Large Language Models meet moral values: A comprehensive assessment of moral abilities

Abstract

Automatic moral classification in textual data is crucial for various fields including Natural Language Processing (NLP), social sciences, and ethical AI development. Despite advancements in supervised models, their performance often suffers when faced with real-world scenarios due to overfitting to specific data distributions. To address these limitations, we propose leveraging state-of-the-art Large Language Models (LLMs) trained on extensive common-sense data for unsupervised moral classification. We introduce an innovative evaluation framework that directly compares model outputs with human annotations, ensuring an assessment of model performance. Our methodology explores the effectiveness of different LLM sizes and prompt designs in moral value detection tasks, considering both multi-label and binary classification scenarios. We present experimental results using the Moral Foundation Reddit Corpus (MFRC) and discuss implications for future research in ethical AI development and human–computer interaction. Experimental results demonstrate that GPT-4 achieves superior performance, followed by GPT-3.5, Llama-70B, Mixtral-8x7B, Mistral-7B and Llama-7B. Additionally, the study reveals significant variations in model performance across different moral domains, particularly between everyday morality and political contexts. Our work provides meaningful insights into the use of zero-shot and few-shot models for moral value detection and discusses the potential and limitations of current technology in this task.