AI & Enterprise
Users fail to spot AI errors as fluent answers drive \'cognitive surrender\', study finds
A study found 73.2 percent of large language model chatbot users accepted AI answers even when they were wrong. Researchers at the University of Pennsylvania called the phenomenon “cognitive surrender”, in which people accept AI reasoning without checking it. In tests involving 1,372 participants and more than 9,500 trials, acceptance remained high even when the chatbot was wrong. Rewards and immediate feedback helped users overturn errors, while time pressure reduced corrections.
