ResearchCisent
"I don't know", penalized three times: a study shows AI takes it away from people too
With an AI at hand, people almost stop saying "I don't know": they're right a third as often and twice as confident. It's the third penalty, after the two Cisent's paper describes.

Knowing when to say “I don’t know” is part of good judgment. A study published in July by researchers at the École Normale Supérieure, Sapienza University of Rome and the University of Milan-Bicocca measured what happens to that ability when an AI is at hand: it almost disappears.1 It’s a preprint and hasn’t been peer reviewed yet.
What they measured
They ran five experiments with 3,132 people, four of them preregistered. Participants answered difficult questions where they could always say “I don’t know”, and the AI in the experiment was wrong on purpose.1
With the AI available, “I don’t know” answers fell from between 17% and 44% to between 1% and 8%. People answered more, but they were right a third as often as without AI, from 27.5% to 9.2%, and their confidence rose from 30 to 76 out of 100. Paying for correct answers helped little: “I don’t know” stayed far below normal. And the effect held when the AI showed up on its own, without anyone asking for it.1
The authors’ explanation is that AI assistants have a fluent answer for almost anything, and that shifts the threshold at which a person decides they know enough to answer.1
The third penalty
In September, Sebastián Salaberry-Rivera, Cisent co-founder, AI researcher and creator of the Manthys models, published “‘I Don’t Know’, Penalized Twice”.2 His thesis is that an AI model is punished twice for saying “I don’t know”: training rewards guessing, and the benchmarks that dominate the leaderboards have no place for it. It builds on the work of Adam Tauman Kalai and his team, published this year in Nature, which shows that evaluating models only on accuracy pushes them to make things up.3
The July study measures, on the human side, what that paper argues on the machine side. The penalty reaches whoever uses the AI: if the model never says “I don’t know”, the person stops saying it too. That’s the third penalty.
The other half of Salaberry-Rivera’s thesis is that “I don’t know” can be trained: a model can learn to abstain when it doesn’t know, instead of guessing. The full explanation is in Research.
What we don’t know yet
Nobody has measured what happens to people when the AI does say “I don’t know”. Do they get their own “I don’t know” back? That’s the next question.
Sources
Chiara Marcoccia, Walter Quattrociocchi and Valerio Capraro, “AI advice suppresses people’s willingness to say ‘I don’t know’, even when the advice is wrong and accuracy is incentivized”, arXiv 2607.13562, July 15, 2026. Preprint, not peer reviewed. https://arxiv.org/abs/2607.13562↩︎
Sebastián Salaberry-Rivera, “‘I Don’t Know’, Penalized Twice: Routed abstention as a trained disposition, and the evaluation apparatus that punishes it”, Zenodo, 2026. https://doi.org/10.5281/zenodo.23032426↩︎
Adam Tauman Kalai, Ofir Nachum, Santosh S. Vempala and Edwin Zhang, “Evaluating large language models for accuracy incentivizes hallucinations”, Nature 653, 1047-1051, April 22, 2026. https://doi.org/10.1038/s41586-026-10549-w↩︎