Open Collections

Abstract

Traditional models of perception are ill-equipped for the high-dimensional data, such as text embeddings, that are central to modern AI and psychological science. To address this, we introduce the Double Machine Learning Lens Model Equation (DML-LME), an integrated framework combining a high-dimensional lens model with a suite of interpretability techniques. We apply this framework to analyze how an AI perceives social class from 9,513 aspirational essays, comparing a 384-dimension embedding model (all-MiniLM-L6-v2) with a 4,096-dimension model (NV-Embed-v2). While both models achieved similar mediation of the AI’s judgment (64.7% vs. 68.6%), our analysis revealed that both failed to effectively predict the actual environmental criterion (e.g., R² = −.05). Crucially, our interpretability suite uncovered a systematic linguistic bias: essays with poor writing quality were 2.6 times more likely to receive a low social class rating from the AI than to actually originate from a low social class background. This bias was strong enough to override other valid cues, causing the AI to misjudge essays discussing traditional upper-class markers like equestrian activities when they were paired with grammatical errors. The DML-LME, combined with robust interpretability tools, thus enables researchers to not only quantify perception in high-dimensional settings but also to uncover the specific, and potentially discriminatory, heuristics that guide AI judgment.