Measuring and Analyzing Entropy in Large Language Models

2025-03-11

Summary: A sourced summary of a benchmark study on how random or predictable different LLM outputs are under number-generation prompts.

This sourced technical post studies randomness in large language models. The author tested many models with different prompts and compared how unpredictable their outputs were. The post says models often show patterns and biases even when asked for random outputs, such as preferring certain numbers.

The practical lesson is that LLMs are not good sources of true randomness. Their responses can look random while still being shaped by training data, prompt wording, and sampling settings. For security or gambling-like randomness, use proper random-number tools instead of asking an AI model.

Free Basics version: AI can explain randomness, but it should not be trusted as a secure random generator.

Source: Daniil Sedov, Gusarich's thoughts, gusarich.com/blog/measuring-llm-entropy