Framing Instability in LLM Ethical Stance: Auditing Negation Sensitivity in Moral Dilemmas
Updated
Updated · arxiv.org · Jul 8
Framing Instability in LLM Ethical Stance: Auditing Negation Sensitivity in Moral Dilemmas
1 articles · Updated · arxiv.org · Jul 8
Summary
A new audit reveals that large language models (LLMs) show significant instability in their ethical stances when moral questions are phrased differently.
Testing 16 models across 14 dilemmas, researchers found endorsement rates could swing by up to 76 percentage points depending on question framing.
This framing sensitivity raises concerns for deploying LLMs in advisory roles, as their moral judgments may depend more on wording than substance.