Researchers have discovered that people systematically misjudge how confident artificial intelligence systems actually are in their responses, leading users to place unwarranted trust in potentially unreliable answers.

The findings emerge from experiments examining how people interpret confidence signals from conversational AI agents like ChatGPT and Google's Gemini. When these systems present information with apparent certainty, users tend to overestimate the accuracy of those responses, even when the AI's actual confidence level is lower than perceived.

This gap between perceived and actual confidence poses a real problem. AI language models generate text based on statistical patterns in training data, not genuine understanding. They can produce convincing-sounding answers to questions they have no reliable basis for answering. Users who overestimate an AI's confidence are more likely to act on incorrect information without verification.

The mismatch occurs partly because people apply human social norms when interpreting AI communication. When a person speaks with conviction, confidence typically correlates with accuracy. Conversational AI systems, however, generate all outputs through the same mechanism regardless of whether they possess valid information. A chatbot discussing documented facts uses the same confident tone as when inventing plausible-sounding but false details.

The experimental work highlights a critical usability challenge as AI becomes embedded in everyday decision-making. Users consulting AI for medical advice, legal information, financial guidance, or academic research face particular risks if they overestimate system reliability. The researchers suggest interface designers should make AI uncertainty more transparent and visible to users.

Current AI systems do generate uncertainty estimates internally during processing, but these rarely appear in user-facing outputs. Making confidence scores explicit could help recalibrate user expectations. Alternatively, interfaces could prompt skepticism through design choices that signal when AI responses require human verification.

This research underscores the importance of AI literacy as these systems proliferate. Users need to understand that confident-sounding AI