Researchers Expose Hidden Developer Choices in Moral AI Elicitation
A recent study published on arXiv (2608.14522) indicates that participatory moral AI is inherently biased, as developers make three crucial decisions—feature scoping, voter sampling, and question framing—that determine the preferences generated through moral AI elicitation. This research, which included 809 participants across two phases and three contexts (AI kidney allocation, AI agents mimicking absent workers, and generative AI representations of th), shows that these often hidden and undocumented choices can greatly affect results. The authors assert that these decisions are normative, not just technical, and advocate for increased transparency and accountability in moral AI system development. The identities of the researchers are not disclosed in the summary.
Key facts
- The paper is titled 'Participatory Moral AI Is Not Neutral: The Invisible Hand of Developers'.
- It was announced on arXiv with ID 2608.14522.
- The study involved 809 participants across two phases.
- Three deployment contexts were examined: AI kidney allocation, AI agents simulating absent workers, and generative AI depictions of th.
- Developers make three key choices: feature scoping, voter sampling, and question framing.
- These choices are often opaque, undocumented, and treated as technical rather than normative.
- Each choice can shape the preferences produced by moral AI elicitation.
- The paper calls for recognition of these choices as normative and for greater transparency.
Entities
Institutions
- arXiv