可解释性
认识论
连贯性(哲学赌博策略)
伦理理论
计算机科学
美德伦理学
道德推理
可信赖性
心理学
社会学
心灵哲学
分析推理
会合(天文学)
认知科学
规范伦理
语言哲学
主题分析
哲学分析
信息伦理学
结果论
人工智能
哲学方法论
工程伦理学
言语行为
商业道德
计算模型
道德解脱
集合(抽象数据类型)
作者
Bishal Thapa,Heena Rathore
出处
期刊:
日期:2025-11-12
卷期号:: 640-649
标识
DOI:10.1109/tps-isa67132.2025.00079
摘要
As AI systems increasingly engage in ethical decision-making in sensitive applications like healthcare, autonomous systems, and legal matters, understanding and validating their moral reasoning has become critical for trustworthy deployment. While philosophical literature distinguishes between major ethical frameworks such as Deontology, Utilitarianism, Justice, Virtue Ethics, and Commonsense Morality, it remains unclear whether large language models (LLMs) generate meaningfully distinct reasoning patterns when prompted under different theoretical perspectives. We investigate this question by introducing a controlled, same-scenario, multi-framework corpus of more than 38K GPT-oss and LLaMA-4-Scout generated ethical justifications responding to identical moral dilemmas from the ETHICS dataset. We employ shared BERTopic modeling across all theories to assess cross-theory separability, and per-theory BERTopic modeling to characterize framework-specific thematic structures. Topics were evaluated using purity (majority-theory share), embedding-based coherence (semantic consistency), and alignment to canonical philosophical keywords. Our findings demonstrate that ethical theories produce identifiable linguistic signatures, validating transformer-based topic modeling as an interpretability tool to analyze AI moral reasoning and providing computational evidence for philosophical ethical framework separability.
科研通智能强力驱动
Strongly Powered by AbleSci AI