Probing Ethical Framework Representations in Large Language Models: Structure, Entanglement, and Methodological Challenges

Researchers analyzed how large language models (4B-72B parameters) internally represent different ethical frameworks, finding that models create distinct ethical subspaces but with asymmetric transfer patterns between frameworks. The study reveals structural insights into AI ethics processing while highlighting methodological limitations in probing techniques.

• →Large language models maintain differentiated internal representations for various ethical frameworks rather than collapsing ethics into a single dimension.
• →Ethical framework probes show asymmetric transfer patterns, with deontology generalizing to virtue scenarios while commonsense fails on justice scenarios.
• →Disagreement between deontological and utilitarian approaches correlates with higher behavioral uncertainty across different model architectures.
• →Probing methods partially depend on surface features of benchmark templates, requiring cautious interpretation of results.
• →The research provides structural insights into AI ethics processing while acknowledging significant epistemological limitations.