Study says AI access nearly ends people admitting they don't know
Researchers found that giving people AI answers, even wrong ones, made them almost three times less accurate while making them far more confident, The Decoder reported.
A study of 3,132 people found that access to an AI chatbot nearly eliminated their willingness to say "I don't know," The Decoder reported. That held even when the AI was almost always wrong.
Researcher Marcoccia and colleagues ran five experiments using trivia questions about minor visual details in films, such as uniform colors in Bend It Like Beckham. They chose questions like these because AI models tend to hallucinate on them. Without AI access, 36 to 44 percent of participants said they did not know an answer. With it, that fell to 3 to 6 percent, and to 1 percent when the AI's answer displayed without a prompt.
Accuracy fell as confidence rose. Participants without AI answered correctly 27.5 percent of the time, against 9.2 percent for those with AI access, about a third as accurate. Their stated confidence rose by a factor of 2.5, according to the study.
The authors wrote that the AI model was "almost always wrong" on the questions tested. They said the drop in "I don't know" answers "can't be explained as reasonable delegation to a reliable tool." Whether the same pattern holds with a more accurate model is a question the study leaves open.