Large Language Models Detect Amplified Abstract Interpretation in Trait Anxiety

Research on the relation of interpretation biases to internalizing symptoms has primarily focused on emotional valence (i.e., negativity bias), leaving other aspects of cognitive processing underexplored. Guided by dual-coding theory (Kousta et al., 2011; Paivio, 1991), we examined how the balance between concrete, perceptual detail and abstract, conceptual elaboration shapes social interpretations in adults with varying levels of trait anxiety and depressive symptoms. A confirmatory sample of 158 U.S. participants wrote interpretations of 24 ambiguous social images (12 negative, 12 non-negative) presented with gradually decreasing blur (Stage 1=80% blur, Stage 2=20%, Stage 3=0%). Large Language Model (LLM; GPT-4.1) analyses were used to compute emotional valence and the proportion of abstract versus concrete language in each response. Trait anxiety symptoms emerged as the most robust correlates of language-based interpretation biases. As preregistered, higher trait anxiety symptoms correlated with more negative interpretations. This pattern was stronger for negative than non-negative scenarios, replicating previous negativity effects. Second, in exploratory analyses, individuals high in trait anxiety symptoms showed an amplified increase in abstract relative to concrete language across stages, indicating a greater shift toward conceptual representation as visual ambiguity in the social scene resolved. The result held even when accounting for non-LLM, lexicon-based relative abstractness scores. These results extend valence-based theories by suggesting that trait anxiety symptoms are associated with amplified abstraction during social interpretations. LLM-derived linguistic indices provide a nuanced, automated method for capturing dual-coding dynamics in natural language, offering new avenues for elucidating latent cognitive processing patterns in psychopathology.