The Vanishing Evidence Link: A Preliminary Content Analysis of Evidence-Link Attenuation in AI-Generated Explanations of Public Subsidy Information Pages

This technical note presents a preliminary content analysis of evidence-link attenuation in AI-generated explanations of public subsidy information pages. Using the same three public subsidy information pages examined in previous “Vanishing” series notes, this study compares a normal explanation condition with an evidence-link-preserving explanation condition. The analysis introduces the Evidence Link Unit (ELU) as an operational unit for examining whether explanations, judgments, and guidance in AI-generated outputs can be traced back to specific source-text conditions, exceptions, cautions, procedures, required documents, contact points, or related-document references. Across 36 ELUs, the normal explanation condition preserved 32 ELUs and weakened 4, while the evidence-link-preserving condition preserved all 36 ELUs. The weakened cases mainly involved composite evidence links, semi-explicit evidence, conditions spanning multiple contexts, and role differences between contact points and administrative departments. The results do not suggest that evidence links uniformly disappear in AI-generated explanations. Rather, they suggest that some evidence links may become less traceable, especially when they depend on multiple source-text statements or semi-explicit source cues. The note also discusses how evidence-link-preserving prompts may improve traceability while introducing explanatory reconstruction. This work is a small-scale preliminary observation and should not be generalized without further validation.