Stance is not a Construct: Validity Gaps in LLM Annotation of Political Attitudes

In Prep

A critical examination of how LLMs are used to annotate political attitudes, arguing that stance detection conflates distinct constructs and produces systematic validity gaps. The paper develops a framework for evaluating when LLM annotation captures genuine political attitudes versus surface-level textual features.

Methods: LLM annotation, computational text analysis · Data: Political text corpora