Automated and statistical methods for estimating latent political traits and classes from textual data hold great promise, because virtually every political act involves the production of text. Statistical models of natural language features, however, are heavily laden with unrealistic assumptions about the process that generates these data, including the stochastic process of text generation, the functional link between political variables and observed text, and the nature of the variables (and dimensions) on which observed text should be conditioned. While acknowledging statistical models of latent traits to be “wrong,” political scientists nonetheless treat their results as sufficiently valid to be useful. In this article, we address the issue of substantive validity in the face of potential model failure, in the context of unsupervised scaling methods of latent traits. We critically examine one popular parametric measurement model of latent traits for text and then compare its results to systematic human judgments of the texts as a benchmark for validity.