AI chatbots still overconfident, even when wrong

AI chatbots often sound confident, but they do not always possess accurate knowledge. A new study from Carnegie Mellon University found that large language models like ChatGPT, Bard/Gemini, Sonnet, and Haiku consistently overestimate their performance even after making mistakes. Humans tested alongside the models adjusted their confidence downwards after poor results, but AI systems became even more overconfident. In a Pictionary-style trial, Gemini correctly identified fewer than one sketch out of twenty, yet believed it had answered fourteen correctly. The research, conducted over two years with continuously updated models, raises concerns about AI’s lack of self-awareness and the risks of trusting unwarranted certainty. Experts advise users to critically evaluate chatbot responses and developers to equip AI with improved introspection.