Does the label on advice shape how clinicians judge it?

Clinicians' judgments followed the label on the advice more than the advice's actual author. When a reply was perceived as expert-written, it scored higher on scientific quality and on all three empathy components, and it was preferred. The excerpt reports coefficients of −1.89 for scientific quality, −3.62 for emotional empathy, −3.02 for cognitive empathy and −2.74 for motivational empathy, all at p < .001. On preference, actual authorship (β = 6.96, p = .002) and perceived authorship (β = 6.26, p = .001) were both significant. Participants showed a "93.55 % preference for perceived expert advice," and they chose it "regardless of whether the answer was actually authored by AI or an expert."

The excerpt's own term is influence: "perceived authorship influenced ratings." Its highlights name "potential biases in the acceptance of AI-generated mental health support." The mechanism it offers is belief about authorship, not the text itself. The identification result sharpens this. Clinicians were at chance (45 % accuracy), so the beliefs that moved their scores were not reliable readings of who had written the text. The excerpt also reports a significant interaction between perceived and actual authorship (β = −12.29, p = .001), shown in its Fig. 1. The text does not unpack that interaction beyond the 93.55 % figure. The sign convention for the rating coefficients is not explained, so this note reports direction only. Perceived authorship was a rater's reported guess, not an assigned condition, so the analysis describes an association between belief and score.

This is the bias the paper names, and it qualifies the parity result in Can clinicians tell GPT-4 advice apart from expert advice?. That parity was measured with authorship hidden from the rater's judgment. The excerpt does not test what happens when the label is visible, so parity cannot be assumed to survive disclosure. The pattern also echoes Can language models truly understand therapeutic ruptures?, where a match to a reference label can hide a different method underneath. Here the rater's score tracks a label (who wrote the text) rather than the text. Both cautions point at the same evaluation risk: a score can faithfully reflect a label while telling you little about the thing the label names.

The excerpt does not show whether disclosure would change the result, because it does not state the order in which the authorship guess and the ratings were collected. It does not show that patients or other lay readers would show the same pattern. The raters were 43 licensed clinicians, 40 of them psychologists. The practical question is open. Before disclosure rules are set for AI-written mental health advice, ratings would need to be compared with authorship shown and hidden. The excerpt does not report that comparison. The bias is best read as a documented tendency in one blinded study, not as a measured effect size for practice.