Dice Question Streamline Icon: https://streamlinehq.com

Evaluate transitivity assumptions in moral graph wisdom relations

Ascertain whether "wiser than" relations between values in the moral graph are transitive beyond a single step and determine the implications for PageRank-based aggregation and downstream training when transitivity may fail.

Information Square Streamline Icon: https://streamlinehq.com

Background

The moral graph encodes directed edges indicating that one value is wiser than another for a given context. PageRank is used to identify winning values, implicitly leveraging transitive information across edges.

The authors note they assume some transitivity of moral judgments and observe limited empirical support (for one step), but emphasize that this assumption requires further evaluation.

References

Note that we assume a level of transitivity here for moral values. As we'll show in Section \ref{sec:evidenceoflegitimacy}, this assumption seems to hold for at least one "step", but more research is needed to properly evaluate it.

What are human values, and how do we align AI to them? (2404.10636 - Klingefjord et al., 27 Mar 2024) in Section 4.2 (What is a moral graph?)