LLM Voting: LLM Collective Decision Making and Human Choices

This paper examines the voting behaviours of Large Language Models (LLMs), specifically OpenAI's GPT-4 and LLaMA-2, in comparison to human voting patterns. Our methodology involved conducting a human voting experiment to establish baseline human preferences, followed by a similar experiment with LLMs. We analysed both collective outcomes and individual preferences, uncovering notable differences in decision-making processes and biases between humans and LLMs. Our findings reveal a significant disparity in preference diversity, with LLMs displaying more homogeneity compared to humans. These results suggest that using LLMs for voting assistance could potentially influence collective outcomes in a uniform direction, highlighting the importance of considering these differences in democratic processes.