This research identifies LLM delusions - a critical subset of hallucinations where models generate incorrect information with abnormally high confidence, making these errors particularly difficult to detect and mitigate.

• Delusions persist despite standard uncertainty measurements
• Unlike regular hallucinations, delusions show remarkably low uncertainty
• This phenomenon occurs across different model families and sizes
• Presents significant challenges for security and reliability of AI systems

From a security perspective, these high-confidence errors undermine trust in AI systems and create vulnerabilities that cannot be easily identified through conventional confidence checks, posing serious risks for applications requiring factual accuracy.

Delusions of Large Language Models