When AI Answers, Nobody Says "I Don't Know"

The AI industry promises tools that make us smarter, faster, more capable. This week, a study of 3,132 people delivered a different verdict: the more AI access people have, the less willing they are to admit they don't know something. Even when the AI is almost always wrong.

The researchers call it Epistemia. The finding is visceral, not academic.

The 44-to-3 Percent Collapse

Researchers at a major university ran five experiments with more than 3,000 participants to test one question: does access to an AI assistant change how people handle uncertainty?

They designed the questions carefully. Participants were asked about fine visual details from movies, such as the color of a team uniform in Bend It Like Beckham, specific scene details that rarely appear in online text. These are exactly the kind of questions an AI model almost always gets wrong, because the training data simply does not contain that information.

The model they used was a current-generation system. It did get most ordinary questions right. But on these deliberately obscure visual facts, it was systematically wrong.

The results were stark. In the control group without AI access, participants withheld judgment on 44% of questions: they responsibly acknowledged they did not know. In the group with AI access, that number collapsed to 3%.

But here is the more troubling finding: participants with AI access rated their confidence at 75.9 points out of 100, compared to 29.6 without AI. That is roughly two and a half times more confident. At the same time, the share of correct answers fell from 27.6% to 10%. People were far more confident in answers that were far less likely to be right.

Across all five experiments, the pattern held. Participants with AI access got 9.2% of questions right versus 27.5% without AI. They answered more questions but were correct only about a third as often.

The Problem That Cannot Be Fixed

One might argue this is a design issue. If the AI were more accurate, the effect would disappear or even reverse. But Study 4 tested this by showing AI answers automatically, without participants having to ask. The effect barely changed. Judgment suspension dropped from 35% without AI to 1% with automatically displayed AI answers.

This suggests the effect is not about choosing to use a bad AI. It is about the availability of AI answers fundamentally rewiring how people handle uncertainty.

The researchers ran a fifth study with financial incentives. Even when participants were paid to be accurate, the effect weakened but did not disappear. People who had just seen an AI answer were measurably less willing to say "I don't know" moments later, even when the AI was no longer available and the new question was unrelated.

What Epistemia Means

The researchers frame their results around Epistemia: the tendency to accept AI answers because they sound convincing rather than because they are actually correct. The term draws on the Greek "episteme" (knowledge) and the suffix "-emia" (a condition or state). It describes a condition of knowledge acceptance without verification.

This is distinct from simple laziness or automation bias. Participants in the study were not passive. They engaged actively with the AI output. But the mere presence of an authoritative-sounding answer suppressed the metacognitive signal that normally triggers "I don't know."

The study builds on a growing body of research showing that AI access correlates with reduced critical thinking. Previous work found that AI use for just 10-15 minutes measurably reduces persistence on problem-solving tasks. Users become less willing to engage in open-ended exploration when an AI can provide immediate answers.

The Human Cost of Never Saying "I Don't Know"

"I don't know" is not a failure state. It is a cornerstone of intellectual humility, scientific thinking, and responsible decision-making. Doctors who cannot say "I don't know" misdiagnose. Engineers who cannot say "I don't know" build on shaky assumptions. Policymakers who cannot say "I don't know" regulate blind.

If AI systems systematically suppress that response, not by improving knowledge but by creating an illusion of certainty. The long-term implications extend far beyond a laboratory experiment. The researchers put it bluntly: "As AI answers become ubiquitous and increasingly appear unsolicited, the willingness to say 'I don't know' could be among the first casualties of human-AI interaction."

The finding does not mean AI is bad or that we should stop using it. It means the industry has focused almost entirely on accuracy metrics and benchmark scores while neglecting a subtler, harder-to-measure effect: what AI does to the person using it. An AI that gets every answer right but trains its users to stop questioning is not an unqualified success. It is a new kind of risk, one that standard evaluation frameworks do not capture.

The study is a reminder that the most important question about AI may not be whether its answers are correct. It may be whether we are still willing to ask the question at all.

Sources

• The Decoder: "AI access makes people almost entirely unwilling to say 'I don't know,' study finds". https://the-decoder.com/ai-access-makes-people-almost-entirely-unwilling-to-say-i-dont-know-study-finds/