Year: 2025
Type: conference-paper
Abstract: Large Language Models (LLMs) are increasingly deployed in contexts requiring moral judgment, yet existing benchmarks largely emphasize surface-level acceptability (e.g., “Is it okay to...”), overlooking the complex ethical trade-offs that humans consider. We introduce Conscience Conflict, a new evaluation suite comprising six human authored, high-stakes moral dilemmas grounded in ethical theory. Each vignette presents three mutually exclusive decisions, requiring both a choice and a free-text justification. Model responses are annotated using a taxonomy of ethical frameworks (e.g., deontology, utilitarianism, virtue ethics). Evaluating five open-source LLMs, we observe sharp shifts in moral reasoning across vignettes, with no consistent ethical orientation. These findings expose gaps in moral coherence that simpler tests fail to detect. Our evaluation framework emphasizes the need for more nuanced and transparent assessments of moral reasoning in LLMs.
Authors: Hota, Asutosh, Jokinen, Jussi P. P.
Institution: University of Jyväskylä
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7212342094