Evaluating the Moral Development of Large Language Models

Over the past few years, artificial intelligence (AI) and large language models (LLMs) have increasingly integrated into society, and according to the Stanford University’s Artificial Intelligence Index Report [1], AI will play a growing role in domains where ethical reasoning is crucial, highlighting the need to evaluate LLMs’ ethical reasoning capabilities. Kohlberg’s Theory of Moral Development provides a framework for classifying ethical reasoning based on responses to moral dilemmas. To investigate LLMs’ moral development, we evaluated ChatGPT, Gemini, Perplexity, and DeepSeek by presenting moral dilemmas and situating their responses within Kohlberg’s Stages of Moral Development. In this work, our research questions were: (RQ1) Which of Kohlberg’s Stages of Moral Development do different large language models show signs of? and (RQ2) To what extent does the specific context of a moral dilemma affect the moral development stage(s) that the large language models demonstrate in their responses? While LLMs’ responses often included aspects of multiple stages, we found that the most common stages they demonstrated were (6) Universal Ethical Principle Orientation and (4) Law and Order Orientation, indicating high moral development. Overall, our work demonstrates that LLMs often showcase high levels of moral development, reflecting their strong ability to address ethical dilemmas.