AI access makes people almost entirely unwilling to say "I don't know," study finds
decoder ·
A recent study involving 3,132 participants reveals that access to AI language models significantly reduces people's willingness to admit uncertainty. Researchers found that when AI advice was available, the rate of participants saying "I don't know" dropped from over 36% to just 3%, despite the AI providing mostly incorrect answers on specific trivia questions. This shift resulted in a sharp increase in user confidence but a corresponding decrease in answer accuracy, with correct response rates falling by more than half. The effect persisted even when AI suggestions were displayed automatically without user request, mirroring modern search engine behaviors. While financial incentives modestly reduced reliance on AI and improved accuracy, they did not fully restore the tendency to abstain from answering. The authors attribute this phenomenon to "Epistemia," where users accept convincing-sounding AI outputs without verification. Consequently, the research suggests that human judgment may suffer as AI becomes ubiquitous, regardless of model accuracy.
decoder