Is there consistency in ethical sensitivity in artificial intelligence? A review of language models

This study examines the structural and content consistency of large language models (LLMs) in ethical decision making with a qualitative approach. Responses to basic ethical themes such as “justice”, “non-maleficence”, “autonomy”, “impartiality”, and “goodness” were evaluated using the thematic analysis method of Braun and Clarke (2006). The three-stage coding process analyzed empathy patterns, contextual transitions, and relationships between themes. The findings, supported by Python-supported frequency and variation analyses, revealed that the models exhibited high empathy and solution determination in the themes of “inclusiveness” and “communication” but low structural consistency in the themes of “religion” and “disability”. The responses to the same ethical theme in different contexts were determined to carry semantic shifts. This original study emphasizes that ethical sensitivity should be evaluated based on patterns.