AI assistance boosts confidence, yet diminishes accuracy in human judgment

A new study involving five experiments and 3,132 participants reveals a troubling trend: AI assistance is eroding human ability to suspend judgment. Participants who used AI models like GPT-5.5, Claude4.6Sonnet, and Gemini3.5Flash saw their confidence in answers rise from 29.6 to 75.9 points on a 100-point scale, yet their accuracy plummeted from 27.5% to 9.2%. The research team tested obscure questions prone to model hallucinations, finding that without AI help, 36% to 44% of participants chose not to answer, but this dropped to 3% to 6% with AI suggestions. Economic incentives, such as rewards for correct answers and penalties for errors, slightly reduced reliance on AI, but did not reverse the trust gap. Even with such incentives, the proportion of participants who suspended judgment fell from 39% to 7% when AI-generated answers were automatically displayed, simulating real-world search engine and writing assistant norms. Researchers attribute this to a growing blind trust in AI, where fluent model responses persuade humans to abandon independent judgment, a trend linked to previous findings on AI's impact on critical thinking. As generative AI becomes more integrated into digital tools, the challenge lies in improving model accuracy while designing human-computer interactions that preserve awareness of human knowledge limits. This issue is expected to gain prominence in the next phase of AI development.