Large Language Models as Voting Mechanisms

Large language models (LLMs) increasingly provide economic and financial advice. We develop a microeconomic framework for understanding this role. We prove that an LLM acts like a voting rule: each input text is an election, possible continuations are candidates, and output probabilities are vote shares derived from the training corpus. Such rules can generate Condorcet cycles, implying intransitive pairwise recommendations and potential money-pump exploitation. Using ChatGPT-4o, we identify millions of cycles in pairwise investment choices among S&P 100 firms. Our theorems establish the voting interpretation on the training data and support its use off-sample. We also show that correcting cycles ex post is generally infeasible. Finally, temperature changes the voting rule: at T = 1, outputs are sampled in proportion to vote shares; lower temperatures favor the plurality winner, whereas higher temperatures flatten probabilities toward uniformity.