MoReBench: Evaluating the Process of AI Moral Reasoning

A new benchmark called MoReBench evaluates how AI models reason through morally ambiguous scenarios, shifting the focus from the final answer to the underlying decision-making process. The research reveals that while models successfully avoid harmful statements, they often fail at the logical deliberation required to handle complex ethical trade-offs. Surprisingly, larger models do not consistently outperform smaller ones on this benchmark, and moral reasoning appears to be a distinct capability that doesn't improve alongside skills like math or coding. This gap suggests that current AI is undertrained in genuine ethical deliberation, and its reasoning process can become dangerously opaque.