A little wrong advice from an AI can wreck a nurse's judgment far more than good advice ever helps it.

This explores why a nurse's judgment about whether a patient is in trouble shifts when an AI tool gives wrong advice, and why the damage is so much bigger than the benefit when the tool is right. The collection's most direct evidence is an ICU simulation. In it, misleading AI predictions cut nurse performance by 96–120%, while correct predictions improved it by only 53–67% Do wrong AI predictions hurt more than right ones help?. The more useful finding is about how this was hidden. Standard accuracy metrics average the gains and losses together, so a tool can look like a modest net positive while wrong calls on individual patients are doing serious harm. A note of caution: the corpus doesn't break down which kinds of emergencies nurses get wrong, or in which direction (missing a real crisis versus raising a false alarm). The 'why' has to be pieced together from nearby work.

Mammography is the clearest parallel. When radiologists were shown wrong breast-cancer risk categories presented as AI output, experienced readers dropped from 82% to 45.5% accuracy, and inexperienced readers fell below 20% How much does wrong AI advice harm radiologist accuracy?. That suggests an answer to the 'differently' in the question: wrong advice doesn't hit every clinician equally. People with less settled judgment of their own have less to push back with, so they follow the AI furthest. That pattern would likely carry over to nurses with different levels of acute-care experience, though the nurse study itself doesn't test it.

One result looks like a contradiction. Other radiology work finds that clinicians actually underweight AI predictions. They treat their own read and the AI's as if they were independent, so on average the AI adds nothing Why don't radiologists benefit from AI predictions?. How can clinicians both discount AI and get dragged down by it? A third study helps explain this. Radiologists rated advice lower when it was labeled 'AI', but their accuracy still followed whether the advice was right or wrong, not who it came from Does labeling advice as AI change how clinicians use it?. Put together, these suggest that stated skepticism is not real protection. Clinicians can distrust the AI in principle and still have their actual decisions pulled by it.

Why might wrong advice be so 'sticky'? The Rose-Frame work names three cognitive traps that make each other worse: treating the model's output as if it were the situation itself, mistaking a quick fluent answer for careful reasoning, and having existing hunches confirmed Why do people trust AI outputs they shouldn't?. In triage, an AI that labels a deteriorating patient 'stable' feeds all three traps at once. A related argument says expert observation is about choosing which differences matter, such as a subtle change in color or breathing, and that statistical pattern-matching can't do that choosing Can AI distinguish which differences actually matter?. Misleading AI may cause harm less by adding false information than by pulling the nurse's attention away from the cue that would have caught the problem.

The surprising part for anyone designing these systems is that putting a human in the loop doesn't automatically catch errors. In a study of consultants, people who fact-checked and pushed back on GPT-4 often got stronger persuasion back instead of a correction Does validating AI output make models more defensive?. That study wasn't about clinical prediction tools. But taken with the nurse and radiologist findings, it suggests a better question to ask of a clinical AI: not 'how accurate is it on average?' but 'what happens to the human when it's wrong?'