Human judgement
Would You Trust the Same Advice With a Different Name on It?
A study found that an AI label changed how doctors rated advice, without a clear change in accuracy. Test the evidence your team checks, whoever appears to give the recommendation.

Before you rely on an “AI recommended” label, check what the reviewer still needs to know. Suppose a dispatch clerk must approve a route that crosses a road with a weight limit. The route recommendation can carry a supervisor’s name or a software label. Neither label tells the clerk whether the loaded van fits the limit.
Human experts wrote every recommendation
In a 2021 study in npj Digital Medicine, Susanne Gaube and colleagues asked 138 radiologists and 127 internal or emergency medicine physicians to diagnose eight cases from chest X-rays and short case notes. Each doctor saw six cases with correct advice and two with incorrect advice. The researchers told some doctors the advice came from an AI system and others that it came from an experienced radiologist. Human experts prepared all of it.
Radiologists and the other physicians were less accurate on cases with wrong advice. The researchers found no clear difference in diagnostic accuracy between the two labels. Radiologists rated the advice lower in quality when it carried the AI label. Every participant received advice. The experiment measured answers to selected cases and did not measure outcomes for patients in care.
Give different reviewers the same route and advice
Use a mock dispatch screen to show the same map, vehicle weight and recommendation to two groups of staff, assigning each reviewer to a group at random. Keep the position and wording of the recommendation fixed and change only the source label. Tell staff the exercise is simulated. Give each reviewer a mix of valid routes and routes whose proposed vehicles exceed a limit. Each reviewer should see each route once.
For each label, count the routes staff approved even though the vehicle exceeded the road limit. Keep the total number of over-limit routes beside that count, along with the evidence reviewers cited. Check whether they still approve valid routes. If the restriction goes unnoticed under either label, put the loaded weight beside the road limit on the approval screen and retest on fresh examples.
Sources and further reading
Sources checked 9 November 2026. Edited and reviewed by AI Boss.