When a trusted source says something wrong
AI
of answers models already had right flipped after one note, styled as a "verified source," endorsed a wrong answer. Seven of eight models; general-knowledge questions; the note was always wrong. More authoritative-sounding notes got more compliance.
Humans
Clinicians given wrong decision-support advice were 26% more likely to decide incorrectly than controls (pooled RR 1.26, 95% CI 1.11–1.44). In prospective studies they abandoned their own correct judgment in 6–11% of cases.
The words: AI side, authority bias, sycophancy, compliance. Human side, automation bias, over-reliance, commission errors. Both clinical. No relabel in this pair, and saying so when it isn't there is what makes it count when it is.
Why not comparable: the denominators differ. The model figure is a share of correct answers flipped when the advice was always wrong; the clinician figures are a share of all cases (mostly with correct advice) or a relative risk against a control group. The subjects differ too: trained experts in their own field versus models answering trivia. The gap is wide enough that a matched study (same questions, same always-wrong source, same "had it right first" denominator) would plausibly come back humans do better. That's written down here so it can be tested and proven wrong.
Human: Goddard, Roudsari & Wyatt (2012), "Automation bias: a systematic review of frequency, effect mediators, and mitigators," JAMIA 19(1):121–127, doi:10.1136/amiajnl-2011-000089 (open access, PMC3240751; full text read).
Left out (closed access, no open copy, so not read): Skitka, Mosier & Burdick 1999; Mosier et al. 1998.