Are LLMs Bad at Moral Reasoning?

The paper evaluates the moral reasoning capabilities of large language models (LLMs) using the MoReBench dataset, which includes gold-standard human-authored rubrics for 1,000 cases. Initial benchmarks showed underwhelming results, but the authors propose a novel approach where LLMs generate their own scoring rubrics for moral analysis, revealing that these generated rubrics are better aligned with human standards. This suggests that LLMs may possess greater moral reasoning capabilities than previously assessed, which is critical for developing AI systems that can operate safely in complex environments.