Authority bias

The canonical citation does not say what the design advice says it says. In Milgram (1963), 26 of 40 participants in a staged learning task went to the shock generator's last switch, labelled 450 volts. That is one condition of 23: Haslam, Loughnan and Perry (2014) pooled 21 of them, covering 740 participants, and found full obedience in 43.6% of cases rather than two thirds, with eight of 14 coded features of the situation independently predicting obedience: it fell with a non-directive or inconsistent experimenter, with group pressure to disobey, and when the learner was closer at hand. Perry, Brannigan, Wanner and Stam (2020) recovered an unpublished analysis in which, across 18 of 23 variations, participants who fully believed they were inflicting pain chose lower shock levels than those who doubted the setup. Gibson (2013) counted every use on the Yale tapes of the fourth and final prod, “You have no other choice, you must go on,” the only line that is an order: used on 23 participants, it produced further shocks from two, one of whom became fully obedient, and the most common reply was a flat rebuttal. There is no obedience rate here to borrow, and nothing in it about a badge on a screen.

Uniforms shift behaviour in small field experiments, without anyone being able to say why. Bickman (1974) had an experimenter dressed as a civilian, a milkman or a guard make a trivial street request. Compliance was higher for the guard, and unaffected by whether he stayed to watch, which points at legitimacy rather than fear of being caught; yet accompanying questionnaire studies found that students “did not perceive the guard as having either more power or more legitimacy than the milkman or civilian.” The other pillar of the popular account, nurses accepting a telephoned drug overdose order, has a published non-replication in Rank and Jacobson (1977). What attire reliably moves is ratings: in Petrilli and colleagues' (2018) patient survey, 53% said physician dress mattered to them and a white coat with formal attire rated highest for knowledge and trustworthiness, though scrubs were preferred for surgeons and patients 65 and over leaned hardest on the coat. That is a preference study about clothes, not evidence that a coat in a product photo sells a subscription.

The best-measured interface version is the trust seal, and the honest summary is small and conditional. Özpolat and Jank (2015) analysed 288,169 transactions across 493 retailers, from field experiments run by a trust seal provider that then shared the data — provenance worth weighing. Seals helped more for small retailers and for new shoppers, substituting for shopper experience and for the seller's own sales volume; they helped on higher-value carts, and only late in the shopping cycle; and showing more than two seals lowered the likelihood of completing a purchase. A badge is a crutch for a store with nothing else to show, not a lever on a store that already has order history and reviews.

The badge also has to be seen and understood, and mostly it is not. Kirlappos, Sasse and Harvey (2012) eye-tracked 60 experienced shoppers: 38% noticed none of the three seals present. Trust ratings rose where a seal was noticed — and when the researchers asked what the seals meant, every stated meaning was wrong, including the belief that a seal's presence “automatically legitimizes any website.” Edelman (2011) shows the cost: across roughly half a million sites, 5.4% of TRUSTe-certified sites were untrustworthy against 2.5% of uncertified ones, while BBBOnLine's certified sites ran the other way, 1.0% against 3.0%. The mark carries almost no information about what a user is trying to judge, and a badge that only asserts identity carries none: across studies of 748 and 2,041 participants, Vaidya, Votipka, Mazurek and Sherr (2019) found a verified-account indicator had no effect on willingness to share or act on a tweet from an unfamiliar account.

A credential can also lose to relatability outright. Alsan and Eichmeyer randomised men without a college degree to vaccine messages from different senders. The lay messengers, rated 0.54 standard deviations less qualified and trustworthy than the physicians, raised stated willingness to take the COVID-19 vaccine by 8.8 percentage points (20%), and 15 percentage points (39%) more reported a flu shot at follow-up — the only arm that differed significantly from the others, on a comparison with substantial attrition. Where trust in the institution is low, pick the messenger your audience resembles over the one with the title.

The large, well documented effect is the failure mode. Goddard, Roudsari and Wyatt (2012) reviewed 74 studies of automation bias — people accepting a system's advice when it is wrong — and Jacobs and colleagues (2021) priced it: 220 antidepressant-prescribing clinicians scored 0.357 on expert concordance unaided, were not helped by correct machine recommendations (0.384), and fell to 0.299 when the recommendation was wrong, below their own unaided accuracy, and to 0.262 when the error carried a feature-based explanation. Confidence did not move with accuracy, and the clinicians most familiar with machine learning scored worst. Buçinca, Malaya and Gajos (2021) show overreliance is reducible — cognitive forcing beat plain explanations — but the designs that reduced it most were the ones participants rated worst. The same cue serves attackers: Williams, Hinds and Joinson (2018) found authority cues in simulated spear-phishing sent to 62,000 employees raised clicks on a suspicious link.

So the tendency is real and the design claim built on it is weak. Pornpitakpan's (2004) review of five decades finds high-credibility sources beating low-credibility ones on average, with interaction studies showing credibility to be “a liability under certain conditions.” The numbers you can trust here describe trust being set wrongly: seals on unsafe sites, confident models degrading expert judgement, authority cues winning clicks for phishing. The defensible build is narrow — one specific, checkable, relevant credential, shown where the doubt is to a user with no other way to judge — and what to measure is whether people believed the right things, not whether they converted.