Out of the (Black)Box: AI as Conditional Probability

Abstract

We explore the economic significance and interpretability of the distribution of conditional probabilities behind LLM's text generation. Using a dataset of news and returns, we find that conditional probabilities are interpretable and correlate with model accuracy. Conversely, measures of declared confidence used in the literature are opaque, structurally biased, unstable, and more model-dependent, indicating that LLMs cannot assess their own confidence. Using conditional probabilities, we analyze LLM biases and provide insights into the internal mechanisms driving model decisions. Our results indicate that conditional probabilities provide a reliable and transparent reflection of LLM beliefs, particularly for economic applications.