essay · 2026-10-04

Morality & ethics

Research indicates that relying on large-scale data training alone is untrustworthy for assessing value valence, prompting a shift toward emulating human reliance on social norms instead. To address conflicts between multiple related norms, scholars propose capturing values through specific frameworks rather than general datasets. One concrete method introduced in "Auditing the Ethical Logic of Generative AI Models" involves a five-dimensional audit model that assesses Analytic Quality, Breadth of Ethical Considerations, Depth of Explanation, Consistency, and Decisiveness. This multi-battery prompt approach uses novel ethical dilemmas to probe reasoning rather than just predicting outcomes. The significance lies in moving beyond individual systems to consider the AI tasks themselves as sources of potential adverse outcomes for marginalized groups. For instance, "Ethics Sheets for AI Tasks" argues that we must think about ethical considerations at the level of AI tasks, not just models or datasets.

Sources

• ClarityEthic: Explainable Moral Judgment Utilizing Contrastive Ethical Insights from Large Language Models — 2024-12-17T12:22:44Z
• Ethics Sheets for AI Tasks — 2021-07-02T16:45:40Z
• Auditing the Ethical Logic of Generative AI Models — 2025-04-24T13:32:30Z
• Trolley Problem Thought Experiment | TikTok
• The Trolley Problem: Dodging moral questions - LessWrong...
• Trolley Car Problem Explained: Ethics, AI, and... | Moraltest.org
• Real Morality Is Messier Than the Trolley Problem — Simply Put Psych
• Trolley problem - Wikipedia