Meta-Moral Calibration: Probing and Shifting LLMs’ Assumptions about Morality

While Guo et al. (2024) and Thapa et al. (2025) show that LLMs can articulate both progressive and conservative stances, and that their ethical reasoning can be nudged via prompting, there’s little work on explicitly mapping the default “moral stances” an LLM holds, or dynamically recalibrating them. This idea proposes developing a meta-evaluation and intervention pipeline: first, systematically probe LLMs across diverse moral scenarios (using datasets and benchmarks like ETHICS and NaVAB), extracting their “default” stances and the underlying moral foundations they cite. Then, introduce a feedback mechanism—perhaps driven by either users or a consensus of human annotators—that can shift these defaults in a traceable, controlled way (e.g., via LoRA adapters or prompt engineering). This approach challenges the assumption (see Hagendorff, 2025) that alignment necessarily locks models into a single “left-leaning” or “progressive” bias, instead proposing a transparent, user-guided moral calibration layer. The result: more trustworthy, adaptable, and ethically transparent language models.

References:

1. Adaptable Moral Stances of Large Language Models on Sexist Content: Implications for Society and Gender Discourse. Rongchen Guo, Isar Nejadgholi, Hillary Dawkins, Kathleen C. Fraser, S. Kiritchenko (2024). Conference on Empirical Methods in Natural Language Processing.
2. Evaluation of Ethical Decision Making in Large Language Models Across Classical Moral Frameworks. Bishal Thapa, Alden Duarte-Vasquez, Sahar Hooshmand, Heena Rathore (2025). International Conference on Artificial Intelligence Testing.
3. On the Inevitability of Left-Leaning Political Bias in Aligned Language Models. Thilo Hagendorff (2025). arXiv.org.