AI advice reduces accuracy while increasing confidence, study finds
Research shows users become less accurate but more confident with AI assistance.
1Surprising findings
A joint study by three universities revealed that AI advice reduced users' tendency to say 'I don't know' from 44% to 3%. Accuracy dropped from 27% to 9%, while confidence surged from 30% to 76%. These contradictory results indicate a significant shift in cognitive processes.
2Trust in wrong answers
The study deliberately used Step 3.5 Flash model for questions it typically gets wrong, like visual details from movies. Participants adopted AI's incorrect advice even when they could answer correctly alone. This challenges the 'smart delegation' argument.
3Limited impact of monetary incentives
Even with financial rewards for correct answers, willingness to admit ignorance only increased from 3% to 8%. Accuracy improved slightly from 9% to 16%, both still far below non-AI baseline levels of 44% and 27% respectively.
4Cognitive surrender
The phenomenon termed 'cognitive surrender' by Wharton researchers shows users accept AI's wrong answers 80% of the time. The new study demonstrates that AI's mere availability suppresses the cognitive habit of recognizing one's knowledge limits.
5Educational risks
Researchers express particular concern for children whose critical thinking skills are still developing. Google's AI search summaries and similar applications have been called an 'unacceptable risk' by Common Sense Media. AI systems are designed to always answer, never to say 'I don't know'.
6False confidence trap
The study reveals AI interactions artificially inflate human confidence. Capraro from University of Milano-Bicocca warns, 'The capacity to recognize what you don't know is fundamental to human learning,' highlighting long-term implications of this trend.
Original article: The Next Web — Yapay zeka tavsiyeleri doğruluğu azaltırken güveni artırıyor