Measuring LLM Reliability With Semantic Entropy in Production Systems

Researchers at Oxford developed semantic entropy, a technique that measures AI model reliability by running the same input multiple times and analyzing whether responses reach consistent conclusions. The method embeds multiple model outputs, clusters them by similarity, and calculates Shannon entropy to identify when models are uncertain—entropy scores above 1.0 indicate unreliable responses that should be routed to humans. In content moderation systems, 31% of problematic posts showed high entropy and routing these cases to human reviewers cut false positives by over half.