Incorrect AI labels influenced how participants reporting themselves as physicians treated fictitious patients—and what they later believed about the treatment—even when outcome feedback could challenge those labels. Two online experiments published in July make a specific oversight problem visible: receiving contradictory information does not ensure that a person uses it to revise a judgment. They do not measure how often doctors make errors in real care. Original study

The research, by Aranzazu Vinas, Fernando Blanco and Helena Matute, separated a classification’s influence from the treatment’s programmed effect. That separation is the useful part of the design.

The label was wrong, but the feedback was available

In the first experiment, participants saw 60 fictitious cases. Half were labeled highly sensitive to a treatment and half were labeled less sensitive. Participants chose whether to administer it, then learned whether the patient recovered.

The labels did not describe a genuine difference. For both categories, the task programmed recovery in 70% of treated cases and 20% of untreated cases. The treatment therefore had the same programmed effect in either group: a 50-percentage-point difference between the recovery probabilities with and without it.

This comparison matters. Counting recoveries after treatment alone would not distinguish a treatment effect from improvement that would have occurred anyway. The experimental feedback gave participants information relevant to that distinction. Experiment 1 procedure

Choices and effectiveness judgments followed the categories

The first experiment analyzed 105 participants after attention-check exclusions. On average, participants chose treatment for 89.3% of cases labeled highly sensitive, compared with 55.7% of cases labeled less sensitive. The paper reports a statistically significant difference.

These percentages describe decisions across fictitious cases. They are not the proportions of doctors who harmed a patient, nor are they recovery rates.

Participants also judged the treatment more effective for the highly sensitive category: average ratings were 69.5 versus 51 on the study’s 0–100 scale. The researchers interpreted the pattern as a continuing influence of the classification, despite equal programmed effectiveness for the two categories. Experiment 1 results

An ineffective treatment tested a different learning problem

The second experiment reported 118 analyzed participants. This time, 70% of fictitious patients recovered whether they received the treatment or not. Its programmed effect was zero, for either label.

Nevertheless, treatment was chosen more often for cases labeled highly sensitive: an average of 78.3%, compared with 36.8% for the other category. Participants also gave higher effectiveness judgments for the highly sensitive group. Experiment 2 methods and results

The contrast explains why a successful-looking outcome can be misleading. When spontaneous recovery is common, seeing recovery after an intervention does not establish that the intervention caused it. Here, both the correct comparison and the incorrect classification were built into a controlled task.

The artificial setting limits the conclusion

Participants were recruited through Prolific, with professional filters and repeated questions about working as a doctor. The study did not independently verify their clinical credentials. It also omitted patient histories, comorbidities and the range of choices available in an actual consultation.

The authors explicitly consider another interpretation: facing an unfamiliar disease and treatment, participants may have deferred to what appeared to be an authoritative protocol. The experiment cannot completely distinguish that response from cognitive bias. It also did not compare specific training or interface safeguards to identify a remedy. Discussion and limitations

For professionals assessing clinical AI, this is evidence about learning under tightly defined uncertainty. It adds a behavioral question to the separate discussion of authority and oversight: what would show that a reviewer noticed and used information contradicting the system? Naming a reviewer and observing effective review are different kinds of evidence.