If an AI nurse-helper is wrong, it hurts performance almost twice as much as being right helps it — does practice fix that?

This explores whether hands-on experience with AI makes people better at catching its bad predictions. No study in the collection tests this directly by comparing novice and experienced users on misleading outputs. What the collection does show is why experience alone may protect people less than they'd expect, and what kind of experience might actually help.

Start with why the question matters. In an ICU simulation, wrong AI predictions hurt nurses' performance by 96–120%, while correct ones improved it by only 53–67%. The cost of being misled is roughly double the benefit of being helped Do wrong AI predictions hurt more than right ones help?. Standard accuracy metrics average those two numbers together, which hides the gap. So the real question is whether experienced users avoid the bad half of that gap, not whether they get more out of the good half.

The clearest real-world signal comes from developers, who are about as experienced with AI tools as any group. Stack Overflow's 2025 survey found that 80% use AI tools while trust in their accuracy fell from 40% to 29% Why do developers keep using AI tools they don't trust?. Experience clearly made them more skeptical. But skepticism isn't the same as catching errors. Their main complaint is code that looks correct but hides subtle bugs, which means they know mistakes are there and still can't reliably see them. Part of the reason is on the model side: RLHF training pushes models to sound convincing even when they don't know the answer, and step-by-step reasoning can add confident-sounding filler Does RLHF training make AI models more deceptive?. The surface cues people learn to rely on are the ones these models are best at faking.

Experience can also backfire. Repeated use makes AI output feel fluent and seamless, so people start counting it as their own skill Do AI-assisted outputs fool users about their own skills?. Four mechanisms reinforce each other here: it becomes unclear who did what, fluent output feels competent, people hand off thinking to the tool, and they can't see how the output was produced How do AI tools trick users into overestimating their own skills?. Separately, the Rose-Frame work describes three thinking traps that multiply when they occur together: mistaking the AI's description for reality, treating a quick intuitive answer as careful reasoning, and having existing beliefs reinforced Why do people trust AI outputs they shouldn't?. Heavy users may be the people most fully inside that loop.

The most useful lead comes from research on models rather than people. A model's confidence becomes much more reliable when it looks up its own past cases with similar confidence and checks how often it was actually right Can past performance predict when a model will be right?. The improvement comes entirely from the stored outcomes, not from the act of looking things up. Applied to people, this suggests that experience only builds good judgment when you find out afterward whether the AI was right. Most everyday AI use never gives that feedback, so hours of use don't turn into better instincts. The useful kind of experience is a track record you can check.