Researchers conducted five experiments involving 3,132 participants to analyze how access to language models influences decision-making and the willingness to admit uncertainty. The study utilized models including Step 3.5 Flash, GPT-5.5, Claude 4.6 Sonnet, and Gemini 3.5 Flash, focusing on questions where models frequently hallucinate, such as specific visual details from films.
In initial tests, participants without AI access withheld judgment on 36 to 44 percent of questions. With AI access, this figure dropped to 3 to 6 percent. Despite a significant increase in confidence—rising from 29.6 to 75.9 points on a 100-point scale—the accuracy of responses fell from 27.6 percent to 10.0 percent when AI was used.
Financial incentives, where participants earned or lost 10 cents per answer, did not eliminate the reliance on AI. While incentives slightly increased the willingness to abstain from answering, the effect remained significantly lower compared to the control group without AI. Even when AI suggestions were provided automatically, judgment suspension remained near zero.
The authors suggest that users may adopt the AI's tendency to provide an answer regardless of its accuracy, a phenomenon they link to 'Epistemia.' This behavior contrasts with traditional advice-taking, where individuals typically underweight external input. The study indicates that AI access can lead to more frequent, yet less accurate, responses.
Broader research supports these findings, noting a negative correlation between AI usage and critical thinking. Studies from the Swiss Business School and Microsoft indicate that delegating cognitive tasks to AI can reduce problem-solving persistence and place high demands on metacognition, with higher education levels serving as a potential protective factor.
