Researchers introduce MoralAltDataset, a collection of 307 moral dilemmas augmented with compromise and reframed alternatives, to study how large language models handle options beyond binary choices. The study examines whether humans and LLMs shift their judgments when these alternatives are presented.

  • The dataset includes narrative Advisor dilemmas and AI-facing Agent dilemmas.
  • Across 15 LLMs, compromise alternatives were often preferred over original options, substantially reshaping moral choice.
  • LLM-generated alternatives were evaluated against human-authored ones using pairwise preference and expert-based criteria.
  • Results indicate that LLM-generated alternatives are often preferred and better satisfy fine-grained structural and ethical criteria.

The findings reveal trade-offs between the structural quality of generated alternatives and their practical feasibility.