Nudging people to stop rubber-stamping AI advice cuts blind trust — but why doesn't the team actually decide better?

This explores why designs that get people to stop rubber-stamping AI recommendations don't turn into better joint decisions. The clearest evidence comes from a 199-person study of cognitive forcing functions. These are deliberate bits of friction, such as making you commit to your own answer before the AI's appears, or making you ask for its recommendation instead of having it shown automatically. They cut overreliance more than explainable-AI displays did. Yet the human-AI teams still scored below the AI working alone, and participants liked these designs least Do forcing functions reduce AI overreliance without harming team performance?. The puzzle comes from a hidden assumption: that overreliance is the only thing standing between a team and good performance. Less blind trust also means more second-guessing, and second-guessing a model that is usually right costs accuracy.

The benefit also isn't spread evenly. The same experiment found that forcing functions helped people who enjoy effortful thinking (high 'Need for Cognition') much more than everyone else Do cognitive forcing functions help all decision makers equally?. So the average result blends a group that gains real judgment with a group that mostly gets slowed down and annoyed. An intervention can look fine on overreliance numbers while it quietly widens the gap between strong and weak decision-makers. That is one reason the team-level gains don't show up.

A deeper explanation is that friction treats a symptom and leaves the cause alone. One framework argues that people over-trust AI because of compounding cognitive traps: mistaking the model's fluent output for reality, confusing quick intuition with careful reasoning, and having existing beliefs echoed back Why do people trust AI outputs they shouldn't?. Forcing a pause interrupts the moment of acceptance. It doesn't give the person a better way to judge when the AI is wrong. So people end up either overriding the AI at random or deferring after a delay.

The most interesting counter-proposal changes what the AI hands over. 'Learning to Guide' has the machine stop offering a verdict and instead point out which parts of the input matter. The human still makes the call, but with better perception Can AI guidance reduce anchoring bias better than AI decisions?. If there's no AI answer to anchor on, there's no overreliance to fight, and the effort goes into improving human judgment instead of policing trust. This suggests the real problem isn't the amount of friction. It's that 'AI recommends, human approves' may simply be a weak way to split the work.

One caveat: the corpus's direct evidence here rests mostly on a single experimental setup. The broader claim that reducing overreliance rarely improves team performance is a plausible pattern, not a well-replicated finding.