Study finds AI access nearly eliminates willingness to say 'I don't know'

By Splitfeed EditorialLast updated Sep 26, 2026, 18:54 UTC

A study published in late September 2026, drawing on more than 3,000 participants across multiple experiments, measured what happened to people's epistemic caution when they were given access to AI-generated answers. In the sharpest of the experiments, the share of participants willing to say they did not know an answer fell from 44 percent to 3 percent once AI access was introduced. The troubling detail embedded in that finding is that the AI used in the experiment was described as almost always wrong, and participants who relied on it were correct only about one third as often as those who answered without it.

The significance lies not simply in AI producing errors, which is a familiar complaint, but in what the study suggests about how people calibrate their own uncertainty in the presence of a confident-sounding system. Saying "I don't know" is a form of intellectual honesty that protects against acting on bad information. When that behavior collapses from 44 percent to 3 percent regardless of the AI's actual accuracy, the gap between felt confidence and real accuracy becomes a practical risk in any domain where decisions follow from answers, including medicine, law, or financial advice. The study does not settle whether this effect is specific to certain kinds of questions, certain populations, or certain AI interfaces, and those remain open questions.

None of the three camps had published reactions to this story at the time of writing, so what follows reflects the positions each camp would typically bring to a finding like this. The Pro-AI camp would likely argue that the study uses an artificially degraded AI to produce a worst-case scenario, and that well-designed systems with accurate outputs and clear uncertainty flags would not produce the same effect. The Anti-AI camp would treat the result as evidence that AI creates a kind of epistemic dependency that erodes the habits of mind people need to think critically, independent of whether any given AI is accurate. The Middle Ground camp would probably call for better interface design and user education, arguing that the problem is not AI access itself but the absence of guardrails that prompt users to weigh AI output against their own judgment.

The most useful follow-up would be studies testing whether accuracy-calibrated AI systems, ones that express their own uncertainty openly, reduce or eliminate the suppression of "I don't know" responses. Research designs that vary AI accuracy levels and interface styles rather than holding the AI fixed at near-total inaccuracy would help clarify how much of this effect is a feature of AI tools broadly versus a feature of poorly designed or unreliable ones specifically.