Assess cross-cultural generalizability of MGE

Determine how the Moral Graph Elicitation values-elicitation and reconciliation process performs in countries with significantly different cultures—including Nigeria, China, and the United Arab Emirates—and whether observed value convergence and legitimacy persist outside the United States.

Background

The case study was conducted with a representative sample of Americans and showed value convergence and encouraging legitimacy metrics.

The authors explicitly flag uncertainty about how well MGE transfers to other cultural contexts and cite examples of countries with different cultures as targets for future assessment.

References

Even though we found value convergence for contentious topics like abortion across the political spectrum, it remains to be seen how well our values elicitation process works in countries with very different cultures, like Nigeria, China, and the United Arab Emirates.

— What are human values, and how do we align AI to them?  (2404.10636 - Klingefjord et al., 2024) in Section 6 (Limitations) – Case Study

A second, complementary mitigation comes from recent translation research: \citet{skorski_moral_2026} shows that moral semantics are largely preserved under careful LLM-based translation (EN$\to$PL, AUROC gaps of 0.01--0.02 across foundations), pointing to a practical route for testing --- and eventually correcting --- this bias without new from-scratch annotation: translate an existing English corpus and re-run the Moral Entropy audit on it directly. We leave this cross-lingual test to future work.

— Moral Entropy: Auditing Bias and Uncertainty in Moral Judgment  (2609.21992 - Skorski, 18 Sep 2026) in Limitations, paragraph “Cultural and annotator-pool bias”