Access to an AI answer made participants much less likely to say “I don’t know,” even when the system was usually wrong. Across five experiments involving 3,132 people, researchers tested obscure visual details from films that rarely appear in online text and are difficult for language models to answer reliably.
In two initial studies, people without AI declined to answer 36% and 44% of questions. When they could consult a model, those rates fell to 6% and 3%. In another test without financial incentives, reported confidence rose from 29.6 to 75.9 on a 100-point scale, while accuracy dropped from 27.6% to 10%. Automatically displaying AI advice produced a similarly large effect. Small rewards for correct answers helped, but did not restore the caution seen in the control groups.
The study’s design is intentionally adversarial, so it does not show that AI advice always harms judgment. Movie-detail questions are also far removed from many workplace decisions. Still, the result identifies a practical risk: fluent answers can encourage a decision where a person would otherwise pause. Interfaces may need to make uncertainty visible and preserve an easy way to abstain, not merely provide a confident-looking suggestion.