Does time pressure make AI advice more persuasive to experts?

In a web-based experiment with 28 trained pathology experts estimating tumor cell percentage on H&E-stained slides, AI integration gave a statistically significant rise in overall performance and also a 7% automation bias rate, in which "initially correct evaluations were overturned by erroneous AI advice." That 7% is the authors' own measurement, which they set against the 6% to 11% acceptance rates they cite from Goddard et al. Time pressure did not change how often automation bias occurred, but it appeared to raise its severity: heavier reliance on negative system consultations and a sharper performance decline.

The authors define automation bias as treating "automated cues as infallible," and count it only as commission errors: the independent estimate matched ground truth, and both the AI advice and the AI-aided estimate were wrong. Their proposed mechanism for time pressure is strain on cognitive resources, which increases alignment with AI advice, "beneficial when the AI is accurate, but detrimental when the system errs." The complicating result is that pathologists were "largely unwilling to adopt model recommendations" that contradicted their own judgments, so the authors suggest automation bias "may not be the primary cognitive bias" in this setting. The first hypothesis is fully accepted; the second is accepted only for severity.

This sits against Why do patients distrust medical AI systems?, which locates resistance on the user side, in beliefs about what AI can do. Here the expert resistance was real, but the costly failure ran the other way, through agreement with wrong advice. The rating study in Can clinicians tell GPT-4 advice apart from expert advice? judged AI output by perceived quality; this paper measures what clinicians do with it, and together they suggest that judging AI output on its face is a weak safeguard. The cost also appears faster than in Does AI assistance weaken our brain's ability to think independently?, which tracks a cost building over months; here it shows within one session.

The excerpt does not establish a clinical error rate. The sample is "modest" (28 experts, with few automation bias incidents), so the non-significant frequency effect is weak evidence that time pressure does not matter. Time pressure was simulated as a condition rather than a countdown, and the omitted clinical background may have made participants less diligent than in routine work. The task was one visual estimate on 20 patches from public datasets, with AI advice from an FCOS-based detector. The supportable claim is narrower than a prevalence figure: under these conditions AI advice can induce a measurable error, and time strain makes that error more damaging. Whether time pressure raises the frequency of automation bias in clinical practice would need field data.