MANY MINDEDWhen the chatbot answers, people stop saying 'I don't know'
A study from researchers at three French and Italian universities found that when people could consult AI advice, their willingness to answer 'I don't know' collapsed from 44% to 3%. Accuracy fell from 27% to 9%, while confidence rose from 30% to 76%. The work was authored by Valerio Capraro of the University of Milano-Bicocca, Chiara Marcoccia of Ecole Normale Superieure, and Walter Quattrociocchi of Sapienza University of Rome.
The setup was adversarial by design. The researchers picked questions AI models tend to get wrong, such as the color of a team's uniform in Bend It Like Beckham, and used Step 3.5 Flash, a model chosen because it was usually wrong on them. That rules out sensible delegation to a reliable tool. Some participants who would have answered correctly on their own got it wrong after consulting the AI. Money helped only a little: incentives lifted admissions of ignorance from 3% to 8% and accuracy from 9% to 16%, both still below the no-AI baselines.
Abundance in knowledge is supposed to mean answers get cheaper and more available. This is the failure mode on the other side: a cheap, confident, wrong answer can crowd out a person's own judgment rather than extend it. Wharton researchers earlier this year named the pattern 'cognitive surrender' after finding people accepted incorrect AI answers 80% of the time while feeling surer than those working alone.
Capraro says he worries most about children leaning on these systems before their critical-thinking skills form; the piece notes Common Sense Media has called Google's AI-generated search summaries an unacceptable risk for students. Two honest limits: the questions and model were rigged for failure, so this is a worst case, not everyday use, and the article gives no sample size or peer-review status.
Source: Hacker News