AI Access Cut 'I Don't Know' Answers From 44% to 3% in 3,000-Person Study
People given an AI assistant stopped admitting uncertainty even when the model was wrong almost every time, researchers found.
Handing people an AI assistant appears to wipe out their willingness to admit ignorance. In an experiment with more than 3,000 participants, the share of answers of "I don't know" collapsed from 44 percent to 3 percent once an AI tool was available, according to a study reported by The Decoder. The drop held even in a trial where the AI's answers were wrong almost every time.
The setup matters more than the headline number. When participants could consult the model, they answered more questions and reported feeling more sure of themselves. Their actual accuracy went the other way: they were correct roughly one-third as often as people working without AI. Confidence rose while correctness fell, a split that the researchers treat as the central finding.
That pattern points to a specific failure mode, not general laziness. Admitting you don't know is normally a cheap, honest signal that flags where your knowledge stops. Remove it and the people evaluating the answers lose their clearest cue for spotting bad information. The AI's errors were not the only problem; the disappearance of doubt did much of the damage.
The scale of the study gives the result weight. With over 3,000 participants across the experiments, the effect is not a small-sample artifact, and the 44-to-3 shift is large enough to rule out noise. The design also isolates access to AI as the variable, since the comparison groups faced the same questions under the same conditions.
The findings land as AI assistants are folded into search, office software and classrooms, often framed as tools that help users learn. This study suggests the opposite can happen: the tool can substitute for the reasoning it was meant to support, leaving users more certain and less right. That is a hard trade to spot in the moment, because a confident answer feels like a good one.
The open question is whether the effect fades with training or better interfaces. Prompts that force a confidence rating, models that state their own uncertainty, or workflows that reward flagging gaps could all change the outcome. Until that is tested, the study stands as a warning that access to answers is not the same as knowledge.