Can making an AI's advice harder to see stop people from blindly trusting it, without making the human-AI team worse at the job?

This explores whether holding back or delaying AI advice can stop people from rubber-stamping it without making the human-AI team worse. The corpus's short answer is that hiding advice partly works on overreliance but doesn't fix team performance. The clearest evidence comes from a 199-person study of "cognitive forcing functions," designs that add friction such as making people commit to their own answer before seeing the AI's. These designs cut overreliance more than explanations did. But the human-AI teams still scored worse than the AI working alone, and participants liked these designs least of all the conditions tested Do forcing functions reduce AI overreliance without harming team performance?. So friction weakens anchoring on the AI's answer, but it doesn't turn people into good judges of when the AI is right.

The obvious alternative is to show the AI's advice and explain its reasoning, and that turns out to be worse. A review of 35 studies found that explanations often deepen misplaced trust instead of reducing it, especially among people with little AI literacy. A confident-sounding rationale reads as a reason to agree. What actually helped was active engagement and checking the answer yourself, not more transparency Does explaining AI decisions actually reduce user overreliance on it?. Read together, the two findings suggest a problem with the framing: whether advice is shown or hidden matters less than whether the person is doing real thinking when they meet it.

That points to a third option the question doesn't mention: change what the AI gives you instead of hiding it. "Learning to Guide" has the machine point out which parts of a case deserve attention instead of handing over a verdict. That removes the anchor entirely while still improving the person's judgment, and responsibility stays with the human Can AI guidance reduce anchoring bias better than AI decisions?. A related finding concerns time. When people are told their partner is an AI, they first avoid it, but that bias reverses once they repeatedly see how its advice turns out. Without that outcome feedback, simply knowing it's an AI doesn't help anyone calibrate their trust Does revealing AI identity help or hurt user trust?. That suggests good reliance is learned from seeing results, not produced by one well-designed screen.

One more thing is worth knowing. Friction is fighting the way assistants are trained. Optimizing for user satisfaction makes agreement central to how assistants behave Is sycophancy in AI systems a training flaw or intentional design?. Users also rate friction-heavy designs poorly. So the pull is toward smooth, agreeable advice that invites overreliance, and the designs that resist it are the ones people want to switch off. The corpus doesn't have a study showing that hiding advice keeps or improves team performance. The more promising directions are reshaping advice into guidance and building in feedback loops.