Does labeling advice as AI change how clinicians use it?

Gaube et al. tested whether telling physicians that diagnostic advice came from "an AI system" changes how they use it. Radiologists (n = 138) and internal or emergency medicine physicians (n = 127) reviewed eight chest X-ray cases. Each case came with advice that was either accurate or inaccurate and labeled as coming from an AI system or an experienced radiologist. All of the advice was written by human experts, so only the label varied. The label moved ratings in one group: "only participants with higher task expertise showed algorithmic aversion by rating the quality of advice to be significantly lower when it came from the AI in comparison to the human." It did not move diagnoses: "the purported source of the advice did not affect participants' performance." What did move accuracy was the advice itself. Task experts performed 40.10% better, and non-experts 37.53% better, when they received accurate rather than inaccurate advice.

The paper's account of the diagnostic result is that advice pulls judgment toward itself. The authors say the advice "could have engaged cognitive biases, by anchoring participants to a particular diagnosis, and triggering confirmatory hypothesis testing," and they describe "a general tendency for participants to agree with advice," stronger among physicians with less task expertise. They define clinical susceptibility as "the propensity to follow incorrect advice." On that measure, 41.73% of IM/EM physicians and 27.54% of radiologists always gave the wrong diagnosis when the advice was wrong. Susceptibility was not limited to weaker performers: 28.26% of radiologists and 17.32% of IM/EM physicians refuted all the incorrect advice they saw. The by-source breakdown of susceptibility sits in a supplementary figure that the excerpt does not include.

This sits against the claim that labels shape reliance. Does the label on advice shape how clinicians judge it? reports a label moving clinicians' preferences. This source agrees that a label moves evaluation, but it finds the movement stops at the rating: the same label that lowered radiologists' quality scores left their diagnoses unchanged. The split resembles the gap in Can self-ratings replace objective performance scores for AI competence?, where a judgment about the work and the work itself came apart. Here the judgment is advice quality and the work is the diagnosis. The accuracy result also suggests that the ceiling on assisted performance is set by the advice. That concern is raised from another direction by Why does assisted accuracy capture only half the LLM gain?.

The excerpt does not establish how the AI label was assigned or checked, and it omits the methods section and the supplementary by-source results. It covers eight cases drawn from one database, so it says nothing about other imaging tasks or about live clinical workflows. It also contrasts an AI label with a human-expert label on the same advice, which is a different question from how clinicians respond to an AI system that is actually wrong in its own ways. The implication, at the strength the evidence allows, is that a clinical AI tool should be evaluated on what clinicians do with its incorrect outputs, not only on how they rate it. On this evidence, a high quality rating from radiologists would not show that they would catch its errors.