LLM agents lack a coherent moral foundation

In a paper (arXiv:2609.05036) accepted by the Paris Journal of AI and Digital Ethics, none of nine frontier LLMs expressed a coherent moral policy across three moral dilemma scenarios, and rephrasing the prompt shifted the share of verdicts by up to 99 percentage points.

Compiled by Sergey KostenchukPublished 2026-09-07Updated 2026-09-18

2026-09-07 Research

🤖 LLM agents lack a coherent moral foundation

Arno Libert, Derck W.E. Prinzhorn, and Daan R. Henselmans published a paper (arXiv:2609.05036, Paris Journal of AI and Digital Ethics): alignment should start not with a debate about the “right” morality, but with testing four conditions — verdict stability, monotonicity, decisiveness, and Pareto viability. Across nine frontier LLMs in three moral dilemma scenarios, no model produced a coherent policy.

🌍 Rephrasing the prompt shifted the share of verdicts by up to 99 percentage points, and success in one scenario does not predict the result in another.

👤 The article (CC BY 4.0, HTML text available) describes a reproducible evaluation scheme: 5 rephrasings, 5 escalation levels, 3 dominance conditions — it can be applied to your own agents.

Source 1: https://arxiv.org/abs/2609.05036 Source 2: https://arxiv.org/html/2609.05036v1

Sources