AI access makes people almost entirely unwilling to say "I don't know," study finds
Study of 3,132 people finds AI access nearly eliminates "I don't know" answers while sharply reducing accuracy.
Researchers ran five experiments with 3,132 participants and found that access to Step 3.5 Flash, which was almost always wrong on fine movie-detail questions, cut judgment suspension from 36–44% to 3–6%. Without incentives, confidence rose from 29.6 to 75.9 while correct answers fell from 27.6% to 10.0%; across studies, AI-access participants were correct 9.2% of the time versus 27.5% without AI. Financial incentives reduced advice-seeking but did not restore abstention to no-AI levels, and automatically displayed answers produced a similar collapse, from 35% to 1% without incentives. The authors call the pattern Epistemia and cite related work linking AI use to weaker critical thinking and metacognition.