The Vanishing “I Don’t Know”

A few minutes with AI can make an unfamiliar subject seem pretty well understood. A clear explanation and a few follow-up questions are often enough to feel ready to discuss it with confidence. It’s hard, though, to know how much of that confidence comes from understanding the subject and how much comes from an answer that feels convincing.

A recent study looked at what that does to people’s willingness to admit they don’t know the answer to a given question. Across five experiments with more than 3,000 participants, researchers asked questions about obscure movie details. People could use an AI assistant or answer without one, and they were always free to leave a question unanswered. The questions were deliberately chosen because the assistant often got them wrong.

In the first experiment, people without AI declined to answer about a third of the time. With access to AI, that dropped to roughly one in twenty. The pattern held in later experiments, including times when advice appeared without being requested. People also became more confident despite getting more answers wrong. Small financial rewards for accuracy helped, but didn’t restore their willingness to say they didn’t know.

Those rewards were small, though, and getting movie trivia wrong had little consequence. People might be more careful when more is at stake, or respond differently to a more reliable assistant. The study leaves that open.

Still, I think the point is valid.  it’s easy to get comfortable with an answer that makes sense, especially on an unfamiliar subject. A good explanation might be enough to carry on a conversation about it, even with little understanding of the details; at that point, checking the answer might feel unnecessary.

That’s what makes the familiar advice to “verify the output” perhaps harder to follow than it sounds. Verification takes work, and part of deciding to do that work is recognizing that you might be wrong.

And knowing that AI sometimes makes things up doesn’t make its mistakes easy to spot. An answer might sound perfectly reasonable to someone unfamiliar with the subject. Checking it means doing some of the work the AI was supposed to save, even when there’s no obvious reason to suspect a problem.

That extra work can be hard to justify when there’s a meeting in ten minutes. An AI summary can help someone quickly get up to speed on an unfamiliar topic and follow the discussion. The problem arises when a conclusion from the summary gets repeated without anyone checking the evidence behind it.

There’s plenty of value in getting help with something unfamiliar. The problem is when “I used AI to look into it” gets treated as if it means there’s nothing left to understand or verify.

Algorithm and Blues publishes Sundays.

Reference

1. Marcoccia, C., Quattrociocchi, W., & Capraro, V. (2026). AI advice suppresses people’s willingness to say “I don’t know”, even when the advice is wrong and accuracy is incentivized. arXiv preprint.