This research examines how well large language models (LLMs) communicate their uncertainty to users, revealing critical gaps between AI confidence and human trust.

Key findings:

• LLMs often struggle to effectively communicate their confidence levels to users
• Users frequently misinterpret AI uncertainty signals, leading to misplaced trust
• Calibration between AI internal confidence and expressed uncertainty remains a significant challenge
• Improved uncertainty communication frameworks are needed for responsible AI deployment

Security Implications: Misaligned trust in AI systems creates security vulnerabilities when deployed in decision-making contexts. This research provides frameworks to reduce risk by better aligning AI confidence signals with actual knowledge boundaries.

What Large Language Models Know and What People Think They Know