LLM moral choices lack stable value rankings, averaging fakes human-like hierarchy: arXiv study
rohanpaul_ai · x · 2026-11-11
An arXiv paper tests whether LLM moral choices satisfy transitivity — a minimal necessary condition for genuinely weighing competing values. Testing 5 LLMs (incl. GPT-5.2, Gemini) on pairwise moral trade-offs, most choice sets showed only weak or unstable orderings. Worse, averaging an LLM's responses — common practice in moral evaluations — produces a transitive, human-looking hierarchy that matches no individual choice set. The authors argue AI morality evaluations must verify structure before asking about alignment.
Related event: Study: LLMs Lack Stable Moral Value Rankings(2 posts)→
More from Research
- EarthRecord benchmark tests 12 multimodal agents on four decades of raw satellite imagery — AIEyesTech · 2026-11-11
- WaveGen: codec-free end-to-end diffusion hits SOTA text-to-speech via Spectral Forcing — Ajou · 2026-11-11
- Berkeley's Social Humanoid teaches Unitree G1 socially responsive interaction — berkeley_ai · 2026-11-11
- Robotics is not solved: frontier models are ~6 orders of magnitude off real assembly lines — ZeeshanZiaML · 2026-11-11
- Interpretability study: Qwen3's 'say' and 'do' circuits are structurally disconnected, unlike OLMo — Sauers_ · 2026-11-11
- Google's AudioFM Foundation Model Tops Audio Benchmarks, Powers 2B Daily Predictions — umpedronosapato · 2026-11-11