Generalization of SKeB findings to sensitive, non-fictional domains

Ascertain whether the observed relationships between persuasive prompt framing, domain graph entanglement metrics, and factual recall in unlearned models generalize to sensitive domains such as personally identifiable information, harmful content, and copyrighted material, given potential differences in how fictional versus factual/personal information is encoded.

Background

The paper’s experiments focus on the Harry Potter fictional domain to avoid ethical complications, constructing a domain graph and measuring entanglement to predict residual recall after unlearning.

The authors explicitly flag uncertainty about applying these findings to real-world sensitive domains (PII, harmful content, copyrighted material), noting that fictional knowledge may be encoded differently than factual or personal information, which could affect both entanglement structure and susceptibility to persuasive framing.