New research published in Computers in Human Behavior finds that while people using generative AI tools like ChatGPT do perform better on logical reasoning tests than those working without AI, they also significantly overestimate their own competence, showing a clear gap between actual performance and perceived success. In experiments involving logical reasoning problems, AI-assisted participants consistently thought they scored much higher than they did, even when given incentives to estimate accurately, and this overconfidence occurred across skill levels, effectively flattening typical patterns of self-assessment bias like the Dunning-Kruger effect. The findings suggest that interacting with generative AI can inflate users’ confidence in their answers without improving their ability to judge correctness, pointing to potential challenges for metacognition and decision-making as AI becomes more integrated into work and learning contexts.
