recommendation, the gap was 22% larger when the score was labeled as AI-generated; in the lenient case, the fairness gap under AI and human labels was statistically indistinguishable. Mediation analysis reveals that higher perceived ability and responsibility of the algorithm in the harsh scenario explain over half of the effect, while weaker attributions in the lenient case lead to stricter corrections. Thus, deference to AI depends not on automation itself but on the direction of its errors and the credibility it signals-offering design insights for accountable human-AI collaboration.