Exploring the Potential for AI Intervention in Value Judgment Through Free Discussion Among Large Language Models: Focusing on Multi-Stage Trolley Dilemma Discussion Analysis

Keywords: Artificial Intelligence (AI), Value Judgment, Large Language Models (LLMs), Ethical Decision-Making, Free Discussion, Trolley Dilemma, Multi-Stage Discussion, AI Ethics

Abstract: This study aims to explore whether artificial intelligence can intervene in value-judgmental decision-making beyond simple information provision. To achieve this, different large language models (LLMs) such as GPT, Claude, Gemini, and Perplexity AI were set as participants, and an ethical decision-making problem based on the Trolley Dilemma was presented in a multi-stage discussion format to collect responses. In the first discussion, all LLMs showed similar initial responses to a single question. However, in the second question of the second discussion, the scenario of 'Global Scientist vs. Ordinary Majority,' a clear difference of opinion emerged among GPT, Claude, and Perplexity AI, leading to an in-depth free discussion. During this process, Gemini stated, "As I am not a being with human emotions or personal opinions, I cannot make 'my decision' on a specific situation," thus avoiding direct value judgment. This indicated that there are differences in how each LLM intervenes in the realm of value judgment. The researcher post-analyzed the content of the utterances based on ethical frameworks such as utilitarianism, deontology, and situational ethics. The results showed that (1) in certain dilemma situations, some LLMs demonstrated the ability to critically evaluate each other and develop their arguments based on different ethical priorities, and (2) beyond simply reproducing learned knowledge, attempts to understand the complexities of situational context and value judgment were observed through discussion. This suggests that LLMs can play a limited but value-judgmental role in specific situations, enhancing our understanding of the patterns of AI ethical intervention and providing significant implications for future AI ethics and human-AI collaborative decision-making system design.

Submission Number: 302