Confidence Cues in Advice Taking: Comparing AI and Peer Advice
Type
conference paper
Date Issued
2026-06-16
Author(s)
El Fassi Ismail
;
Srinivasan Suraj
Abstract
When people receive advice from AI systems, they often over- or under-weight it relative to equivalent human advice, and both forms of mis-reliance degrade decision quality. Because the accuracy of any single recommendation is rarely observable when it is used, decision-makers rely on cues disclosed alongside the advice to judge how far to trust it. We study one such cue, the level of confidence an advisor states in its recommendation, in two pre-registered experiments (estimation, N = 308, 7,392 decisions; forecasting, N = 398, 1,990 decisions). Holding the recommendation fixed and varying stated confidence, we find that higher advisor confidence sharply increases reliance on algorithm-labeled advice but barely moves reliance on peer advice; the same signal is source-contingent. Reliance on algorithmic advice approaches the equal-weighting benchmark that peer advice never reaches, and confidence raises users' self-confidence more for algorithmic advice, identifying disclosed confidence as decision-facilitating information whose value hinges on its calibration to accuracy.
HSG Classification
contribution to scientific community
Refereed
Yes
Division(s)