Are LLMs random?

This article explores whether Large Language Models exhibit inherent biases when generating supposedly random outputs, revealing that their training data distributions can create unexpected patterns despite theoretical capabilities. By testing coin tosses and number generation across various models, the study highlights that these systems do not behave like true random number generators. Instead, they demonstrate consistent, measurable deviations from statistical norms, suggesting that LLMs reproduce subtle artifacts from their training rather than producing unbiased randomness. The most significant finding is a universal "heads" bias in coin toss simulations, where all tested models favored heads over tails to varying degrees, with some models showing extreme skew. In number generation tasks, a strong preference for odd numbers emerged in most models, though performance varied widely, with one model achieving perfect balance. These results indicate that bias is not uniform but depends heavily on the specific model and the type of random task, challenging the assumption that LLMs are neutral sources of randomness. This experiment is highly relevant to the open data community because it exposes a critical reliability issue for datasets generated or processed by AI. If LLMs introduce systematic biases into data creation, such as survey responses, synthetic data generation, or automated labeling, this can distort analysis and lead to flawed conclusions. Recognizing these inherent limitations is essential for anyone using AI-generated data, as it necessitates rigorous validation and awareness of potential skewness in AI-produced information.

Source: rnikhil.com
Published on 2025-04-27

Are LLMs random?

Related news