# Summary: 2026-08-14_17-43-16Z_ParticipatoryMoralAIIsNotNeutral_TheInvisibleHando.md
Saved: 2026-08-16 22:03
Source: 2026-08-14_17-43-16Z_ParticipatoryMoralAIIsNotNeutral_TheInvisibleHando.md
Original paper: [arXiv](http://arxiv.org/abs/2608.14522v1)
Model: None

---

## Summary  
The paper argues that participatory moral AI is not neutral because the developers who design it make opaque choices that shape how people vote on moral dilemmas. By tracing three key stages—feature scoping, voter sampling, and question framing—the authors demonstrate that each can bias the aggregated preferences used to train an AI system. Their empirical study across multiple deployment contexts shows that these biases are systematic rather than accidental. The work calls for transparency and auditing of every stage in the moral AI elicitation pipeline.

## Key Contributions  
- [Finding 1] Morally relevant features shift across different deployment contexts, indicating that feature schemas should not be assumed to transfer universally.  
- [Finding 2] Preferences differ by political ideology for roughly one‑third of features, and some differences reverse direction depending on the ideological composition of the voter pool.  
- [Finding 3] The wording of the elicitation question can narrow or widen ideological gaps by up to a full scale point and alters how moral foundations are linked to judgments.

## Methodology  
The authors conducted an empirical study with two data collection phases (N = 809) across three deployment contexts: AI‑driven kidney allocation, AI agents simulating absent workers, and generative AI depictions of the deceased. In each phase they varied the three main stages—feature scoping (which dilemmas are presented), voter sampling (who participates), and question framing (how the dilemma is worded)—and recorded how these choices affect aggregated moral preferences.

## Results  
Across contexts, the set of features that participants consider morally relevant varies, confirming a finding 1. For about one‑third of those features, ideological alignment shifts with voter composition, supporting finding 2. Question framing narrows or widens the ideological gap by up to a full scale point and reshapes moral foundation associations, evidencing finding 3.

## Significance  
These results show that voting‑based alignment cannot deliver fair or transparent AI simply because preferences are aggregated; each stage of the pipeline is normative and must be disclosed. The paper underscores that developers bear responsibility for the “invisible hand” that steers moral outcomes, prompting a call for auditing and transparency throughout the elicitation process.

## Related Concepts  
- Moral preference elicitation  
- Feature scoping  
- Voter sampling  
- Question framing  
- Ideological bias in AI design  
- Moral foundations theory  
- Participatory AI  
- Aggregation of preferences