Researchers increasingly construct text-based empirical measures with Large Language Models, often treating model outputs as observed data rather than generated regressors. Using S&P 500 Item 1A risk disclosures from 2005–2024, we show that LLM-derived measures contain non-classical measurement error driven by two distortions: compression and injection. Compression is substantial: summaries retain only 40–63% of source vocabulary and selectively omit source-text language. Injection is also pervasive and not merely editorial: summaries introduce novel terms absent from the source that account for 16–42% of their most discriminative vocabulary, often changing salience, tone, and substantive content. Together, compression and injection directionally reshape narrative emphasis and predict outputs from downstream LLM tasks, including categorization, scoring, and ordinal classification. These distortions predict severity, specificity, complexity, and boilerplate measures, and generate firm-outcome associations consistent with non-classical measurement error. We document these patterns across models and textual constructs and propose diagnostics for structured LLM measurement error.