Measure and mitigate model-induced bias in transition stories

Determine the extent to which GPT-4–generated value transition stories influence participant wisdom judgments and develop methods to measure and reduce such model-induced bias when constructing edges in the moral graph.

Background

Edges in the moral graph are collected by asking participants to judge value transitions presented as GPT-4–generated stories. This design brings potential susceptibility to LLM-generated persuasive content.

The authors explicitly call for further work to quantify how much participant judgments can be swayed by story generation and to improve robustness accordingly.