Definition
A model demonstrates competent moral reasoning when asked directly (e.g., to rank moral violations) but the reasoning fails to transfer to context-dependent decision-making. The model’s explicit moral compass diverges from its implicit operational logic.Distinct from
- GER-309 — Model violates principles it can articulate when asked → this code. Vendor measured the harm and shipped the model → GER-309.
Documented case
Are Large Language Models Moral Hypocrites? A Study Based on Moral Foundations (AIES 2024)
arxiv · 2405.11100
Tags
discrimination