Comparing Vision–Language Models in Psychological Drawing Analysis

Children’s drawings are used in clinical and developmental assessment, yet their interpretation remains subjective. The emergence of vision–language models (VLMs) raises a central question: do different models produce similar psychological narratives when presented with the same drawing? To investigate this, we evaluated five VLMs interpreting 315 children’s drawings under a shared projective-psychology instruction. We compared the models’ outputs across emotional wording, lexical content, linguistic style, and semantic similarity using drawing-adjusted mixed-effects comparisons and drawing-level bootstrap intervals. The systems produced semantically similar accounts but differed in textual polarity, psychological vocabulary, readability, sentence length, and lexical diversity. Responses to the same drawing were more similar than responses paired across different drawings, although substantial similarity remained in the mismatched comparison. The results show that semantic proximity can coexist with systematic differences in how a psychological narrative is expressed. An exploratory comparison with clinician interpretations provided a professional reference for thematic, emotional, and lexical correspondence. The study contributes a comparative account of VLM-generated language under a common protocol and motivates further evaluation of its use in psychologically sensitive settings, including drawing-based dialogue in children’s language education.