AI models appear to recognize moral complexity — then ignore it, new study by researchers affiliated with Harvard Kennedy School’s Allen Lab finds
Publish Date: 2026-05-27 14:03:00
Source Domain: ash.harvard.edu
- AI models, when faced with complex ethical dilemmas, express uncertainty but make consistent decisions, indicating an implicit value hierarchy rather than true ethical deliberation.
- The study “Crocodile Tears: Can the Ethical-Moral Intelligence of AI Models Be Trusted?” found that leading AI models resolved tragic ethical tradeoffs with uniformity, favoring options related to worker safety over environmental protections or vocational training.
- The authors introduce an ethical-moral intelligence framework focusing on expertise, sensitivity, coherence, and transparency, and suggest existing benchmarks are inadequate for evaluating moral reasoning in AI.
- The research calls for greater transparency in AI models’ ethical reasoning, suggesting systems should alert users to conflicting values, and proposes a “badging” system instead of monolithic benchmarks for model evaluation.
- Sarah Hubbard warns against trusting AI models as genuine ethical-moral agents and stresses the need for high standards of ethical-moral intelligence before these models are entrusted with decisions carrying moral weight.
